Skip to content

API

Table of Contents

torch algorithm utils

lm_opt 

lm_opt(f, theta0, ytrue, batch_dims=0, iters=10, residtol=None, loss_mult=1, loss_shift=0, f_kwargs_vec={}, f_kwargs_no_vec={}, lam0=1e-06, alpha0=1.0, lam_factors=[[1 / 2, 1, 2]], alpha_factors=[[1 / 2, 1, 2]], vmap_chunk_size=None, jacmode='auto', verbose=False, verbose_indent=4, quantiles_losses=[0, 1, 5, 10, 25, 40, 50, 60, 75, 90, 95, 99, 100], quantiles_lams=[0, 1, 5, 10, 25, 40, 50, 60, 75, 90, 95, 99, 100], quantiles_alphas=[0, 1, 5, 10, 25, 40, 50, 60, 75, 90, 95, 99, 100], verbose_quantiles_losses=[5, 25, 50, 75, 90], verbose_quantiles_lams=[5, 25, 50, 75, 90], verbose_quantiles_alphas=[5, 25, 50, 75, 90], verbose_times=True, warn=True, store_data_iters=False, store_all_data=False)

Levenberg--Marquardt optimization

Parameters:

Name Type Description Default
f func

Residual function.

required
theta0 Tensor

Initial guess for parameters \(\theta\).

required
ytrue Tensor

True y values, i.e. f(theta_true).

required
batch_dims int

Number of batch dimension.

0
iters int

Number of iterations.

10
residtol float

Non-negative tolerance on the maximum residual for early stopping, defaults to 1e-12 for torch.float64 and 2.5e-4 for torch.float32.

 None
loss_mult bool

Scalar amount by which to multiply the loss so loss = loss_mult*torch.sum(resid**2,dim=-1)+loss_shift.

 1
loss_shift bool

Scalar amount by which to shift the loss so loss = loss_mult*torch.sum(resid**2,dim=-1)+loss_shift.

 0
f_kwargs_vec dict

Keyword arguments to f which will be vectorized over the first dimension.

{}
f_kwargs_no_vec dict

Keyword arguments to f which will not be vectorized over the first dimension.

{}
lam0 float

Initial positive relaxation parameter \(\lambda\).

 1e-06
alpha0 float

Initial positive step size \(\alpha\).

 1.0
lam_factors Tensor

Either a 1D torch.Tensor or list of 1d torch.Tensor.

  • Passing in a float for lam_factors is equivalent to passing in torch.tensor([lam_factors]) on the correct device
  • If a 1D torch.Tensor for lam_factors will consider all lam*lam_factors options at each step.
  • If a list of 1D torch.Tensors are passed in for lam_factors, iterations will cycle through the list and then return to the start after exhausting the list.
 [[1 / 2, 1, 2]]
alpha_factors Tensor

Either a 1D torch.Tensor or list of 1d torch.Tensor.

  • Passing in a float for alpha_factors is equivalent to passing in torch.tensor([alpha_factors]) on the correct device
  • If a 1D torch.Tensor for alpha_factors will consider all lam*alpha_factors options at each step.
  • If a list of 1D torch.Tensors are passed in for alpha_factors, iterations will cycle through the list and then return to the start after exhausting the list.
[[1 / 2, 1, 2]]
vmap_chunk_size int

Parameter chunksize to pass to torch.vmap.

 None
jacmode bool

Choose between torch.func.jacfwd and torch.func.jacrev using options:

  • "fwd": Use torch.func.jacfwd.
  • "rev": Use torch.func.jacrev.
  • "auto": Choose between torch.func.jacfwd and torch.func.jacrev depending on the size of the inputs and outputs.
'auto'
verbose int

Controls logging verbosity

  • If True, perform logging.
  • If a positive int, only log every verbose iterations.
  • If None, set to a reasonable positive int based on the maximum number of iterations
  • If False, don't log.
False
verbose_indent int

Non-negative number of indentation spaces for logging.

 4
quantiles_losses list

Loss quantiles to record.

 [0, 1, 5, 10, 25, 40, 50, 60, 75, 90, 95, 99, 100]
quantiles_lams list

\(\lambda\) quantiles to record.

 [0, 1, 5, 10, 25, 40, 50, 60, 75, 90, 95, 99, 100]
quantiles_alphas list

\(\alpha\) quantiles to record.

 [0, 1, 5, 10, 25, 40, 50, 60, 75, 90, 95, 99, 100]
verbose_quantiles_losses list

Loss quantiles to show in verbose log.

 [5, 25, 50, 75, 90]
verbose_quantiles_lams list

\(\lambda\) quantiles to show in verbose log.

 [5, 25, 50, 75, 90]
verbose_quantiles_alphas list

\(\alpha\) quantiles to show in verbose log.

 [5, 25, 50, 75, 90]
verbose_times bool

If False, do not show the times in the verbose log. This is mostly for testing where timing is not reproducible.

True
warn bool

If False, then suppress warnings.

 True
store_data_iters int

Controls storage iterations with the same options as verbose. If store_data_iters==0, then the data is not collected or returned.

  • If True, store every iteration.
  • If a positive int, only store every store_data_iters iterations.
  • If None, set to a reasonable positive int based on the maximum number of iterations
  • If False, don't store data, and do not return data
False
store_all_data bool

If True, store the x values as well as the metrics.

False

Returns:

Name Type Description
theta Tensor

Optimized parameters.
data dict

Iteration data, only returned when store_data_iters>0

Examples:

>>> torch.set_default_dtype(torch.float64)
>>> rng = torch.Generator().manual_seed(7)

Standard example

>>> x = torch.rand((10,4,),generator=rng)
>>> theta_true = torch.rand((4,),generator=rng)
>>> ytrue = torch.exp((x*theta_true).sum(-1)) # (10,)
>>> def f(theta):
...     yhat = torch.exp((x*theta[...,None,:]).sum(-1)) # (...,10)
...     return yhat
>>> theta_hat,data = lm_opt(
...     f = f, 
...     theta0 = torch.rand_like(theta_true,generator=rng),
...     ytrue = ytrue,
...     iters = 3,
...     batch_dims = 0,
...     verbose = True,
...     verbose_times = False,
...     store_data_iters = None,
...     store_all_data = True,
...     )
    iter i     | losses_quantiles                                          | lams_quantiles                                            | alphas_quantiles                                          
               | 5         | 25        | 50        | 75        | 90        | 5         | 25        | 50        | 75        | 90        | 5         | 25        | 50        | 75        | 90        
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    0          | 2.3e+01   | 2.3e+01   | 2.3e+01   | 2.3e+01   | 2.3e+01   | 1.0e-06   | 1.0e-06   | 1.0e-06   | 1.0e-06   | 1.0e-06   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
    1          | 3.4e+00   | 3.4e+00   | 3.4e+00   | 3.4e+00   | 3.4e+00   | 5.0e-07   | 5.0e-07   | 5.0e-07   | 5.0e-07   | 5.0e-07   | 5.0e-01   | 5.0e-01   | 5.0e-01   | 5.0e-01   | 5.0e-01   
    2          | 4.5e-02   | 4.5e-02   | 4.5e-02   | 4.5e-02   | 4.5e-02   | 1.0e-06   | 1.0e-06   | 1.0e-06   | 1.0e-06   | 1.0e-06   | 5.0e-01   | 5.0e-01   | 5.0e-01   | 5.0e-01   | 5.0e-01   
    3          | 4.8e-06   | 4.8e-06   | 4.8e-06   | 4.8e-06   | 4.8e-06   | 5.0e-07   | 5.0e-07   | 5.0e-07   | 5.0e-07   | 5.0e-07   | 5.0e-01   | 5.0e-01   | 5.0e-01   | 5.0e-01   | 5.0e-01   
>>> torch.allclose(theta_hat,theta_true,atol=5e-2)
True
>>> print_data_signatures(data)
    data['theta'].shape = (4,)
    data['iterrange'].shape = (4,)
    data['times'].shape = (4,)
    data['losses_quantiles']
        data['losses_quantiles']['0'].shape = (4,)
        data['losses_quantiles']['1'].shape = (4,)
        data['losses_quantiles']['5'].shape = (4,)
        data['losses_quantiles']['10'].shape = (4,)
        data['losses_quantiles']['25'].shape = (4,)
        data['losses_quantiles']['40'].shape = (4,)
        data['losses_quantiles']['50'].shape = (4,)
        data['losses_quantiles']['60'].shape = (4,)
        data['losses_quantiles']['75'].shape = (4,)
        data['losses_quantiles']['90'].shape = (4,)
        data['losses_quantiles']['95'].shape = (4,)
        data['losses_quantiles']['99'].shape = (4,)
        data['losses_quantiles']['100'].shape = (4,)
    data['lams_quantiles']
        data['lams_quantiles']['0'].shape = (4,)
        data['lams_quantiles']['1'].shape = (4,)
        data['lams_quantiles']['5'].shape = (4,)
        data['lams_quantiles']['10'].shape = (4,)
        data['lams_quantiles']['25'].shape = (4,)
        data['lams_quantiles']['40'].shape = (4,)
        data['lams_quantiles']['50'].shape = (4,)
        data['lams_quantiles']['60'].shape = (4,)
        data['lams_quantiles']['75'].shape = (4,)
        data['lams_quantiles']['90'].shape = (4,)
        data['lams_quantiles']['95'].shape = (4,)
        data['lams_quantiles']['99'].shape = (4,)
        data['lams_quantiles']['100'].shape = (4,)
    data['alphas_quantiles']
        data['alphas_quantiles']['0'].shape = (4,)
        data['alphas_quantiles']['1'].shape = (4,)
        data['alphas_quantiles']['5'].shape = (4,)
        data['alphas_quantiles']['10'].shape = (4,)
        data['alphas_quantiles']['25'].shape = (4,)
        data['alphas_quantiles']['40'].shape = (4,)
        data['alphas_quantiles']['50'].shape = (4,)
        data['alphas_quantiles']['60'].shape = (4,)
        data['alphas_quantiles']['75'].shape = (4,)
        data['alphas_quantiles']['90'].shape = (4,)
        data['alphas_quantiles']['95'].shape = (4,)
        data['alphas_quantiles']['99'].shape = (4,)
        data['alphas_quantiles']['100'].shape = (4,)
    data['thetas'].shape = (4, 4)
    data['losses'].shape = (4,)
    data['lams'].shape = (4,)
    data['alphas'].shape = (4,)

Batched example 

>>> x = torch.rand((3,3,3,2,2),generator=rng)
>>> theta_true = torch.rand((4,4,2,2),generator=rng)
>>> ytrue = torch.exp((x*theta_true[...,None,None,None,:,:]).sum((-2,-1))) # (4,4,3,3,3)
>>> def f(theta):
...     yhat = torch.exp((x*theta[...,None,None,None,:,:]).sum((-2,-1))) # (...,3,3,3)
...     return yhat
>>> theta_hat,data = lm_opt(
...     f = f, 
...     theta0 = torch.rand_like(theta_true,generator=rng),
...     ytrue = ytrue,
...     iters = 2,
...     batch_dims = 2,
...     lam_factors = [torch.tensor([1/4,1/2,1,2,4])],
...     alpha_factors = [torch.tensor([2/3,1,3/2])],
...     verbose = True,
...     verbose_times = False,
...     store_data_iters = None,
...     store_all_data = True,
...     )
    iter i     | losses_quantiles                                          | lams_quantiles                                            | alphas_quantiles                                          
               | 5         | 25        | 50        | 75        | 90        | 5         | 25        | 50        | 75        | 90        | 5         | 25        | 50        | 75        | 90        
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    0          | 5.5e+00   | 2.5e+01   | 5.3e+01   | 1.3e+02   | 5.1e+02   | 1.0e-06   | 1.0e-06   | 1.0e-06   | 1.0e-06   | 1.0e-06   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
    1          | 7.4e-02   | 6.4e-01   | 1.1e+00   | 2.2e+00   | 7.7e+01   | 2.5e-07   | 2.5e-07   | 2.5e-07   | 2.5e-07   | 4.0e-06   | 6.7e-01   | 6.7e-01   | 1.0e+00   | 1.1e+00   | 1.5e+00   
    2          | 1.0e-05   | 6.9e-04   | 1.8e-03   | 3.9e-03   | 1.2e+00   | 6.2e-08   | 6.2e-08   | 1.0e-06   | 1.0e-06   | 1.0e-06   | 6.7e-01   | 6.7e-01   | 1.0e+00   | 1.1e+00   | 1.5e+00   
>>> torch.allclose(theta_hat,theta_true,atol=5e-2)
False
>>> print_data_signatures(data)
    data['theta'].shape = (4, 4, 2, 2)
    data['iterrange'].shape = (3,)
    data['times'].shape = (3,)
    data['losses_quantiles']
        data['losses_quantiles']['0'].shape = (3,)
        data['losses_quantiles']['1'].shape = (3,)
        data['losses_quantiles']['5'].shape = (3,)
        data['losses_quantiles']['10'].shape = (3,)
        data['losses_quantiles']['25'].shape = (3,)
        data['losses_quantiles']['40'].shape = (3,)
        data['losses_quantiles']['50'].shape = (3,)
        data['losses_quantiles']['60'].shape = (3,)
        data['losses_quantiles']['75'].shape = (3,)
        data['losses_quantiles']['90'].shape = (3,)
        data['losses_quantiles']['95'].shape = (3,)
        data['losses_quantiles']['99'].shape = (3,)
        data['losses_quantiles']['100'].shape = (3,)
    data['lams_quantiles']
        data['lams_quantiles']['0'].shape = (3,)
        data['lams_quantiles']['1'].shape = (3,)
        data['lams_quantiles']['5'].shape = (3,)
        data['lams_quantiles']['10'].shape = (3,)
        data['lams_quantiles']['25'].shape = (3,)
        data['lams_quantiles']['40'].shape = (3,)
        data['lams_quantiles']['50'].shape = (3,)
        data['lams_quantiles']['60'].shape = (3,)
        data['lams_quantiles']['75'].shape = (3,)
        data['lams_quantiles']['90'].shape = (3,)
        data['lams_quantiles']['95'].shape = (3,)
        data['lams_quantiles']['99'].shape = (3,)
        data['lams_quantiles']['100'].shape = (3,)
    data['alphas_quantiles']
        data['alphas_quantiles']['0'].shape = (3,)
        data['alphas_quantiles']['1'].shape = (3,)
        data['alphas_quantiles']['5'].shape = (3,)
        data['alphas_quantiles']['10'].shape = (3,)
        data['alphas_quantiles']['25'].shape = (3,)
        data['alphas_quantiles']['40'].shape = (3,)
        data['alphas_quantiles']['50'].shape = (3,)
        data['alphas_quantiles']['60'].shape = (3,)
        data['alphas_quantiles']['75'].shape = (3,)
        data['alphas_quantiles']['90'].shape = (3,)
        data['alphas_quantiles']['95'].shape = (3,)
        data['alphas_quantiles']['99'].shape = (3,)
        data['alphas_quantiles']['100'].shape = (3,)
    data['thetas'].shape = (3, 4, 4, 2, 2)
    data['losses'].shape = (3, 4, 4)
    data['lams'].shape = (3, 4, 4)
    data['alphas'].shape = (3, 4, 4)
Source code in agsutil/algos.py 
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
def lm_opt(
        f,
        theta0,
        ytrue,
        batch_dims = 0, 
        iters = 10,
        residtol = None,
        loss_mult = 1,
        loss_shift = 0,
        f_kwargs_vec = {},
        f_kwargs_no_vec = {},
        lam0 = 1e-6,
        alpha0 = 1e0,
        lam_factors = [[1/2,1,2]],
        alpha_factors = [[1/2,1,2]],
        vmap_chunk_size = None,
        jacmode = "auto",
        verbose = False, 
        verbose_indent = 4,
        quantiles_losses = [0,1,5,10,25,40,50,60,75,90,95,99,100],
        quantiles_lams =   [0,1,5,10,25,40,50,60,75,90,95,99,100],
        quantiles_alphas = [0,1,5,10,25,40,50,60,75,90,95,99,100],
        verbose_quantiles_losses = [5,25,50,75,90],
        verbose_quantiles_lams =   [5,25,50,75,90],
        verbose_quantiles_alphas = [5,25,50,75,90],
        verbose_times = True, 
        warn = True,
        store_data_iters = False,
        store_all_data = False, 
        ):
    r"""
    Levenberg--Marquardt optimization 

    Args:
        f (func): Residual function. 
        theta0 (torch.Tensor): Initial guess for parameters $\theta$. 
        ytrue (torch.Tensor): True `y` values, i.e. `f(theta_true)`. 
        batch_dims (int): Number of batch dimension. 
        iters (int): Number of iterations. 
        residtol (float): Non-negative tolerance on the maximum residual for early stopping, defaults to `1e-12` for `torch.float64` and `2.5e-4` for `torch.float32`.
        loss_mult (bool): Scalar amount by which to multiply the loss so `loss = loss_mult*torch.sum(resid**2,dim=-1)+loss_shift`.
        loss_shift (bool): Scalar amount by which to shift the loss so `loss = loss_mult*torch.sum(resid**2,dim=-1)+loss_shift`.
        f_kwargs_vec (dict): Keyword arguments to `f` which will be vectorized over the first dimension. 
        f_kwargs_no_vec (dict): Keyword arguments to `f` which will not be vectorized over the first dimension. 
        lam0 (float): Initial positive relaxation parameter $\lambda$.
        alpha0 (float): Initial positive step size $\alpha$.
        lam_factors (torch.Tensor): Either a 1D `torch.Tensor` or list of 1d `torch.Tensor`. 

            - Passing in a `float` for `lam_factors` is equivalent to passing in `torch.tensor([lam_factors])` on the correct device
            - If a 1D `torch.Tensor` for `lam_factors` will consider all `lam*lam_factors` options at each step. 
            - If a list of 1D `torch.Tensor`s are passed in for `lam_factors`, iterations will cycle through the list and then return to the start after exhausting the list.

        alpha_factors (torch.Tensor): Either a 1D `torch.Tensor` or list of 1d `torch.Tensor`. 

            - Passing in a `float` for `alpha_factors` is equivalent to passing in `torch.tensor([alpha_factors])` on the correct device
            - If a 1D `torch.Tensor` for `alpha_factors` will consider all `lam*alpha_factors` options at each step. 
            - If a list of 1D `torch.Tensor`s are passed in for `alpha_factors`, iterations will cycle through the list and then return to the start after exhausting the list. 

        vmap_chunk_size (int): Parameter `chunksize` to pass to `torch.vmap`.
        jacmode (bool): Choose between `torch.func.jacfwd` and `torch.func.jacrev` using options:

            - `"fwd"`: Use `torch.func.jacfwd`.
            - `"rev"`: Use `torch.func.jacrev`.
            - `"auto"`: Choose between `torch.func.jacfwd` and `torch.func.jacrev` depending on the size of the inputs and outputs. 

        verbose (int): Controls logging verbosity

            - If `True`, perform logging. 
            - If a positive int, only log every verbose iterations. 
            - If `None`, set to a reasonable positive int based on the maximum number of iterations
            - If `False`, don't log. 

        verbose_indent (int): Non-negative number of indentation spaces for logging.
        quantiles_losses (list): Loss quantiles to record.
        quantiles_lams (list): $\lambda$ quantiles to record.
        quantiles_alphas (list): $\alpha$ quantiles to record.
        verbose_quantiles_losses (list): Loss quantiles to show in verbose log.
        verbose_quantiles_lams (list): $\lambda$ quantiles to show in verbose log.
        verbose_quantiles_alphas (list): $\alpha$ quantiles to show in verbose log.
        verbose_times (bool): If `False`, do not show the times in the verbose log. This is mostly for testing where timing is not reproducible. 
        warn (bool): If `False`, then suppress warnings.
        store_data_iters (int): Controls storage iterations with the same options as verbose. If `store_data_iters==0`, then the data is not collected or returned. 

            - If `True`, store every iteration. 
            - If a positive int, only store every `store_data_iters` iterations. 
            - If `None`, set to a reasonable positive int based on the maximum number of iterations
            - If `False`, don't store data, and do not return data 

        store_all_data (bool): If `True`, store the `x` values as well as the metrics. 

    Returns:
        theta (torch.Tensor): Optimized parameters.
        data (dict): Iteration data, only returned when `store_data_iters>0`

    Examples:

        >>> torch.set_default_dtype(torch.float64)
        >>> rng = torch.Generator().manual_seed(7)


    Standard example

        >>> x = torch.rand((10,4,),generator=rng)
        >>> theta_true = torch.rand((4,),generator=rng)
        >>> ytrue = torch.exp((x*theta_true).sum(-1)) # (10,)
        >>> def f(theta):
        ...     yhat = torch.exp((x*theta[...,None,:]).sum(-1)) # (...,10)
        ...     return yhat
        >>> theta_hat,data = lm_opt(
        ...     f = f, 
        ...     theta0 = torch.rand_like(theta_true,generator=rng),
        ...     ytrue = ytrue,
        ...     iters = 3,
        ...     batch_dims = 0,
        ...     verbose = True,
        ...     verbose_times = False,
        ...     store_data_iters = None,
        ...     store_all_data = True,
        ...     )
            iter i     | losses_quantiles                                          | lams_quantiles                                            | alphas_quantiles                                          
                       | 5         | 25        | 50        | 75        | 90        | 5         | 25        | 50        | 75        | 90        | 5         | 25        | 50        | 75        | 90        
            ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            0          | 2.3e+01   | 2.3e+01   | 2.3e+01   | 2.3e+01   | 2.3e+01   | 1.0e-06   | 1.0e-06   | 1.0e-06   | 1.0e-06   | 1.0e-06   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
            1          | 3.4e+00   | 3.4e+00   | 3.4e+00   | 3.4e+00   | 3.4e+00   | 5.0e-07   | 5.0e-07   | 5.0e-07   | 5.0e-07   | 5.0e-07   | 5.0e-01   | 5.0e-01   | 5.0e-01   | 5.0e-01   | 5.0e-01   
            2          | 4.5e-02   | 4.5e-02   | 4.5e-02   | 4.5e-02   | 4.5e-02   | 1.0e-06   | 1.0e-06   | 1.0e-06   | 1.0e-06   | 1.0e-06   | 5.0e-01   | 5.0e-01   | 5.0e-01   | 5.0e-01   | 5.0e-01   
            3          | 4.8e-06   | 4.8e-06   | 4.8e-06   | 4.8e-06   | 4.8e-06   | 5.0e-07   | 5.0e-07   | 5.0e-07   | 5.0e-07   | 5.0e-07   | 5.0e-01   | 5.0e-01   | 5.0e-01   | 5.0e-01   | 5.0e-01   
        >>> torch.allclose(theta_hat,theta_true,atol=5e-2)
        True
        >>> print_data_signatures(data)
            data['theta'].shape = (4,)
            data['iterrange'].shape = (4,)
            data['times'].shape = (4,)
            data['losses_quantiles']
                data['losses_quantiles']['0'].shape = (4,)
                data['losses_quantiles']['1'].shape = (4,)
                data['losses_quantiles']['5'].shape = (4,)
                data['losses_quantiles']['10'].shape = (4,)
                data['losses_quantiles']['25'].shape = (4,)
                data['losses_quantiles']['40'].shape = (4,)
                data['losses_quantiles']['50'].shape = (4,)
                data['losses_quantiles']['60'].shape = (4,)
                data['losses_quantiles']['75'].shape = (4,)
                data['losses_quantiles']['90'].shape = (4,)
                data['losses_quantiles']['95'].shape = (4,)
                data['losses_quantiles']['99'].shape = (4,)
                data['losses_quantiles']['100'].shape = (4,)
            data['lams_quantiles']
                data['lams_quantiles']['0'].shape = (4,)
                data['lams_quantiles']['1'].shape = (4,)
                data['lams_quantiles']['5'].shape = (4,)
                data['lams_quantiles']['10'].shape = (4,)
                data['lams_quantiles']['25'].shape = (4,)
                data['lams_quantiles']['40'].shape = (4,)
                data['lams_quantiles']['50'].shape = (4,)
                data['lams_quantiles']['60'].shape = (4,)
                data['lams_quantiles']['75'].shape = (4,)
                data['lams_quantiles']['90'].shape = (4,)
                data['lams_quantiles']['95'].shape = (4,)
                data['lams_quantiles']['99'].shape = (4,)
                data['lams_quantiles']['100'].shape = (4,)
            data['alphas_quantiles']
                data['alphas_quantiles']['0'].shape = (4,)
                data['alphas_quantiles']['1'].shape = (4,)
                data['alphas_quantiles']['5'].shape = (4,)
                data['alphas_quantiles']['10'].shape = (4,)
                data['alphas_quantiles']['25'].shape = (4,)
                data['alphas_quantiles']['40'].shape = (4,)
                data['alphas_quantiles']['50'].shape = (4,)
                data['alphas_quantiles']['60'].shape = (4,)
                data['alphas_quantiles']['75'].shape = (4,)
                data['alphas_quantiles']['90'].shape = (4,)
                data['alphas_quantiles']['95'].shape = (4,)
                data['alphas_quantiles']['99'].shape = (4,)
                data['alphas_quantiles']['100'].shape = (4,)
            data['thetas'].shape = (4, 4)
            data['losses'].shape = (4,)
            data['lams'].shape = (4,)
            data['alphas'].shape = (4,)

    Batched example 

        >>> x = torch.rand((3,3,3,2,2),generator=rng)
        >>> theta_true = torch.rand((4,4,2,2),generator=rng)
        >>> ytrue = torch.exp((x*theta_true[...,None,None,None,:,:]).sum((-2,-1))) # (4,4,3,3,3)
        >>> def f(theta):
        ...     yhat = torch.exp((x*theta[...,None,None,None,:,:]).sum((-2,-1))) # (...,3,3,3)
        ...     return yhat
        >>> theta_hat,data = lm_opt(
        ...     f = f, 
        ...     theta0 = torch.rand_like(theta_true,generator=rng),
        ...     ytrue = ytrue,
        ...     iters = 2,
        ...     batch_dims = 2,
        ...     lam_factors = [torch.tensor([1/4,1/2,1,2,4])],
        ...     alpha_factors = [torch.tensor([2/3,1,3/2])],
        ...     verbose = True,
        ...     verbose_times = False,
        ...     store_data_iters = None,
        ...     store_all_data = True,
        ...     )
            iter i     | losses_quantiles                                          | lams_quantiles                                            | alphas_quantiles                                          
                       | 5         | 25        | 50        | 75        | 90        | 5         | 25        | 50        | 75        | 90        | 5         | 25        | 50        | 75        | 90        
            ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            0          | 5.5e+00   | 2.5e+01   | 5.3e+01   | 1.3e+02   | 5.1e+02   | 1.0e-06   | 1.0e-06   | 1.0e-06   | 1.0e-06   | 1.0e-06   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
            1          | 7.4e-02   | 6.4e-01   | 1.1e+00   | 2.2e+00   | 7.7e+01   | 2.5e-07   | 2.5e-07   | 2.5e-07   | 2.5e-07   | 4.0e-06   | 6.7e-01   | 6.7e-01   | 1.0e+00   | 1.1e+00   | 1.5e+00   
            2          | 1.0e-05   | 6.9e-04   | 1.8e-03   | 3.9e-03   | 1.2e+00   | 6.2e-08   | 6.2e-08   | 1.0e-06   | 1.0e-06   | 1.0e-06   | 6.7e-01   | 6.7e-01   | 1.0e+00   | 1.1e+00   | 1.5e+00   
        >>> torch.allclose(theta_hat,theta_true,atol=5e-2)
        False
        >>> print_data_signatures(data)
            data['theta'].shape = (4, 4, 2, 2)
            data['iterrange'].shape = (3,)
            data['times'].shape = (3,)
            data['losses_quantiles']
                data['losses_quantiles']['0'].shape = (3,)
                data['losses_quantiles']['1'].shape = (3,)
                data['losses_quantiles']['5'].shape = (3,)
                data['losses_quantiles']['10'].shape = (3,)
                data['losses_quantiles']['25'].shape = (3,)
                data['losses_quantiles']['40'].shape = (3,)
                data['losses_quantiles']['50'].shape = (3,)
                data['losses_quantiles']['60'].shape = (3,)
                data['losses_quantiles']['75'].shape = (3,)
                data['losses_quantiles']['90'].shape = (3,)
                data['losses_quantiles']['95'].shape = (3,)
                data['losses_quantiles']['99'].shape = (3,)
                data['losses_quantiles']['100'].shape = (3,)
            data['lams_quantiles']
                data['lams_quantiles']['0'].shape = (3,)
                data['lams_quantiles']['1'].shape = (3,)
                data['lams_quantiles']['5'].shape = (3,)
                data['lams_quantiles']['10'].shape = (3,)
                data['lams_quantiles']['25'].shape = (3,)
                data['lams_quantiles']['40'].shape = (3,)
                data['lams_quantiles']['50'].shape = (3,)
                data['lams_quantiles']['60'].shape = (3,)
                data['lams_quantiles']['75'].shape = (3,)
                data['lams_quantiles']['90'].shape = (3,)
                data['lams_quantiles']['95'].shape = (3,)
                data['lams_quantiles']['99'].shape = (3,)
                data['lams_quantiles']['100'].shape = (3,)
            data['alphas_quantiles']
                data['alphas_quantiles']['0'].shape = (3,)
                data['alphas_quantiles']['1'].shape = (3,)
                data['alphas_quantiles']['5'].shape = (3,)
                data['alphas_quantiles']['10'].shape = (3,)
                data['alphas_quantiles']['25'].shape = (3,)
                data['alphas_quantiles']['40'].shape = (3,)
                data['alphas_quantiles']['50'].shape = (3,)
                data['alphas_quantiles']['60'].shape = (3,)
                data['alphas_quantiles']['75'].shape = (3,)
                data['alphas_quantiles']['90'].shape = (3,)
                data['alphas_quantiles']['95'].shape = (3,)
                data['alphas_quantiles']['99'].shape = (3,)
                data['alphas_quantiles']['100'].shape = (3,)
            data['thetas'].shape = (3, 4, 4, 2, 2)
            data['losses'].shape = (3, 4, 4)
            data['lams'].shape = (3, 4, 4)
            data['alphas'].shape = (3, 4, 4)
    """
    if warn and (not torch.get_default_dtype()==torch.float64): warnings.warn('''
            torch.get_default_dtype() = %s, but lm_opt often requires high precision updates. We recommend using:
                torch.set_default_dtype(torch.float64)'''%str(torch.get_default_dtype()))
    assert torch.get_default_dtype() in [torch.float32,torch.float64]
    default_dtype = torch.get_default_dtype()
    device = str(theta0.device)
    default_device = str(torch.get_default_device())
    assert iters%1==0, "iters should be an int"
    assert iters>=0
    assert callable(f) 
    assert batch_dims>=0
    assert isinstance(theta0,torch.Tensor)
    batch_shape = tuple(theta0.shape[:batch_dims])
    R = int(torch.tensor(batch_shape).prod())
    nonbatch_theta_dims = theta0.ndim-batch_dims
    nonbatch_theta_shape = tuple(theta0.shape[batch_dims:])
    nonbatch_y_dims = ytrue.ndim-batch_dims
    nonbatch_y_shape = tuple(ytrue.shape[batch_dims:])
    K = int(torch.tensor(nonbatch_y_shape).prod())
    T = int(torch.tensor(nonbatch_theta_shape).prod())
    if batch_dims==0:
        theta = theta0[None,...]
    else: # batch_dims>0:
        theta = theta0.flatten(end_dim=batch_dims-1)
    if batch_dims==0:
        ytrue = ytrue[None,...]
    else: # batch_dims>0:
        ytrue = ytrue.flatten(end_dim=batch_dims-1)
    assert isinstance(f_kwargs_vec,dict)
    assert isinstance(f_kwargs_no_vec,dict)
    f_kwargs_vec_names = list(f_kwargs_vec.keys())
    f_kwargs_vec_vals = []
    for key in f_kwargs_vec_names:
        assert f_kwargs_vec[key].shape[:batch_dims]==batch_shape, "f_kwargs_vec['%s'].shape[:%d] = %s but bs = %s"%(key,batch_dims,f_kwargs_vec[key].shape[:batch_dims])
        if batch_dims==0:
            f_kwargs_vec_vals.append(f_kwargs_vec[key][None,...])
        else: # batch_dims>0
            f_kwargs_vec_vals.append(f_kwargs_vec[key].flatten(end_dim=batch_dims-1))
    f_kwargs_vec_names = ["ytrue"]+f_kwargs_vec_names
    f_kwargs_vec_vals = [ytrue]+f_kwargs_vec_vals
    if verbose is None: 
        verbose = max(1,iters//20)
    assert verbose%1==0
    assert verbose>=0 
    if store_data_iters is None: 
        store_data_iters = max(1,iters//1000)
    assert store_data_iters%1==0
    assert store_data_iters>=0 
    assert isinstance(store_all_data,bool)
    loss_mult = float(loss_mult)
    assert loss_mult!=0 
    signminimize = -1 if loss_mult<0 else 1
    loss_shift = float(loss_shift)
    if residtol is None: 
        if default_dtype==torch.float64:
            residtol = 1e-12
        elif default_dtype==torch.float32:
            residtol = 2.5e-4
        else:
            raise Exception("default_dtype = %s not parsed"%str(default_dtype))
    assert residtol>=0
    assert lam0>0
    assert alpha0>0
    if np.isscalar(lam_factors):
        lam_factors = [torch.tensor([lam_factors],device=device)]
    elif isinstance(lam_factors,torch.Tensor):
        lam_factors = [lam_factors.to(device)]
    lam_factors = [torch.tensor(list(lam_factors[i])).to(device) for i in range(len(lam_factors))]
    assert isinstance(lam_factors,list)
    assert all(isinstance(lam_factors[i],torch.Tensor) for i in range(len(lam_factors)))
    assert all(lam_factors[i].ndim==1 for i in range(len(lam_factors)))
    if np.isscalar(alpha_factors):
        alpha_factors = [torch.tensor([alpha_factors],device=device)]
    elif isinstance(alpha_factors,torch.Tensor):
        alpha_factors = [alpha_factors.to(device)]
    alpha_factors = [torch.tensor(list(alpha_factors[i])).to(device) for i in range(len(alpha_factors))]
    assert isinstance(alpha_factors,list)
    assert all(isinstance(alpha_factors[i],torch.Tensor) for i in range(len(alpha_factors)))
    assert all(alpha_factors[i].ndim==1 for i in range(len(alpha_factors)))
    if len(alpha_factors)==1:
        alpha_factors = alpha_factors*len(lam_factors)
    if len(lam_factors)==1:
        lam_factors = lam_factors*len(alpha_factors)
    assert len(lam_factors)==len(alpha_factors)
    assert isinstance(quantiles_losses,list)
    assert all(0<=qt<=100 for qt in quantiles_losses)
    assert isinstance(quantiles_lams,list)
    assert all(0<=qt<=100 for qt in quantiles_lams)
    assert isinstance(quantiles_alphas,list)
    assert all(0<=qt<=100 for qt in quantiles_alphas)
    assert isinstance(verbose_quantiles_losses,list)
    assert all(qt in quantiles_losses for qt in verbose_quantiles_losses)
    assert isinstance(verbose_quantiles_lams,list)
    assert all(qt in quantiles_lams for qt in verbose_quantiles_lams)
    assert isinstance(verbose_quantiles_alphas,list)
    assert all(qt in quantiles_alphas for qt in verbose_quantiles_alphas)
    assert verbose_indent%1==0 
    assert verbose_indent>=0
    assert isinstance(verbose_times,bool)
    if store_data_iters:
        iterrange = []
        times = []
        losses = []
        losses_quantiles = {str(qt):[] for qt in quantiles_losses}
        lams_quantiles = {str(qt):[] for qt in quantiles_lams}
        alphas_quantiles = {str(qt):[] for qt in quantiles_alphas}
        if store_all_data:
            thetas = []
            lams = []
            alphas = []
    def f_resid(theta, *f_kwargs_vec_vals):
        assert len(f_kwargs_vec_vals)==len(f_kwargs_vec_names)
        ytrue = f_kwargs_vec_vals[0]
        f_kwargs_vec = {f_kwargs_vec_names[i]:f_kwargs_vec_vals[i] for i in range(1,len(f_kwargs_vec_names))}
        all_args = f(theta,**f_kwargs_vec,**f_kwargs_no_vec)
        assert isinstance(all_args,tuple) or isinstance(all_args,torch.Tensor), "f must return a tuple or torch.Tensor"
        if isinstance(all_args,torch.Tensor):
            yhat = all_args
            args = ()
        else: #  isinstance(all_args,tuple)
            assert all(isinstance(arg,torch.Tensor) for arg in all_args)
            yhat = all_args[0]
            args = all_args[1:]
        resid = yhat-ytrue
        return resid,(resid,yhat,*args)
    assert jacmode in ["auto","fwd","rev"]
    if jacmode=="auto":
        if T<=K:
            jacfunc = torch.func.jacfwd
        else:
            jacfunc = torch.func.jacrev
    elif jacmode=="fwd":
        jacfunc = torch.func.jacfwd
        if warn and T>K: warnings.warn('''
            For T the number of inputs and K the number of outputs:
                torch.func.jacfwd performs best when T << K. 
                torch.func.jacrev performs best when K << T.
            You are using jacmode=='fwd' but T = %d > %d = K. 
            Try using jacmode=='rev' by setting jacfwd = False.'''%(T,K))
    elif jacmode=="rev":
        jacfunc = torch.func.jacrev
        if warn and T<K: warnings.warn('''
            For T the number of inputs and K the number of outputs:
                torch.func.jacfwd performs best when T << K. 
                torch.func.jacrev performs best when K << T.
            You are using jacmode=='rev' but T = %d < %d = K. 
            Try using jacmode=='fwd' for better performance.'''%(T,K))
    jac_ftilde = jacfunc(f_resid,argnums=(0,),has_aux=True)
    vjac_ftilde = torch.func.vmap(jac_ftilde,in_dims=(0,)+(0,)*len(f_kwargs_vec_names),chunk_size=vmap_chunk_size)
    eyeT = torch.eye(T,device=device)
    Rrange = torch.arange(R,device=device)
    lam = lam0*torch.ones(R,device=device)
    alpha = alpha0*torch.ones(R,device=device)
    if verbose:
        _h_iter = "%-10s "%"iter i"
        _h_times = "| %-10s"%"times" if verbose_times else ""
        _s_losses_qt = ("| %-9s "*len(verbose_quantiles_losses))%tuple(str(qt) for qt in verbose_quantiles_losses)
        _s_lams_qt = ("| %-9s "*len(verbose_quantiles_lams))%tuple(str(qt) for qt in verbose_quantiles_lams)
        _s_alphas_qt = ("| %-9s "*len(verbose_quantiles_alphas))%tuple(str(qt) for qt in verbose_quantiles_alphas)
        _h_losses_qt = "| losses_quantiles"+" "*(len(_s_losses_qt)-len("| losses_quantiles"))
        _h_lams_qt   = "| lams_quantiles"  +" "*(len(_s_lams_qt)  -len("| lams_quantiles"))
        _h_alphas_qt = "| alphas_quantiles"+" "*(len(_s_alphas_qt)-len("| alphas_quantiles"))
        _h = _h_iter+_h_losses_qt+_h_lams_qt+_h_alphas_qt+_h_times
        _s = " "*len(_h_iter)+_s_losses_qt+_s_lams_qt+_s_alphas_qt+("|"+" "*(len(_h_times)-1) if verbose_times else " "*len(_h_times))
        print(" "*verbose_indent+_h)
        print(" "*verbose_indent+_s)
        print(" "*verbose_indent+"~"*len(_s))
    timer = Timer(device=device)
    timer.tic()
    for i in range(iters+1):
        if i==iters:
            _,(resid,yhat,*args) = f_resid(theta,*f_kwargs_vec_vals)
        else:
            (Jfull,),(resid,yhat,*args) = vjac_ftilde(theta,*f_kwargs_vec_vals)
        assert Jfull.shape==(R,*nonbatch_y_shape,*nonbatch_theta_shape)
        breakcond = i==iters or resid.abs().amax()<=residtol
        loss = loss_mult*(resid**2).flatten(start_dim=1).sum(-1)+loss_shift
        times_i = timer.toc()
        losses_quantiles_i = {str(qt): loss.nanquantile(qt/100) for qt in quantiles_losses}
        lams_quantiles_i = {str(qt): lam.nanquantile(qt/100) for qt in quantiles_lams}
        alphas_quantiles_i = {str(qt): alpha.nanquantile(qt/100) for qt in quantiles_alphas}
        if store_data_iters and (i%store_data_iters==0 or breakcond):
            iterrange.append(i)
            losses.append(loss.reshape(batch_shape).to(default_device))
            times.append(times_i)
            for qt in quantiles_losses:
                losses_quantiles[str(qt)].append(losses_quantiles_i[str(qt)].to(default_device))
            for qt in quantiles_lams:
                lams_quantiles[str(qt)].append(lams_quantiles_i[str(qt)].to(default_device))
            for qt in quantiles_alphas:
                alphas_quantiles[str(qt)].append(alphas_quantiles_i[str(qt)].to(default_device))
            if store_all_data:
                thetas.append(theta.reshape(*batch_shape,*nonbatch_theta_shape).to(default_device))
                lams.append(lam.reshape(batch_shape).to(default_device))
                alphas.append(alpha.reshape(batch_shape).to(default_device))
        if verbose and (i%verbose==0 or i==iters):
            _s_iter = "%-10d "%i
            _s_losses_qt = ("| %-9.1e "*len(verbose_quantiles_losses))%tuple(losses_quantiles_i[str(qt)] for qt in verbose_quantiles_losses)
            _s_lams_qt = ("| %-9.1e "*len(verbose_quantiles_lams))%tuple(lams_quantiles_i[str(qt)] for qt in verbose_quantiles_lams)
            _s_alphas_qt = ("| %-9.1e "*len(verbose_quantiles_alphas))%tuple(alphas_quantiles_i[str(qt)] for qt in verbose_quantiles_alphas)
            _s_times = "| %-10.1f "%(times_i) if verbose_times else ""
            print(" "*verbose_indent+_s_iter+_s_losses_qt+_s_lams_qt+_s_alphas_qt+_s_times)
        if breakcond: break
        J = Jfull.reshape((R,K,T))
        residf = resid.reshape((R,K))
        gamma = torch.einsum("rij,ri->rj",J,residf)
        JtJ = torch.einsum("rij,ril->rjl",J,J) # (R,T,T)
        lam_factors_i = lam_factors[i%len(lam_factors)]
        alpha_factors_i = alpha_factors[i%len(alpha_factors)]
        Q_lams = len(lam_factors_i)
        Q_alphas = len(alpha_factors_i)
        lams_try = lam_factors_i[:,None]*lam # (Q_lams,R)
        alphas_try = alpha_factors_i[:,None]*alpha # (Q_alphas,R)
        JtJplam = JtJ[None,:,:,:]+lams_try[:,:,None,None]*eyeT # (Q_lams,R,T,T)
        L,fails = torch.linalg.cholesky_ex(JtJplam,upper=False) # L.shape==(Q_lams,R,T,T) and fails.shape==(Q_lams,R)
        success = ~fails.to(bool) # (Q_lams,R)
        deltaf = torch.nan*torch.ones((Q_lams,R,T),device=device)
        gammas = torch.ones((Q_lams,1,1),device=device)*gamma[None,:,:] # (Q_lams,R,T)
        deltaf[success] = torch.linalg.solve_triangular(L[success].transpose(dim0=-2,dim1=-1),torch.linalg.solve_triangular(L[success],gammas[success].unsqueeze(-1),upper=False),upper=True)[...,0] # (Q_lam,R,T)
        thetasf = torch.ones((Q_lams,1,1),device=device)*theta.reshape((1,R,T)) # (Q_lam,R,T)
        thetasf_new = torch.nan*torch.ones((Q_alphas,Q_lams,R,T),device=device)
        thetasf_new[:,success] = thetasf[success]-signminimize*alpha_factors_i[:,None,None]*deltaf[success]
        thetas_new = thetasf_new.reshape((Q_alphas,Q_lams,R,*nonbatch_theta_shape))
        f_kwargs_vec_vals_success = [(torch.ones((Q_alphas,Q_lams)+(1,)*f_kwargs_vec_vals[l].ndim,device=device)*f_kwargs_vec_vals[l][None,None,...])[:,success] for l in range(len(f_kwargs_vec_vals))]
        residf_new = torch.inf*torch.ones((Q_alphas,Q_lams,R,K),device=device)
        _,(resid_new_success,*_) = f_resid(thetas_new[:,success],*f_kwargs_vec_vals_success)
        residf_new[:,success] = resid_new_success.reshape((Q_alphas,resid_new_success.size(1),K))
        losses_new = loss_mult*(residf_new**2).sum(-1)+loss_shift
        ibest = losses_new.reshape((Q_alphas*Q_lams,R)).argmin(0) # (R,)
        ibest_alpha,ibest_lam = ibest//Q_lams,ibest%Q_lams
        lam_best_new = lams_try[ibest_lam,Rrange] # (R,)
        alpha_best_new = alphas_try[ibest_alpha,Rrange] # (R,)
        loss_best_new = losses_new[ibest_alpha,ibest_lam,Rrange] # (R,)
        thetas_best_new = thetas_new[ibest_alpha,ibest_lam,Rrange] # (R,*nonbatch_theta_shape)
        improved = loss_best_new<loss # (R,)
        lam[improved] = lam_best_new[improved]
        alpha[improved] = alpha_best_new[improved]
        theta[improved] = thetas_best_new[improved]
        lam[~improved] = lams_try[:,~improved].amax(0)
        alpha[~improved] = alphas_try[:,~improved].amin(0)
    theta = theta.reshape((*batch_shape,*nonbatch_theta_shape))
    if batch_shape==():
        args = [arg[0] for arg in args]
    if store_data_iters==0:
        if args==[]:
            return theta
        else:
            return theta,*args
    else:
        data = {
            "theta": theta.to(default_device), 
            "iterrange": torch.tensor(iterrange,dtype=int), 
            "times": torch.tensor(times), 
            "losses_quantiles": {str(qt):torch.tensor(losses_quantiles[str(qt)]) for qt in quantiles_losses},
            "lams_quantiles": {str(qt):torch.tensor(lams_quantiles[str(qt)]) for qt in quantiles_lams},
            "alphas_quantiles": {str(qt):torch.tensor(alphas_quantiles[str(qt)]) for qt in quantiles_alphas},
            }
        if store_all_data:
            data["thetas"] = torch.stack(thetas,dim=0)
            data["losses"] = torch.stack(losses,dim=0)
            data["lams"] = torch.stack(lams,dim=0)
            data["alphas"] = torch.stack(alphas,dim=0)
        if args==[]:
            return theta,data
        else:
            return theta,*args,data

pcg 

pcg(A, B, X0=None, precond_solver=None, iters=None, residtol=None, verbose=False, verbose_indent=4, quantiles_losses=[0, 1, 5, 10, 25, 40, 50, 60, 75, 90, 95, 99, 100], verbose_quantiles_losses=[5, 25, 50, 75, 90], verbose_times=True, warn=True, store_data_iters=False, store_all_data=False)

Preconditioned conjugate gradient algorithm for solving linear systems \(AX=B\) where \(A\) is positive semi-definite

Parameters:

Name Type Description Default
A Union[Tensor, callable]

Matrix A with shape (...,n,n), or
callable(A) where A(X) should return the batch matrix multiplication of A and X,

required
B Tensor

Right hand side tensor \(B\) with shape (...,n,k)

 required
X0 Tensor

Initial guess for \(X\) with shape (...,n,k), defaults to zeros.

None
precond_solver Union[Tensor, callable]

Preconditioning solver, either a matrix M with shape (...,n,n), or callable(M) where M(X) should return the batch matrix multiplication of M and X

 None
iters int

number of minres iterations, defaults to 5n.

 None
residtol float

Non-negative tolerance on the maximum residual for early stopping, defaults to 1e-12 for torch.float64 and 2.5e-4 for torch.float32.

 None
verbose int

Controls logging verbosity

  • If True, perform logging.
  • If a positive int, only log every verbose iterations.
  • If None, set to a reasonable positive int based on the maximum number of iterations
  • If False, don't log.
False
verbose_indent int

Non-negative number of indentation spaces for logging.

 4
quantiles_losses list

Loss quantiles to record.

 [0, 1, 5, 10, 25, 40, 50, 60, 75, 90, 95, 99, 100]
verbose_quantiles_losses list

Loss quantiles to show in verbose log.

 [5, 25, 50, 75, 90]
verbose_times bool

If False, do not show the times in the verbose log. This is mostly for testing where timing is not reproducible.

True
warn bool

If False, then suppress warnings.

 True
store_data_iters int

Controls storage iterations with the same options as verbose. If store_data_iters==0, then the data is not collected or returned.

  • If True, store every iteration.
  • If a positive int, only store every store_data_iters iterations.
  • If None, set to a reasonable positive int based on the maximum number of iterations
  • If False, don't store data, and do not return data
False
store_all_data bool

If True, store the x values as well as the metrics.

False

Returns:

Name Type Description
x Tensor

Optimized \(X\).
data dict

Iteration data, only returned when store_data_iters>0

Examples:

>>> torch.set_default_dtype(torch.float64)
>>> rng = torch.Generator().manual_seed(7)

Real SPD example with column vector \(b\) 

>>> n = 5
>>> L = torch.rand(n,n,generator=rng).tril()
>>> A = L@L.T
>>> b = torch.rand(n,generator=rng)
>>> x_true = torch.linalg.solve(A,b[...,None])[...,0]
>>> x_true
tensor([3445.9564, -777.4224, 3706.6588, -663.9501,   23.4735])
>>> torch.allclose(A@x_true-b,torch.zeros_like(b))
True
>>> x_pcg = pcg(A,b[...,None],verbose=None,verbose_times=False)[...,0]
    iter i     | losses_quantiles                                          
               | 5         | 25        | 50        | 75        | 90        
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
    1          | 9.0e-01   | 9.0e-01   | 9.0e-01   | 9.0e-01   | 9.0e-01   
    2          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
    3          | 3.7e+00   | 3.7e+00   | 3.7e+00   | 3.7e+00   | 3.7e+00   
    4          | 1.5e+00   | 1.5e+00   | 1.5e+00   | 1.5e+00   | 1.5e+00   
    5          | 4.0e-08   | 4.0e-08   | 4.0e-08   | 4.0e-08   | 4.0e-08   
    6          | 1.8e-12   | 1.8e-12   | 1.8e-12   | 1.8e-12   | 1.8e-12   
    7          | 3.2e-13   | 3.2e-13   | 3.2e-13   | 3.2e-13   | 3.2e-13   
>>> torch.allclose(x_pcg,x_true)
True

Complex SPD example with column vector \(b\) 

>>> n = 5
>>> L = torch.randn(n,n,generator=rng,dtype=torch.complex128).tril(-1)+torch.rand(n,generator=rng).diag()
>>> A = L@L.adjoint()
>>> b = torch.rand(n,dtype=torch.complex128,generator=rng)
>>> x_true = torch.linalg.solve(A,b[...,None])[...,0]
>>> x_true
tensor([ 139.4855-450.7870j,  221.0729+203.4796j, -220.4211+69.1949j,
         107.2952+6.4764j,    4.0780-7.4403j])
>>> torch.allclose(A@x_true-b,torch.zeros_like(b))
True
>>> x_pcg = pcg(A,b[...,None],verbose=None,verbose_times=False)[...,0]
    iter i     | losses_quantiles                                          
               | 5         | 25        | 50        | 75        | 90        
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
    1          | 1.4e+00   | 1.4e+00   | 1.4e+00   | 1.4e+00   | 1.4e+00   
    2          | 1.2e+00   | 1.2e+00   | 1.2e+00   | 1.2e+00   | 1.2e+00   
    3          | 9.4e-01   | 9.4e-01   | 9.4e-01   | 9.4e-01   | 9.4e-01   
    4          | 6.5e-02   | 6.5e-02   | 6.5e-02   | 6.5e-02   | 6.5e-02   
    5          | 1.9e-08   | 1.9e-08   | 1.9e-08   | 1.9e-08   | 1.9e-08   
    6          | 2.7e-10   | 2.7e-10   | 2.7e-10   | 2.7e-10   | 2.7e-10   
    7          | 1.1e-11   | 1.1e-11   | 1.1e-11   | 1.1e-11   | 1.1e-11   
    8          | 6.0e-14   | 6.0e-14   | 6.0e-14   | 6.0e-14   | 6.0e-14   
>>> torch.allclose(x_pcg,x_true)
True

Matrix \(B\)

>>> n = 5
>>> k = 3
>>> L = torch.rand(n,n,generator=rng).tril()
>>> A = L@L.T
>>> B = torch.rand(n,k,generator=rng)
>>> X_true = torch.linalg.solve(A,B)
>>> X_true
tensor([[-4.7668e+03, -1.6571e+03, -1.2753e+03],
        [-2.7778e+03, -9.6609e+02, -7.4529e+02],
        [ 1.3876e+04,  4.8280e+03,  3.7223e+03],
        [ 1.4911e+00,  3.0229e-01, -8.1048e-01],
        [-4.1842e+01, -1.4783e+01, -1.1143e+01]])
>>> torch.allclose(A@X_true-B,torch.zeros_like(B))
True
>>> X_pcg = pcg(A,B,verbose=None,verbose_times=False)
    iter i     | losses_quantiles                                          
               | 5         | 25        | 50        | 75        | 90        
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
    1          | 3.7e-01   | 4.2e-01   | 4.8e-01   | 5.2e-01   | 5.4e-01   
    2          | 2.3e-01   | 3.1e-01   | 4.2e-01   | 8.5e-01   | 1.1e+00   
    3          | 2.1e-01   | 2.1e-01   | 2.2e-01   | 2.7e+00   | 4.1e+00   
    4          | 1.7e-01   | 2.8e-01   | 4.1e-01   | 3.7e+00   | 5.7e+00   
    5          | 2.0e-10   | 2.4e-10   | 2.9e-10   | 4.0e-10   | 4.7e-10   
    6          | 5.8e-13   | 1.5e-12   | 2.7e-12   | 4.5e-12   | 5.5e-12   
    7          | 2.4e-13   | 2.4e-13   | 2.5e-13   | 2.8e-12   | 4.4e-12   
    8          | 1.0e-13   | 1.5e-13   | 2.2e-13   | 1.9e-12   | 2.9e-12   
    9          | 1.6e-13   | 1.9e-13   | 2.3e-13   | 5.5e-13   | 7.5e-13   
>>> torch.allclose(X_pcg,X_true)
True

Low-rank plus diagonal \(A\) with storage-saving multiplication function 

>>> n = 5
>>> k = 3
>>> Q = torch.randn(n,2,generator=rng)
>>> delta = torch.rand(n,generator=rng)
>>> A = Q@Q.T+delta.diag()
>>> A
tensor([[ 3.7335,  2.5416,  0.4532,  1.8471,  1.2511],
        [ 2.5416,  2.5374,  0.3233,  1.5515,  0.6545],
        [ 0.4532,  0.3233,  0.2774,  0.2474,  0.1450],
        [ 1.8471,  1.5515,  0.2474,  3.7532, -1.1804],
        [ 1.2511,  0.6545,  0.1450, -1.1804,  3.1788]])
>>> B = torch.rand(n,k,generator=rng)
>>> X_true = torch.linalg.solve(A,B)
>>> X_true
tensor([[ 0.0859, -0.9122,  0.5310],
        [-0.1233,  0.7115, -0.7571],
        [ 0.3733,  1.3716,  1.3474],
        [ 0.0620,  0.3954,  0.0723],
        [ 0.1599,  0.3084,  0.1209]])
>>> torch.allclose(A@X_true-B,torch.zeros_like(B))
True
>>> def A_mult(x):
...     return Q@(Q.T@x)+delta[:,None]*x
>>> torch.allclose(A_mult(X_true),A@X_true)
True
>>> X_pcg = pcg(A_mult,B,verbose=None,verbose_times=False)
    iter i     | losses_quantiles                                          
               | 5         | 25        | 50        | 75        | 90        
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
    1          | 3.4e-01   | 4.4e-01   | 5.7e-01   | 5.7e-01   | 5.8e-01   
    2          | 2.6e-01   | 4.4e-01   | 6.6e-01   | 7.0e-01   | 7.3e-01   
    3          | 8.9e-02   | 1.2e-01   | 1.6e-01   | 1.7e-01   | 1.8e-01   
    4          | 3.3e-02   | 5.3e-02   | 7.9e-02   | 8.7e-02   | 9.2e-02   
    5          | 1.1e-14   | 1.3e-14   | 1.5e-14   | 2.4e-14   | 3.0e-14   
>>> torch.allclose(X_pcg,X_true)
True

Batched low-rank plus diagonal \(A\) with storage-saving multiplication function 

>>> n = 100
>>> k = 3
>>> Q = torch.randn(2,1,4,n,50,generator=rng)
>>> delta = torch.rand(2,1,4,n,generator=rng)
>>> A = torch.einsum("...ij,...kj->...ik",Q,Q)+delta[...,None]*torch.eye(n)
>>> B = torch.rand(2,6,1,n,k,generator=rng)
>>> X_true = torch.linalg.solve(A,B)
>>> torch.allclose(torch.einsum("...ij,...jk->...ik",A,X_true)-B,torch.zeros_like(B))
True
>>> def A_mult(x):
...     return torch.einsum("...ij,...jk->...ik",Q,torch.einsum("...ji,...jk->...ik",Q,x))+delta[...,None]*x
>>> torch.allclose(A_mult(X_true),torch.einsum("...ij,...jk->...ik",A,X_true))
True
>>> X_pcg,data = pcg(A_mult,B,verbose=None,iters=200,verbose_times=False,store_data_iters=None,store_all_data=True)
    iter i     | losses_quantiles                                          
               | 5         | 25        | 50        | 75        | 90        
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
    10         | 8.1e-01   | 1.1e+00   | 1.4e+00   | 1.7e+00   | 1.8e+00   
    20         | 7.0e-01   | 9.1e-01   | 1.1e+00   | 1.3e+00   | 1.4e+00   
    30         | 3.4e-01   | 4.2e-01   | 5.1e-01   | 6.0e-01   | 7.1e-01   
    ...   
>>> X_pcg.shape
torch.Size([2, 6, 4, 100, 3])
>>> torch.allclose(X_pcg,X_true)
True
>>> print_data_signatures(data)
    data['x'].shape = (2, 6, 4, 100, 3)
    data['iterrange'].shape = (201,)
    data['times'].shape = (201,)
    data['losses_quantiles']
        data['losses_quantiles']['0'].shape = (201,)
        data['losses_quantiles']['1'].shape = (201,)
        data['losses_quantiles']['5'].shape = (201,)
        data['losses_quantiles']['10'].shape = (201,)
        data['losses_quantiles']['25'].shape = (201,)
        data['losses_quantiles']['40'].shape = (201,)
        data['losses_quantiles']['50'].shape = (201,)
        data['losses_quantiles']['60'].shape = (201,)
        data['losses_quantiles']['75'].shape = (201,)
        data['losses_quantiles']['90'].shape = (201,)
        data['losses_quantiles']['95'].shape = (201,)
        data['losses_quantiles']['99'].shape = (201,)
        data['losses_quantiles']['100'].shape = (201,)
    data['xs'].shape = (201, 2, 6, 4, 100, 3)
    data['losses'].shape = (201, 2, 6, 4, 3)
Source code in agsutil/algos.py 
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
def pcg(
        A,
        B,
        X0 = None,
        precond_solver = None,
        iters = None,
        residtol = None,
        verbose = False, 
        verbose_indent = 4,
        quantiles_losses = [0,1,5,10,25,40,50,60,75,90,95,99,100],
        verbose_quantiles_losses = [5,25,50,75,90],
        verbose_times = True, 
        warn = True,
        store_data_iters = False, 
        store_all_data = False,
        ):
    r"""
    [Preconditioned conjugate gradient algorithm](https://en.wikipedia.org/wiki/Conjugate_gradient_method#The_preconditioned_conjugate_gradient_method) for solving linear systems $AX=B$ where $A$ is positive semi-definite

    Args:
        A (Union[torch.Tensor,callable]): Matrix `A` with shape `(...,n,n)`, or  
            `callable(A)` where `A(X)` should return the batch matrix multiplication of `A` and `X`,  
        B (torch.Tensor): Right hand side tensor $B$ with shape `(...,n,k)`
        X0 (torch.Tensor): Initial guess for $X$ with shape `(...,n,k)`, defaults to zeros. 
        precond_solver (Union[torch.Tensor,callable]): Preconditioning solver, either a matrix `M` with shape `(...,n,n)`, or 
            `callable(M)` where `M(X)` should return the batch matrix multiplication of `M` and `X`
        iters (int): number of minres iterations, defaults to `5n`.
        residtol (float): Non-negative tolerance on the maximum residual for early stopping, defaults to `1e-12` for `torch.float64` and `2.5e-4` for `torch.float32`.
        verbose (int): Controls logging verbosity

            - If `True`, perform logging. 
            - If a positive int, only log every `verbose` iterations. 
            - If `None`, set to a reasonable positive int based on the maximum number of iterations
            - If `False`, don't log. 

        verbose_indent (int): Non-negative number of indentation spaces for logging.
        quantiles_losses (list): Loss quantiles to record.
        verbose_quantiles_losses (list): Loss quantiles to show in verbose log.
        verbose_times (bool): If `False`, do not show the times in the verbose log. This is mostly for testing where timing is not reproducible. 
        warn (bool): If `False`, then suppress warnings.
        store_data_iters (int): Controls storage iterations with the same options as verbose. If `store_data_iters==0`, then the data is not collected or returned. 

            - If `True`, store every iteration. 
            - If a positive int, only store every `store_data_iters` iterations. 
            - If `None`, set to a reasonable positive int based on the maximum number of iterations
            - If `False`, don't store data, and do not return data 

        store_all_data (bool): If `True`, store the `x` values as well as the metrics. 

    Returns:
        x (torch.Tensor): Optimized $X$.
        data (dict): Iteration data, only returned when `store_data_iters>0`

    Examples:

        >>> torch.set_default_dtype(torch.float64)
        >>> rng = torch.Generator().manual_seed(7)

    Real SPD example with column vector $b$ 

        >>> n = 5
        >>> L = torch.rand(n,n,generator=rng).tril()
        >>> A = L@L.T
        >>> b = torch.rand(n,generator=rng)
        >>> x_true = torch.linalg.solve(A,b[...,None])[...,0]
        >>> x_true
        tensor([3445.9564, -777.4224, 3706.6588, -663.9501,   23.4735])
        >>> torch.allclose(A@x_true-b,torch.zeros_like(b))
        True
        >>> x_pcg = pcg(A,b[...,None],verbose=None,verbose_times=False)[...,0]
            iter i     | losses_quantiles                                          
                       | 5         | 25        | 50        | 75        | 90        
            ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
            1          | 9.0e-01   | 9.0e-01   | 9.0e-01   | 9.0e-01   | 9.0e-01   
            2          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
            3          | 3.7e+00   | 3.7e+00   | 3.7e+00   | 3.7e+00   | 3.7e+00   
            4          | 1.5e+00   | 1.5e+00   | 1.5e+00   | 1.5e+00   | 1.5e+00   
            5          | 4.0e-08   | 4.0e-08   | 4.0e-08   | 4.0e-08   | 4.0e-08   
            6          | 1.8e-12   | 1.8e-12   | 1.8e-12   | 1.8e-12   | 1.8e-12   
            7          | 3.2e-13   | 3.2e-13   | 3.2e-13   | 3.2e-13   | 3.2e-13   
        >>> torch.allclose(x_pcg,x_true)
        True

    Complex SPD example with column vector $b$ 

        >>> n = 5
        >>> L = torch.randn(n,n,generator=rng,dtype=torch.complex128).tril(-1)+torch.rand(n,generator=rng).diag()
        >>> A = L@L.adjoint()
        >>> b = torch.rand(n,dtype=torch.complex128,generator=rng)
        >>> x_true = torch.linalg.solve(A,b[...,None])[...,0]
        >>> x_true
        tensor([ 139.4855-450.7870j,  221.0729+203.4796j, -220.4211+69.1949j,
                 107.2952+6.4764j,    4.0780-7.4403j])
        >>> torch.allclose(A@x_true-b,torch.zeros_like(b))
        True
        >>> x_pcg = pcg(A,b[...,None],verbose=None,verbose_times=False)[...,0]
            iter i     | losses_quantiles                                          
                       | 5         | 25        | 50        | 75        | 90        
            ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
            1          | 1.4e+00   | 1.4e+00   | 1.4e+00   | 1.4e+00   | 1.4e+00   
            2          | 1.2e+00   | 1.2e+00   | 1.2e+00   | 1.2e+00   | 1.2e+00   
            3          | 9.4e-01   | 9.4e-01   | 9.4e-01   | 9.4e-01   | 9.4e-01   
            4          | 6.5e-02   | 6.5e-02   | 6.5e-02   | 6.5e-02   | 6.5e-02   
            5          | 1.9e-08   | 1.9e-08   | 1.9e-08   | 1.9e-08   | 1.9e-08   
            6          | 2.7e-10   | 2.7e-10   | 2.7e-10   | 2.7e-10   | 2.7e-10   
            7          | 1.1e-11   | 1.1e-11   | 1.1e-11   | 1.1e-11   | 1.1e-11   
            8          | 6.0e-14   | 6.0e-14   | 6.0e-14   | 6.0e-14   | 6.0e-14   
        >>> torch.allclose(x_pcg,x_true)
        True

    Matrix $B$

        >>> n = 5
        >>> k = 3
        >>> L = torch.rand(n,n,generator=rng).tril()
        >>> A = L@L.T
        >>> B = torch.rand(n,k,generator=rng)
        >>> X_true = torch.linalg.solve(A,B)
        >>> X_true
        tensor([[-4.7668e+03, -1.6571e+03, -1.2753e+03],
                [-2.7778e+03, -9.6609e+02, -7.4529e+02],
                [ 1.3876e+04,  4.8280e+03,  3.7223e+03],
                [ 1.4911e+00,  3.0229e-01, -8.1048e-01],
                [-4.1842e+01, -1.4783e+01, -1.1143e+01]])
        >>> torch.allclose(A@X_true-B,torch.zeros_like(B))
        True
        >>> X_pcg = pcg(A,B,verbose=None,verbose_times=False)
            iter i     | losses_quantiles                                          
                       | 5         | 25        | 50        | 75        | 90        
            ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
            1          | 3.7e-01   | 4.2e-01   | 4.8e-01   | 5.2e-01   | 5.4e-01   
            2          | 2.3e-01   | 3.1e-01   | 4.2e-01   | 8.5e-01   | 1.1e+00   
            3          | 2.1e-01   | 2.1e-01   | 2.2e-01   | 2.7e+00   | 4.1e+00   
            4          | 1.7e-01   | 2.8e-01   | 4.1e-01   | 3.7e+00   | 5.7e+00   
            5          | 2.0e-10   | 2.4e-10   | 2.9e-10   | 4.0e-10   | 4.7e-10   
            6          | 5.8e-13   | 1.5e-12   | 2.7e-12   | 4.5e-12   | 5.5e-12   
            7          | 2.4e-13   | 2.4e-13   | 2.5e-13   | 2.8e-12   | 4.4e-12   
            8          | 1.0e-13   | 1.5e-13   | 2.2e-13   | 1.9e-12   | 2.9e-12   
            9          | 1.6e-13   | 1.9e-13   | 2.3e-13   | 5.5e-13   | 7.5e-13   
        >>> torch.allclose(X_pcg,X_true)
        True

    Low-rank plus diagonal $A$ with storage-saving multiplication function 

        >>> n = 5
        >>> k = 3
        >>> Q = torch.randn(n,2,generator=rng)
        >>> delta = torch.rand(n,generator=rng)
        >>> A = Q@Q.T+delta.diag()
        >>> A
        tensor([[ 3.7335,  2.5416,  0.4532,  1.8471,  1.2511],
                [ 2.5416,  2.5374,  0.3233,  1.5515,  0.6545],
                [ 0.4532,  0.3233,  0.2774,  0.2474,  0.1450],
                [ 1.8471,  1.5515,  0.2474,  3.7532, -1.1804],
                [ 1.2511,  0.6545,  0.1450, -1.1804,  3.1788]])
        >>> B = torch.rand(n,k,generator=rng)
        >>> X_true = torch.linalg.solve(A,B)
        >>> X_true
        tensor([[ 0.0859, -0.9122,  0.5310],
                [-0.1233,  0.7115, -0.7571],
                [ 0.3733,  1.3716,  1.3474],
                [ 0.0620,  0.3954,  0.0723],
                [ 0.1599,  0.3084,  0.1209]])
        >>> torch.allclose(A@X_true-B,torch.zeros_like(B))
        True
        >>> def A_mult(x):
        ...     return Q@(Q.T@x)+delta[:,None]*x
        >>> torch.allclose(A_mult(X_true),A@X_true)
        True
        >>> X_pcg = pcg(A_mult,B,verbose=None,verbose_times=False)
            iter i     | losses_quantiles                                          
                       | 5         | 25        | 50        | 75        | 90        
            ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
            1          | 3.4e-01   | 4.4e-01   | 5.7e-01   | 5.7e-01   | 5.8e-01   
            2          | 2.6e-01   | 4.4e-01   | 6.6e-01   | 7.0e-01   | 7.3e-01   
            3          | 8.9e-02   | 1.2e-01   | 1.6e-01   | 1.7e-01   | 1.8e-01   
            4          | 3.3e-02   | 5.3e-02   | 7.9e-02   | 8.7e-02   | 9.2e-02   
            5          | 1.1e-14   | 1.3e-14   | 1.5e-14   | 2.4e-14   | 3.0e-14   
        >>> torch.allclose(X_pcg,X_true)
        True

    Batched low-rank plus diagonal $A$ with storage-saving multiplication function 

        >>> n = 100
        >>> k = 3
        >>> Q = torch.randn(2,1,4,n,50,generator=rng)
        >>> delta = torch.rand(2,1,4,n,generator=rng)
        >>> A = torch.einsum("...ij,...kj->...ik",Q,Q)+delta[...,None]*torch.eye(n)
        >>> B = torch.rand(2,6,1,n,k,generator=rng)
        >>> X_true = torch.linalg.solve(A,B)
        >>> torch.allclose(torch.einsum("...ij,...jk->...ik",A,X_true)-B,torch.zeros_like(B))
        True
        >>> def A_mult(x):
        ...     return torch.einsum("...ij,...jk->...ik",Q,torch.einsum("...ji,...jk->...ik",Q,x))+delta[...,None]*x
        >>> torch.allclose(A_mult(X_true),torch.einsum("...ij,...jk->...ik",A,X_true))
        True
        >>> X_pcg,data = pcg(A_mult,B,verbose=None,iters=200,verbose_times=False,store_data_iters=None,store_all_data=True)
            iter i     | losses_quantiles                                          
                       | 5         | 25        | 50        | 75        | 90        
            ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
            10         | 8.1e-01   | 1.1e+00   | 1.4e+00   | 1.7e+00   | 1.8e+00   
            20         | 7.0e-01   | 9.1e-01   | 1.1e+00   | 1.3e+00   | 1.4e+00   
            30         | 3.4e-01   | 4.2e-01   | 5.1e-01   | 6.0e-01   | 7.1e-01   
            ...   
        >>> X_pcg.shape
        torch.Size([2, 6, 4, 100, 3])
        >>> torch.allclose(X_pcg,X_true)
        True
        >>> print_data_signatures(data)
            data['x'].shape = (2, 6, 4, 100, 3)
            data['iterrange'].shape = (201,)
            data['times'].shape = (201,)
            data['losses_quantiles']
                data['losses_quantiles']['0'].shape = (201,)
                data['losses_quantiles']['1'].shape = (201,)
                data['losses_quantiles']['5'].shape = (201,)
                data['losses_quantiles']['10'].shape = (201,)
                data['losses_quantiles']['25'].shape = (201,)
                data['losses_quantiles']['40'].shape = (201,)
                data['losses_quantiles']['50'].shape = (201,)
                data['losses_quantiles']['60'].shape = (201,)
                data['losses_quantiles']['75'].shape = (201,)
                data['losses_quantiles']['90'].shape = (201,)
                data['losses_quantiles']['95'].shape = (201,)
                data['losses_quantiles']['99'].shape = (201,)
                data['losses_quantiles']['100'].shape = (201,)
            data['xs'].shape = (201, 2, 6, 4, 100, 3)
            data['losses'].shape = (201, 2, 6, 4, 3)
    """
    if warn and (not torch.get_default_dtype()==torch.float64): warnings.warn('''
            torch.get_default_dtype() = %s, but pcg often requires high precision updates. We recommend using:
                torch.set_default_dtype(torch.float64)'''%str(torch.get_default_dtype()))
    assert torch.get_default_dtype() in [torch.float32,torch.float64]
    default_dtype = torch.get_default_dtype()
    device = str(B.device)
    default_device = str(torch.get_default_device())
    assert B.ndim>=2, "B should have shape (...,n,k)"
    n = B.size(-2)
    k = B.size(-1)
    if X0 is None: 
        X0 = torch.zeros_like(B)
    if isinstance(A,torch.Tensor):
        assert A.shape[-2:]==(n,n)
        assert torch.allclose(A.adjoint(),A)
        matvec = lambda X: torch.einsum("...ij,...jk->...ik",A,X)
    else:
        assert callable(A)
        matvec = A
    if precond_solver is None: 
        psolve = lambda X: X # TODO: implement more involved preconditioned solver
    elif isinstance(precond_solver,torch.Tensor):
        assert precond_solver.shape[-2:]==(n,n)
        assert torch.allclose(precond_solver.adjoint(),precond_solver)
        psolve = lambda X: torch.einsum("...ij,...jk->...ik",precond_solver,X)
    else:
        assert callable(precond_solver)
        psolve = precond_solver
    if iters is None:
        iters = 5*n
    assert iters>=0
    assert iters%1==0
    if residtol is None: 
        if default_dtype==torch.float64:
            residtol = 1e-12
        elif default_dtype==torch.float32:
            residtol = 2.5e-4
        else:
            raise Exception("default_dtype = %s not parsed"%str(default_dtype))
    assert residtol>=0
    if verbose is None: 
        verbose = max(1,iters//20)
    assert verbose%1==0
    assert verbose>=0 
    if store_data_iters is None: 
        store_data_iters = max(1,iters//1000)
    assert store_data_iters%1==0
    assert store_data_iters>=0 
    assert isinstance(store_all_data,bool)
    assert isinstance(quantiles_losses,list)
    assert all(0<=qt<=100 for qt in quantiles_losses)
    assert isinstance(verbose_quantiles_losses,list)
    assert all(qt in quantiles_losses for qt in verbose_quantiles_losses)
    assert verbose_indent%1==0 
    assert verbose_indent>=0
    assert isinstance(verbose_times,bool)
    if verbose:
        _h_iter = "%-10s "%"iter i"
        _h_times = "| %-10s"%"times" if verbose_times else ""
        _s_losses_qt = ("| %-9s "*len(verbose_quantiles_losses))%tuple(str(qt) for qt in verbose_quantiles_losses)
        _h_losses_qt = "| losses_quantiles"+" "*(len(_s_losses_qt)-len("| losses_quantiles"))
        _h = _h_iter+_h_losses_qt+_h_times
        _s = " "*len(_h_iter)+_s_losses_qt+("|"+" "*(len(_h_times)-1) if verbose_times else " "*len(_h_times))
        print(" "*verbose_indent+_h)
        print(" "*verbose_indent+_s)
        print(" "*verbose_indent+"~"*len(_s))
    timer = Timer(device=device)
    timer.tic()
    inner = lambda a,b: torch.einsum("...ij,...ij->...j",a.conj(),b) # TODO: do we need this? 
    Anorm = 0
    x = X0 
    Ax = matvec(x)
    bnorm = torch.linalg.norm(B,dim=-2) # (...,k)
    assert Ax.shape[-2:]==(n,k)
    batch_shape = tuple(Ax.shape[:-2])
    if store_data_iters:
        iterrange = []
        times = []
        losses = []
        losses_quantiles = {str(qt):[] for qt in quantiles_losses}
        if store_all_data:
            xs = []
    for i in range(iters+1):
        resid = matvec(x)-B 
        breakcond = i==iters or resid.abs().amax()<=residtol
        loss = torch.linalg.norm(resid,dim=-2)/bnorm
        times_i = timer.toc()
        losses_quantiles_i = {str(qt): loss.nanquantile(qt/100) for qt in quantiles_losses}
        if store_data_iters and (i%store_data_iters==0 or breakcond):
            iterrange.append(i)
            losses.append(loss.to(default_device))
            times.append(times_i)
            for qt in quantiles_losses:
                losses_quantiles[str(qt)].append(losses_quantiles_i[str(qt)].to(default_device))
            if store_all_data:
                xs.append(x.expand(resid.shape).to(default_device))
        if verbose and (i%verbose==0 or breakcond):
            _s_iter = "%-10d "%i
            _s_losses_qt = ("| %-9.1e "*len(verbose_quantiles_losses))%tuple(losses_quantiles_i[str(qt)] for qt in verbose_quantiles_losses)
            _s_times = "| %-10.1f "%(times_i) if verbose_times else ""
            print(" "*verbose_indent+_s_iter+_s_losses_qt+_s_times)
        if breakcond: break 
        if i==0:
            r = B-Ax # (...,n,k)
            z = psolve(r) # (...,n,k)
            p = z
            rz = inner(r,z)
        else:
            z = psolve(r)
            rznew = inner(r,z)
            beta = rznew/rz
            rz = rznew 
            p = z+beta[...,None,:]*p
        Ap = matvec(p)
        alpha = rz/inner(p,Ap)
        x = x+alpha[...,None,:]*p
        r = r-alpha[...,None,:]*Ap
    if store_data_iters==0:
        return x 
    else:
        data = {
            "x": x.to(default_device), 
            "iterrange": torch.tensor(iterrange,dtype=int), 
            "times": torch.tensor(times), 
            "losses_quantiles": {str(qt):torch.tensor(losses_quantiles[str(qt)]) for qt in quantiles_losses},
            }
        if store_all_data:
            data["xs"] = torch.stack(xs,dim=0)
            data["losses"] = torch.stack(losses,dim=0)
        return x,data

minres 

minres(A, B, X0=None, iters=None, residtol=None, verbose=False, verbose_indent=4, quantiles_losses=[0, 1, 5, 10, 25, 40, 50, 60, 75, 90, 95, 99, 100], verbose_quantiles_losses=[5, 25, 50, 75, 90], verbose_times=True, warn=True, store_data_iters=False, store_all_data=False)

MINRES algorithm for solving linear systems \(AX=B\) where \(A\) is real-symmetric or complex-Hermitian

A translation of scipy.sparse.linalg.minres.

Parameters:

Name Type Description Default
A Union[Tensor, callable]

Symmetric matrix A with shape (...,n,n), or
callable(A) where A(X) should return the batch matrix multiplication of A and X,

required
B Tensor

Right hand side tensor \(B\) with shape (...,n,k)

 required
X0 Tensor

Initial guess for \(X\) with shape (...,n,k), defaults to zeros.

None
iters int

number of minres iterations, defaults to 5n.

None
residtol float

Non-negative tolerance on the maximum residual for early stopping, defaults to 1e-12 for torch.float64 and 2.5e-4 for torch.float32.

 None
verbose int

Controls logging verbosity

  • If True, perform logging.
  • If a positive int, only log every verbose iterations.
  • If None, set to a reasonable positive int based on the maximum number of iterations
  • If False, don't log.
False
verbose_indent int

Non-negative number of indentation spaces for logging.

 4
quantiles_losses list

Loss quantiles to record.

 [0, 1, 5, 10, 25, 40, 50, 60, 75, 90, 95, 99, 100]
verbose_quantiles_losses list

Loss quantiles to show in verbose log.

 [5, 25, 50, 75, 90]
verbose_times bool

If False, do not show the times in the verbose log. This is mostly for testing where timing is not reproducible.

True
warn bool

If False, then suppress warnings.

 True
store_data_iters int

Controls storage iterations with the same options as verbose. If store_data_iters==0, then the data is not collected or returned.

  • If True, store every iteration.
  • If a positive int, only store every store_data_iters iterations.
  • If None, set to a reasonable positive int based on the maximum number of iterations
  • If False, don't store data, and do not return data
False
store_all_data bool

If True, store the x values as well as the metrics.

False

Returns:

Name Type Description
x Tensor

Optimized \(X\).
data dict

Iteration data, only returned when store_data_iters>0

Examples:

>>> torch.set_default_dtype(torch.float64)
>>> rng = torch.Generator().manual_seed(7)

Real-symmetric example with column vector \(b\) 

>>> n = 5
>>> A = torch.randn(n,n,generator=rng)
>>> A = (A+A.T)/2
>>> b = torch.rand(n,generator=rng)
>>> x_true = torch.linalg.solve(A,b[...,None])[...,0]
>>> x_true
tensor([-0.1402,  0.4565,  0.2920,  0.2470,  0.3251])
>>> torch.allclose(A@x_true-b,torch.zeros_like(b))
True
>>> x_minres = minres(A,b[...,None],verbose=None,verbose_times=False)[...,0]
    iter i     | losses_quantiles                                          
               | 5         | 25        | 50        | 75        | 90        
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
    1          | 6.1e-01   | 6.1e-01   | 6.1e-01   | 6.1e-01   | 6.1e-01   
    2          | 2.1e-01   | 2.1e-01   | 2.1e-01   | 2.1e-01   | 2.1e-01   
    3          | 1.9e-01   | 1.9e-01   | 1.9e-01   | 1.9e-01   | 1.9e-01   
    4          | 7.3e-02   | 7.3e-02   | 7.3e-02   | 7.3e-02   | 7.3e-02   
    5          | 5.4e-16   | 5.4e-16   | 5.4e-16   | 5.4e-16   | 5.4e-16   
>>> torch.allclose(x_minres,x_true)
True

Complex-Hermitian example with column vector \(b\) 

>>> n = 5
>>> A = torch.randn(n,n,dtype=torch.complex128,generator=rng)
>>> A = (A+A.adjoint())/2
>>> b = torch.rand(n,dtype=torch.complex128,generator=rng)
>>> x_true = torch.linalg.solve(A,b[...,None])[...,0]
>>> x_true
tensor([ 0.2207+0.2879j,  0.0928-0.0057j,  0.2681+1.3488j, -1.2520-0.4214j,
        -0.8860-0.6922j])
>>> torch.allclose(A@x_true-b,torch.zeros_like(b))
True
>>> x_minres = minres(A,b[...,None],verbose=None,verbose_times=False)[...,0]
    iter i     | losses_quantiles                                          
               | 5         | 25        | 50        | 75        | 90        
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
    1          | 8.7e-01   | 8.7e-01   | 8.7e-01   | 8.7e-01   | 8.7e-01   
    2          | 5.4e-01   | 5.4e-01   | 5.4e-01   | 5.4e-01   | 5.4e-01   
    3          | 5.4e-01   | 5.4e-01   | 5.4e-01   | 5.4e-01   | 5.4e-01   
    4          | 2.6e-01   | 2.6e-01   | 2.6e-01   | 2.6e-01   | 2.6e-01   
    5          | 3.0e-15   | 3.0e-15   | 3.0e-15   | 3.0e-15   | 3.0e-15   
>>> torch.allclose(x_minres,x_true)
True

Matrix \(B\)

>>> n = 5
>>> k = 3
>>> A = torch.randn(n,n,generator=rng)
>>> A = (A+A.T)/2
>>> B = torch.rand(n,k,generator=rng)
>>> X_true = torch.linalg.solve(A,B)
>>> X_true
tensor([[ 0.8801, -0.0116,  0.4805],
        [-1.1095, -1.6166, -0.7103],
        [-2.9918, -1.9201, -3.5855],
        [-4.1777, -3.6586, -5.1658],
        [ 1.5417,  0.9814,  1.3790]])
>>> torch.allclose(A@X_true-B,torch.zeros_like(B))
True
>>> X_minres = minres(A,B,verbose=None,verbose_times=False)
    iter i     | losses_quantiles                                          
               | 5         | 25        | 50        | 75        | 90        
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
    1          | 6.6e-01   | 7.5e-01   | 8.7e-01   | 8.8e-01   | 8.9e-01   
    2          | 6.2e-01   | 6.4e-01   | 6.6e-01   | 7.3e-01   | 7.6e-01   
    3          | 3.5e-01   | 4.7e-01   | 6.1e-01   | 6.3e-01   | 6.4e-01   
    4          | 1.5e-01   | 1.8e-01   | 2.2e-01   | 3.4e-01   | 4.1e-01   
    5          | 7.0e-15   | 9.1e-15   | 1.2e-14   | 2.2e-14   | 2.8e-14   
>>> torch.allclose(X_minres,X_true)
True

Tri-diagonal \(A\) with storage-saving multiplication function 

>>> n = 5
>>> k = 3
>>> A_diag = torch.randn(n,generator=rng)
>>> A_off_diag = torch.randn(n-1,generator=rng) 
>>> A = torch.zeros(n,n)
>>> A[torch.arange(n),torch.arange(n)] = A_diag 
>>> A[torch.arange(n-1),torch.arange(1,n)] = A_off_diag
>>> A[torch.arange(1,n),torch.arange(n-1)] = A_off_diag
>>> A
tensor([[-0.2728, -0.1545,  0.0000,  0.0000,  0.0000],
        [-0.1545, -0.0275, -0.0120,  0.0000,  0.0000],
        [ 0.0000, -0.0120, -0.4436,  0.2802,  0.0000],
        [ 0.0000,  0.0000,  0.2802, -0.7303,  0.9724],
        [ 0.0000,  0.0000,  0.0000,  0.9724, -0.4180]])
>>> B = torch.rand(n,k,generator=rng)
>>> X_true = torch.linalg.solve(A,B)
>>> X_true
tensor([[-5.5996, -3.1344, -6.0989],
        [ 6.5450, -0.5211,  5.0505],
        [-1.8303, -0.9818, -0.4647],
        [ 0.5862,  0.9233,  1.2342],
        [ 1.3614,  1.4523,  1.9166]])
>>> torch.allclose(A@X_true-B,torch.zeros_like(B))
True
>>> def A_mult(x):
...     y = x*A_diag[:,None]
...     y[1:,:] += x[:-1,:]*A_off_diag[:,None]
...     y[:-1,:] += x[1:,:]*A_off_diag[:,None]
...     return y
>>> torch.allclose(A_mult(X_true),A@X_true)
True
>>> X_minres = minres(A_mult,B,verbose=None,verbose_times=False)
    iter i     | losses_quantiles                                          
               | 5         | 25        | 50        | 75        | 90        
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
    1          | 7.5e-01   | 8.1e-01   | 8.8e-01   | 9.3e-01   | 9.6e-01   
    2          | 4.5e-01   | 5.6e-01   | 7.1e-01   | 8.0e-01   | 8.5e-01   
    3          | 1.5e-01   | 1.9e-01   | 2.4e-01   | 2.9e-01   | 3.2e-01   
    4          | 5.7e-02   | 1.4e-01   | 2.4e-01   | 2.9e-01   | 3.2e-01   
    5          | 4.0e-15   | 5.4e-15   | 7.1e-15   | 1.9e-14   | 2.6e-14   
>>> torch.allclose(X_minres,X_true)
True

Batched tri-diagonal \(A\) with storage-saving multiplication function 

>>> n = 100
>>> k = 3
>>> A_diag = torch.randn(2,1,4,n,generator=rng)
>>> A_off_diag = torch.randn(2,1,4,n-1,generator=rng) 
>>> A = torch.zeros(2,1,4,n,n)
>>> A[...,torch.arange(n),torch.arange(n)] = A_diag 
>>> A[...,torch.arange(n-1),torch.arange(1,n)] = A_off_diag
>>> A[...,torch.arange(1,n),torch.arange(n-1)] = A_off_diag
>>> B = torch.rand(2,6,1,n,k,generator=rng)
>>> X_true = torch.linalg.solve(A,B)
>>> torch.allclose(torch.einsum("...ij,...jk->...ik",A,X_true)-B,torch.zeros_like(B))
True
>>> def A_mult(x):
...     y = x*A_diag[...,:,None]
...     y[...,1:,:] += x[...,:-1,:]*A_off_diag[...,:,None]
...     y[...,:-1,:] += x[...,1:,:]*A_off_diag[...,:,None]
...     return y
>>> torch.allclose(A_mult(X_true),torch.einsum("...ij,...jk->...ik",A,X_true))
True
>>> X_minres,data = minres(A_mult,B,verbose=None,verbose_times=False,store_data_iters=None,store_all_data=True)
    iter i     | losses_quantiles                                          
               | 5         | 25        | 50        | 75        | 90        
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
    25         | 1.8e-01   | 2.3e-01   | 2.6e-01   | 2.9e-01   | 3.2e-01   
    50         | 9.4e-02   | 1.2e-01   | 1.5e-01   | 1.9e-01   | 2.1e-01   
    75         | 3.3e-02   | 5.1e-02   | 8.2e-02   | 1.5e-01   | 1.8e-01   
    ...   
>>> X_minres.shape
torch.Size([2, 6, 4, 100, 3])
>>> torch.allclose(X_minres,X_true)
True
>>> print_data_signatures(data)
    data['x'].shape = (2, 6, 4, 100, 3)
    data['iterrange'].shape = (177,)
    data['times'].shape = (177,)
    data['losses_quantiles']
        data['losses_quantiles']['0'].shape = (177,)
        data['losses_quantiles']['1'].shape = (177,)
        data['losses_quantiles']['5'].shape = (177,)
        data['losses_quantiles']['10'].shape = (177,)
        data['losses_quantiles']['25'].shape = (177,)
        data['losses_quantiles']['40'].shape = (177,)
        data['losses_quantiles']['50'].shape = (177,)
        data['losses_quantiles']['60'].shape = (177,)
        data['losses_quantiles']['75'].shape = (177,)
        data['losses_quantiles']['90'].shape = (177,)
        data['losses_quantiles']['95'].shape = (177,)
        data['losses_quantiles']['99'].shape = (177,)
        data['losses_quantiles']['100'].shape = (177,)
    data['xs'].shape = (177, 2, 6, 4, 100, 3)
    data['losses'].shape = (177, 2, 6, 4, 3)
Source code in agsutil/algos.py 
 897
 898
 899
 900
 901
 902
 903
 904
 905
 906
 907
 908
 909
 910
 911
 912
 913
 914
 915
 916
 917
 918
 919
 920
 921
 922
 923
 924
 925
 926
 927
 928
 929
 930
 931
 932
 933
 934
 935
 936
 937
 938
 939
 940
 941
 942
 943
 944
 945
 946
 947
 948
 949
 950
 951
 952
 953
 954
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
def minres(
        A,
        B,
        X0 = None,
        iters = None,
        residtol = None,
        verbose = False, 
        verbose_indent = 4,
        quantiles_losses = [0,1,5,10,25,40,50,60,75,90,95,99,100],
        verbose_quantiles_losses = [5,25,50,75,90],
        verbose_times = True, 
        warn = True,
        store_data_iters = False, 
        store_all_data = False,
        ):
    r"""
    [MINRES algorithm](https://en.wikipedia.org/wiki/Minimal_residual_method) for solving linear systems $AX=B$ where $A$ is real-symmetric or complex-Hermitian

    A translation of [`scipy.sparse.linalg.minres`](https://github.com/scipy/scipy/blob/v1.17.0/scipy/sparse/linalg/_isolve/minres.py).

    Args:
        A (Union[torch.Tensor,callable]): Symmetric matrix `A` with shape `(...,n,n)`, or  
            `callable(A)` where `A(X)` should return the batch matrix multiplication of `A` and `X`,  
        B (torch.Tensor): Right hand side tensor $B$ with shape `(...,n,k)`
        X0 (torch.Tensor): Initial guess for $X$ with shape `(...,n,k)`, defaults to zeros. 
        iters (int): number of minres iterations, defaults to `5n`. 
        residtol (float): Non-negative tolerance on the maximum residual for early stopping, defaults to `1e-12` for `torch.float64` and `2.5e-4` for `torch.float32`.
        verbose (int): Controls logging verbosity

            - If `True`, perform logging. 
            - If a positive int, only log every `verbose` iterations. 
            - If `None`, set to a reasonable positive int based on the maximum number of iterations
            - If `False`, don't log. 

        verbose_indent (int): Non-negative number of indentation spaces for logging.
        quantiles_losses (list): Loss quantiles to record.
        verbose_quantiles_losses (list): Loss quantiles to show in verbose log.
        verbose_times (bool): If `False`, do not show the times in the verbose log. This is mostly for testing where timing is not reproducible. 
        warn (bool): If `False`, then suppress warnings.
        store_data_iters (int): Controls storage iterations with the same options as verbose. If `store_data_iters==0`, then the data is not collected or returned. 

            - If `True`, store every iteration. 
            - If a positive int, only store every `store_data_iters` iterations. 
            - If `None`, set to a reasonable positive int based on the maximum number of iterations
            - If `False`, don't store data, and do not return data 

        store_all_data (bool): If `True`, store the `x` values as well as the metrics. 

    Returns:
        x (torch.Tensor): Optimized $X$.
        data (dict): Iteration data, only returned when `store_data_iters>0`

    Examples:

        >>> torch.set_default_dtype(torch.float64)
        >>> rng = torch.Generator().manual_seed(7)

    Real-symmetric example with column vector $b$ 

        >>> n = 5
        >>> A = torch.randn(n,n,generator=rng)
        >>> A = (A+A.T)/2
        >>> b = torch.rand(n,generator=rng)
        >>> x_true = torch.linalg.solve(A,b[...,None])[...,0]
        >>> x_true
        tensor([-0.1402,  0.4565,  0.2920,  0.2470,  0.3251])
        >>> torch.allclose(A@x_true-b,torch.zeros_like(b))
        True
        >>> x_minres = minres(A,b[...,None],verbose=None,verbose_times=False)[...,0]
            iter i     | losses_quantiles                                          
                       | 5         | 25        | 50        | 75        | 90        
            ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
            1          | 6.1e-01   | 6.1e-01   | 6.1e-01   | 6.1e-01   | 6.1e-01   
            2          | 2.1e-01   | 2.1e-01   | 2.1e-01   | 2.1e-01   | 2.1e-01   
            3          | 1.9e-01   | 1.9e-01   | 1.9e-01   | 1.9e-01   | 1.9e-01   
            4          | 7.3e-02   | 7.3e-02   | 7.3e-02   | 7.3e-02   | 7.3e-02   
            5          | 5.4e-16   | 5.4e-16   | 5.4e-16   | 5.4e-16   | 5.4e-16   
        >>> torch.allclose(x_minres,x_true)
        True

    Complex-Hermitian example with column vector $b$ 

        >>> n = 5
        >>> A = torch.randn(n,n,dtype=torch.complex128,generator=rng)
        >>> A = (A+A.adjoint())/2
        >>> b = torch.rand(n,dtype=torch.complex128,generator=rng)
        >>> x_true = torch.linalg.solve(A,b[...,None])[...,0]
        >>> x_true
        tensor([ 0.2207+0.2879j,  0.0928-0.0057j,  0.2681+1.3488j, -1.2520-0.4214j,
                -0.8860-0.6922j])
        >>> torch.allclose(A@x_true-b,torch.zeros_like(b))
        True
        >>> x_minres = minres(A,b[...,None],verbose=None,verbose_times=False)[...,0]
            iter i     | losses_quantiles                                          
                       | 5         | 25        | 50        | 75        | 90        
            ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
            1          | 8.7e-01   | 8.7e-01   | 8.7e-01   | 8.7e-01   | 8.7e-01   
            2          | 5.4e-01   | 5.4e-01   | 5.4e-01   | 5.4e-01   | 5.4e-01   
            3          | 5.4e-01   | 5.4e-01   | 5.4e-01   | 5.4e-01   | 5.4e-01   
            4          | 2.6e-01   | 2.6e-01   | 2.6e-01   | 2.6e-01   | 2.6e-01   
            5          | 3.0e-15   | 3.0e-15   | 3.0e-15   | 3.0e-15   | 3.0e-15   
        >>> torch.allclose(x_minres,x_true)
        True

    Matrix $B$

        >>> n = 5
        >>> k = 3
        >>> A = torch.randn(n,n,generator=rng)
        >>> A = (A+A.T)/2
        >>> B = torch.rand(n,k,generator=rng)
        >>> X_true = torch.linalg.solve(A,B)
        >>> X_true
        tensor([[ 0.8801, -0.0116,  0.4805],
                [-1.1095, -1.6166, -0.7103],
                [-2.9918, -1.9201, -3.5855],
                [-4.1777, -3.6586, -5.1658],
                [ 1.5417,  0.9814,  1.3790]])
        >>> torch.allclose(A@X_true-B,torch.zeros_like(B))
        True
        >>> X_minres = minres(A,B,verbose=None,verbose_times=False)
            iter i     | losses_quantiles                                          
                       | 5         | 25        | 50        | 75        | 90        
            ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
            1          | 6.6e-01   | 7.5e-01   | 8.7e-01   | 8.8e-01   | 8.9e-01   
            2          | 6.2e-01   | 6.4e-01   | 6.6e-01   | 7.3e-01   | 7.6e-01   
            3          | 3.5e-01   | 4.7e-01   | 6.1e-01   | 6.3e-01   | 6.4e-01   
            4          | 1.5e-01   | 1.8e-01   | 2.2e-01   | 3.4e-01   | 4.1e-01   
            5          | 7.0e-15   | 9.1e-15   | 1.2e-14   | 2.2e-14   | 2.8e-14   
        >>> torch.allclose(X_minres,X_true)
        True

    Tri-diagonal $A$ with storage-saving multiplication function 

        >>> n = 5
        >>> k = 3
        >>> A_diag = torch.randn(n,generator=rng)
        >>> A_off_diag = torch.randn(n-1,generator=rng) 
        >>> A = torch.zeros(n,n)
        >>> A[torch.arange(n),torch.arange(n)] = A_diag 
        >>> A[torch.arange(n-1),torch.arange(1,n)] = A_off_diag
        >>> A[torch.arange(1,n),torch.arange(n-1)] = A_off_diag
        >>> A
        tensor([[-0.2728, -0.1545,  0.0000,  0.0000,  0.0000],
                [-0.1545, -0.0275, -0.0120,  0.0000,  0.0000],
                [ 0.0000, -0.0120, -0.4436,  0.2802,  0.0000],
                [ 0.0000,  0.0000,  0.2802, -0.7303,  0.9724],
                [ 0.0000,  0.0000,  0.0000,  0.9724, -0.4180]])
        >>> B = torch.rand(n,k,generator=rng)
        >>> X_true = torch.linalg.solve(A,B)
        >>> X_true
        tensor([[-5.5996, -3.1344, -6.0989],
                [ 6.5450, -0.5211,  5.0505],
                [-1.8303, -0.9818, -0.4647],
                [ 0.5862,  0.9233,  1.2342],
                [ 1.3614,  1.4523,  1.9166]])
        >>> torch.allclose(A@X_true-B,torch.zeros_like(B))
        True
        >>> def A_mult(x):
        ...     y = x*A_diag[:,None]
        ...     y[1:,:] += x[:-1,:]*A_off_diag[:,None]
        ...     y[:-1,:] += x[1:,:]*A_off_diag[:,None]
        ...     return y
        >>> torch.allclose(A_mult(X_true),A@X_true)
        True
        >>> X_minres = minres(A_mult,B,verbose=None,verbose_times=False)
            iter i     | losses_quantiles                                          
                       | 5         | 25        | 50        | 75        | 90        
            ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
            1          | 7.5e-01   | 8.1e-01   | 8.8e-01   | 9.3e-01   | 9.6e-01   
            2          | 4.5e-01   | 5.6e-01   | 7.1e-01   | 8.0e-01   | 8.5e-01   
            3          | 1.5e-01   | 1.9e-01   | 2.4e-01   | 2.9e-01   | 3.2e-01   
            4          | 5.7e-02   | 1.4e-01   | 2.4e-01   | 2.9e-01   | 3.2e-01   
            5          | 4.0e-15   | 5.4e-15   | 7.1e-15   | 1.9e-14   | 2.6e-14   
        >>> torch.allclose(X_minres,X_true)
        True

    Batched tri-diagonal $A$ with storage-saving multiplication function 

        >>> n = 100
        >>> k = 3
        >>> A_diag = torch.randn(2,1,4,n,generator=rng)
        >>> A_off_diag = torch.randn(2,1,4,n-1,generator=rng) 
        >>> A = torch.zeros(2,1,4,n,n)
        >>> A[...,torch.arange(n),torch.arange(n)] = A_diag 
        >>> A[...,torch.arange(n-1),torch.arange(1,n)] = A_off_diag
        >>> A[...,torch.arange(1,n),torch.arange(n-1)] = A_off_diag
        >>> B = torch.rand(2,6,1,n,k,generator=rng)
        >>> X_true = torch.linalg.solve(A,B)
        >>> torch.allclose(torch.einsum("...ij,...jk->...ik",A,X_true)-B,torch.zeros_like(B))
        True
        >>> def A_mult(x):
        ...     y = x*A_diag[...,:,None]
        ...     y[...,1:,:] += x[...,:-1,:]*A_off_diag[...,:,None]
        ...     y[...,:-1,:] += x[...,1:,:]*A_off_diag[...,:,None]
        ...     return y
        >>> torch.allclose(A_mult(X_true),torch.einsum("...ij,...jk->...ik",A,X_true))
        True
        >>> X_minres,data = minres(A_mult,B,verbose=None,verbose_times=False,store_data_iters=None,store_all_data=True)
            iter i     | losses_quantiles                                          
                       | 5         | 25        | 50        | 75        | 90        
            ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
            25         | 1.8e-01   | 2.3e-01   | 2.6e-01   | 2.9e-01   | 3.2e-01   
            50         | 9.4e-02   | 1.2e-01   | 1.5e-01   | 1.9e-01   | 2.1e-01   
            75         | 3.3e-02   | 5.1e-02   | 8.2e-02   | 1.5e-01   | 1.8e-01   
            ...   
        >>> X_minres.shape
        torch.Size([2, 6, 4, 100, 3])
        >>> torch.allclose(X_minres,X_true)
        True
        >>> print_data_signatures(data)
            data['x'].shape = (2, 6, 4, 100, 3)
            data['iterrange'].shape = (177,)
            data['times'].shape = (177,)
            data['losses_quantiles']
                data['losses_quantiles']['0'].shape = (177,)
                data['losses_quantiles']['1'].shape = (177,)
                data['losses_quantiles']['5'].shape = (177,)
                data['losses_quantiles']['10'].shape = (177,)
                data['losses_quantiles']['25'].shape = (177,)
                data['losses_quantiles']['40'].shape = (177,)
                data['losses_quantiles']['50'].shape = (177,)
                data['losses_quantiles']['60'].shape = (177,)
                data['losses_quantiles']['75'].shape = (177,)
                data['losses_quantiles']['90'].shape = (177,)
                data['losses_quantiles']['95'].shape = (177,)
                data['losses_quantiles']['99'].shape = (177,)
                data['losses_quantiles']['100'].shape = (177,)
            data['xs'].shape = (177, 2, 6, 4, 100, 3)
            data['losses'].shape = (177, 2, 6, 4, 3)
    """
    if warn and (not torch.get_default_dtype()==torch.float64): warnings.warn('''
            torch.get_default_dtype() = %s, but minres often requires high precision updates. We recommend using:
                torch.set_default_dtype(torch.float64)'''%str(torch.get_default_dtype()))
    assert torch.get_default_dtype() in [torch.float32,torch.float64]
    default_dtype = torch.get_default_dtype()
    device = str(B.device)
    default_device = str(torch.get_default_device())
    assert B.ndim>=2, "B should have shape (...,n,k)"
    n = B.size(-2)
    k = B.size(-1)
    if X0 is None: 
        X0 = torch.zeros_like(B)
    if isinstance(A,torch.Tensor):
        assert A.shape[-2:]==(n,n)
        assert torch.allclose(A.adjoint(),A)
        matvec = lambda X: torch.einsum("...ij,...jk->...ik",A,X)
    else:
        assert callable(A)
        matvec = A
    if iters is None: 
        iters = 5*n 
    assert iters>=0
    assert iters%1==0
    if residtol is None: 
        if default_dtype==torch.float64:
            residtol = 1e-12
        elif default_dtype==torch.float32:
            residtol = 2.5e-4
        else:
            raise Exception("default_dtype = %s not parsed"%str(default_dtype))
    assert residtol>=0
    if verbose is None: 
        verbose = max(1,iters//20)
    assert verbose%1==0
    assert verbose>=0 
    if store_data_iters is None: 
        store_data_iters = max(1,iters//1000)
    assert store_data_iters%1==0
    assert store_data_iters>=0 
    assert isinstance(store_all_data,bool)
    assert isinstance(quantiles_losses,list)
    assert all(0<=qt<=100 for qt in quantiles_losses)
    assert isinstance(verbose_quantiles_losses,list)
    assert all(qt in quantiles_losses for qt in verbose_quantiles_losses)
    assert verbose_indent%1==0 
    assert verbose_indent>=0
    assert isinstance(verbose_times,bool)
    if verbose:
        _h_iter = "%-10s "%"iter i"
        _h_times = "| %-10s"%"times" if verbose_times else ""
        _s_losses_qt = ("| %-9s "*len(verbose_quantiles_losses))%tuple(str(qt) for qt in verbose_quantiles_losses)
        _h_losses_qt = "| losses_quantiles"+" "*(len(_s_losses_qt)-len("| losses_quantiles"))
        _h = _h_iter+_h_losses_qt+_h_times
        _s = " "*len(_h_iter)+_s_losses_qt+("|"+" "*(len(_h_times)-1) if verbose_times else " "*len(_h_times))
        print(" "*verbose_indent+_h)
        print(" "*verbose_indent+_s)
        print(" "*verbose_indent+"~"*len(_s))
    timer = Timer(device=device)
    timer.tic()
    psolve = lambda X: X # TODO: implement more involved preconditioned solver
    inner = lambda a,b: torch.einsum("...ij,...ij->...j",a.conj(),b)
    Anorm = 0
    eps = torch.finfo(B.dtype).eps
    x = X0 
    Ax = matvec(x)
    assert Ax.shape[-2:]==(n,k)
    batch_shape = tuple(Ax.shape[:-2])
    if store_data_iters:
        iterrange = []
        times = []
        losses = []
        losses_quantiles = {str(qt):[] for qt in quantiles_losses}
        if store_all_data:
            xs = []
    r1 = B-Ax # (...,n,k)
    y = psolve(r1) # (...,n,k)
    beta1 = torch.sqrt(inner(r1,y)) # (...,k)
    bnorm = torch.linalg.norm(B,dim=-2) # (...,k)
    oldb = 0
    beta = beta1
    dbar = 0
    epsln = torch.zeros(1,device=device)
    phibar = beta1
    tnorm2 = 0
    cs = -1
    sn = 0
    w = torch.zeros_like(B)
    w2 = torch.zeros_like(B)
    r2 = r1
    shift = 0 # TODO: If shift != 0 then the method solves (A - shift*I)x = b
    for i in range(iters+1):
        resid = matvec(x)-B 
        breakcond = i==iters or resid.abs().amax()<=residtol
        loss = torch.linalg.norm(resid,dim=-2)/bnorm
        times_i = timer.toc()
        losses_quantiles_i = {str(qt): loss.nanquantile(qt/100) for qt in quantiles_losses}
        if store_data_iters and (i%store_data_iters==0 or breakcond):
            iterrange.append(i)
            losses.append(loss.to(default_device))
            times.append(times_i)
            for qt in quantiles_losses:
                losses_quantiles[str(qt)].append(losses_quantiles_i[str(qt)].to(default_device))
            if store_all_data:
                xs.append(x.expand(resid.shape).to(default_device))
        if verbose and (i%verbose==0 or breakcond):
            _s_iter = "%-10d "%i
            _s_losses_qt = ("| %-9.1e "*len(verbose_quantiles_losses))%tuple(losses_quantiles_i[str(qt)] for qt in verbose_quantiles_losses)
            _s_times = "| %-10.1f "%(times_i) if verbose_times else ""
            print(" "*verbose_indent+_s_iter+_s_losses_qt+_s_times)
        if breakcond: break 
        s = 1/beta
        v = s[...,None,:]*y
        y = matvec(v)
        y = y-shift*v
        if i>0:
            y = y-(beta/oldb)[...,None,:]*r1
        alfa = inner(v,y)
        y = y-(alfa/beta)[...,None,:]*r2
        r1 = r2
        r2 = y
        y = psolve(r2)
        oldb = beta
        beta = inner(r2,y)
        beta = torch.sqrt(beta)
        tnorm2 += alfa**2+oldb**2+beta**2
        oldeps = epsln
        delta = cs*dbar+sn*alfa
        gbar = sn*dbar-cs*alfa
        epsln = sn*beta
        dbar = -cs*beta
        gamma = torch.linalg.norm(torch.stack([gbar,beta],dim=-1),dim=-1)
        gamma = torch.maximum(gamma,eps*torch.ones(1,device=device))
        cs = gbar/gamma
        sn = beta/gamma
        phi = cs*phibar
        phibar = sn*phibar
        denom = 1/gamma
        w1 = w2
        w2 = w
        w = (v-oldeps[...,None,:]*w1-delta[...,None,:]*w2)*denom[...,None,:]
        x = x+phi[...,None,:]*w
    if store_data_iters==0:
        return x 
    else:
        data = {
            "x": x.to(default_device), 
            "iterrange": torch.tensor(iterrange,dtype=int), 
            "times": torch.tensor(times), 
            "losses_quantiles": {str(qt):torch.tensor(losses_quantiles[str(qt)]) for qt in quantiles_losses},
            }
        if store_all_data:
            data["xs"] = torch.stack(xs,dim=0)
            data["losses"] = torch.stack(losses,dim=0)
        return x,data

minres_qlp_cs 

minres_qlp_cs(A, B, X0=None, iters=None, residtol=None, verbose=False, verbose_indent=4, quantiles_losses=[0, 1, 5, 10, 25, 40, 50, 60, 75, 90, 95, 99, 100], verbose_quantiles_losses=[5, 25, 50, 75, 90], verbose_times=True, warn=True, store_data_iters=False, store_all_data=False)

MINRES QLP algorith for complex-symmetric matrices.

A translation of the MATLAB version of MINRESQLP.

References

  1. S.-C. Choi, C. C. Paige, and M. A. Saunders, MINRES-QLP: A Krylov subspace method for indefinite or singular symmetric systems, SIAM Journal of Scientific Computing, submitted on March 7, 2010.

  2. S.-C. Choi's PhD Dissertation, Stanford University, 2006: http://www.stanford.edu/group/SOL/dissertations.html

Parameters:

Name Type Description Default
A Union[Tensor, callable]

Symmetric matrix A with shape (...,n,n), or
callable(A) where A(X) should return the batch matrix multiplication of A and X,

required
B Tensor

Right hand side tensor \(B\) with shape (...,n,k)

 required
X0 Tensor

Initial guess for \(X\) with shape (...,n,k), defaults to zeros.

None
iters int

number of minres iterations, defaults to 5n.

None
residtol float

Non-negative tolerance on the maximum residual for early stopping, defaults to 1e-12 for torch.float64 and 2.5e-4 for torch.float32.

 None
verbose int

Controls logging verbosity

  • If True, perform logging.
  • If a positive int, only log every verbose iterations.
  • If None, set to a reasonable positive int based on the maximum number of iterations
  • If False, don't log.
False
verbose_indent int

Non-negative number of indentation spaces for logging.

 4
quantiles_losses list

Loss quantiles to record.

 [0, 1, 5, 10, 25, 40, 50, 60, 75, 90, 95, 99, 100]
verbose_quantiles_losses list

Loss quantiles to show in verbose log.

 [5, 25, 50, 75, 90]
verbose_times bool

If False, do not show the times in the verbose log. This is mostly for testing where timing is not reproducible.

True
warn bool

If False, then suppress warnings.

 True
store_data_iters int

Controls storage iterations with the same options as verbose. If store_data_iters==0, then the data is not collected or returned.

  • If True, store every iteration.
  • If a positive int, only store every store_data_iters iterations.
  • If None, set to a reasonable positive int based on the maximum number of iterations
  • If False, don't store data, and do not return data
False
store_all_data bool

If True, store the x values as well as the metrics.

False

Returns:

Name Type Description
x Tensor

Optimized \(X\).
data dict

Iteration data, only returned when store_data_iters>0

Examples:

>>> torch.set_default_dtype(torch.float64)
>>> rng = torch.Generator().manual_seed(7)

Column vector \(b\) 

>>> n = 5
>>> A = torch.randn(n,n,dtype=torch.complex128,generator=rng)
>>> A = (A+A.T)/2
>>> b = torch.rand(n,dtype=torch.complex128,generator=rng)
>>> x_true = torch.linalg.solve(A,b[...,None])[...,0]
>>> x_true
tensor([-0.6207-0.4121j,  0.5221+0.3249j, -1.0952+0.8594j,  0.9080-1.2110j,
        -0.9799+0.7372j])
>>> torch.allclose(A@x_true-b,torch.zeros_like(b))
True
>>> x_minres = minres_qlp_cs(A,b[...,None],verbose=None,verbose_times=False)[...,0]
    iter i     | losses_quantiles                                          
               | 5         | 25        | 50        | 75        | 90        
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
    1          | 9.6e-01   | 9.6e-01   | 9.6e-01   | 9.6e-01   | 9.6e-01   
    2          | 5.8e-01   | 5.8e-01   | 5.8e-01   | 5.8e-01   | 5.8e-01   
    3          | 4.5e-01   | 4.5e-01   | 4.5e-01   | 4.5e-01   | 4.5e-01   
    4          | 2.8e-01   | 2.8e-01   | 2.8e-01   | 2.8e-01   | 2.8e-01   
    5          | 1.7e-14   | 1.7e-14   | 1.7e-14   | 1.7e-14   | 1.7e-14   
>>> torch.allclose(x_minres,x_true)
True

Matrix \(B\)

>>> n = 5
>>> k = 3
>>> A = torch.randn(n,n,dtype=torch.complex128,generator=rng)
>>> A = (A+A.T)/2
>>> B = torch.rand(n,k,dtype=torch.complex128,generator=rng)
>>> X_true = torch.linalg.solve(A,B)
>>> X_true
tensor([[ 0.0142+0.7190j,  0.0097+0.9734j, -0.3620+0.6413j],
        [ 0.4527+0.7455j,  0.4270+0.4941j,  0.9255+0.5685j],
        [-0.7182-1.0284j,  0.3193-0.4463j, -0.3421-0.8780j],
        [-0.5973+0.4910j, -0.8147+0.8256j, -0.8821+0.7147j],
        [-0.0926+0.5500j, -0.5192+0.2817j, -0.8496+0.6976j]])
>>> torch.allclose(A@X_true-B,torch.zeros_like(B))
True
>>> X_minres = minres_qlp_cs(A,B,verbose=None,verbose_times=False)
    iter i     | losses_quantiles                                          
               | 5         | 25        | 50        | 75        | 90        
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
    1          | 9.0e-01   | 9.1e-01   | 9.2e-01   | 9.5e-01   | 9.7e-01   
    2          | 2.2e-01   | 3.1e-01   | 4.3e-01   | 5.0e-01   | 5.5e-01   
    3          | 2.0e-01   | 2.2e-01   | 2.3e-01   | 2.9e-01   | 3.3e-01   
    4          | 3.4e-02   | 5.0e-02   | 7.1e-02   | 8.1e-02   | 8.7e-02   
    5          | 1.9e-15   | 4.2e-15   | 7.0e-15   | 1.5e-14   | 2.0e-14   
>>> torch.allclose(X_minres,X_true)
True

Tri-diagonal \(A\) with storage-saving multiplication function 

>>> n = 4
>>> k = 3
>>> A_diag = torch.randn(n,dtype=torch.complex128,generator=rng)
>>> A_off_diag = torch.randn(n-1,dtype=torch.complex128,generator=rng) 
>>> A = torch.zeros(n,n,dtype=torch.complex128)
>>> A[torch.arange(n),torch.arange(n)] = A_diag 
>>> A[torch.arange(n-1),torch.arange(1,n)] = A_off_diag
>>> A[torch.arange(1,n),torch.arange(n-1)] = A_off_diag
>>> A
tensor([[ 0.4070+0.4993j, -0.3137-0.5164j,  0.0000+0.0000j,  0.0000+0.0000j],
        [-0.3137-0.5164j, -0.2736+0.1860j, -0.2956-0.1092j,  0.0000+0.0000j],
        [ 0.0000+0.0000j, -0.2956-0.1092j,  0.4033-0.3862j, -0.0085+0.1981j],
        [ 0.0000+0.0000j,  0.0000+0.0000j, -0.0085+0.1981j, -0.1929-0.0194j]])
>>> B = torch.rand(n,k,dtype=torch.complex128,generator=rng)
>>> X_true = torch.linalg.solve(A,B)
>>> X_true
tensor([[-0.0284-1.1816j, -0.6400-0.1975j,  0.7656-1.1306j],
        [-1.6586-0.3570j, -2.4720-0.1880j, -1.1169-0.6640j],
        [-2.4585-1.2378j, -1.0485-2.0358j, -3.1323-0.3632j],
        [ 0.1591-7.4823j,  1.8120-2.6001j, -2.3910-7.6943j]])
>>> torch.allclose(A@X_true-B,torch.zeros_like(B))
True
>>> def A_mult(x):
...     y = x*A_diag[:,None]
...     y[1:,:] += x[:-1,:]*A_off_diag[:,None]
...     y[:-1,:] += x[1:,:]*A_off_diag[:,None]
...     return y
>>> torch.allclose(A_mult(X_true),A@X_true)
True
>>> X_minres = minres_qlp_cs(A_mult,B,verbose=None,verbose_times=False)
    iter i     | losses_quantiles                                          
               | 5         | 25        | 50        | 75        | 90        
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
    1          | 7.8e-01   | 8.3e-01   | 8.9e-01   | 9.4e-01   | 9.7e-01   
    2          | 5.1e-01   | 5.2e-01   | 5.2e-01   | 5.4e-01   | 5.5e-01   
    3          | 3.2e-01   | 3.2e-01   | 3.2e-01   | 3.2e-01   | 3.3e-01   
    4          | 9.1e-16   | 1.1e-15   | 1.3e-15   | 2.1e-15   | 2.6e-15   
>>> torch.allclose(X_minres,X_true)
True

Batched tri-diagonal \(A\) with storage-saving multiplication function 

>>> n = 100
>>> k = 3
>>> A_diag = torch.randn(2,1,4,n,dtype=torch.complex128,generator=rng)
>>> A_off_diag = torch.randn(2,1,4,n-1,dtype=torch.complex128,generator=rng) 
>>> A = torch.zeros(2,1,4,n,n,dtype=torch.complex128)
>>> A[...,torch.arange(n),torch.arange(n)] = A_diag 
>>> A[...,torch.arange(n-1),torch.arange(1,n)] = A_off_diag
>>> A[...,torch.arange(1,n),torch.arange(n-1)] = A_off_diag
>>> B = torch.rand(2,6,1,n,k,dtype=torch.complex128,generator=rng)
>>> X_true = torch.linalg.solve(A,B)
>>> torch.allclose(torch.einsum("...ij,...jk->...ik",A,X_true)-B,torch.zeros_like(B))
True
>>> def A_mult(x):
...     y = x*A_diag[...,:,None]
...     y[...,1:,:] += x[...,:-1,:]*A_off_diag[...,:,None]
...     y[...,:-1,:] += x[...,1:,:]*A_off_diag[...,:,None]
...     return y
>>> torch.allclose(A_mult(X_true),torch.einsum("...ij,...jk->...ik",A,X_true))
True
>>> X_minres,data = minres_qlp_cs(A_mult,B,verbose=None,verbose_times=False,store_data_iters=None,store_all_data=True,iters=40)
    iter i     | losses_quantiles                                          
               | 5         | 25        | 50        | 75        | 90        
    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
    2          | 6.6e-01   | 6.8e-01   | 7.0e-01   | 7.2e-01   | 7.3e-01   
    4          | 5.1e-01   | 5.4e-01   | 5.6e-01   | 5.8e-01   | 5.9e-01   
    6          | 4.2e-01   | 4.5e-01   | 4.6e-01   | 4.9e-01   | 5.0e-01   
    8          | 3.5e-01   | 3.8e-01   | 4.0e-01   | 4.2e-01   | 4.4e-01   
    10         | 3.0e-01   | 3.3e-01   | 3.6e-01   | 3.8e-01   | 4.0e-01   
    12         | 2.6e-01   | 2.9e-01   | 3.2e-01   | 3.4e-01   | 3.6e-01   
    14         | 2.3e-01   | 2.6e-01   | 2.9e-01   | 3.1e-01   | 3.3e-01   
    16         | 2.1e-01   | 2.3e-01   | 2.6e-01   | 2.8e-01   | 3.0e-01   
    18         | 1.9e-01   | 2.1e-01   | 2.4e-01   | 2.6e-01   | 2.8e-01   
    20         | 1.7e-01   | 1.9e-01   | 2.2e-01   | 2.3e-01   | 2.6e-01   
    22         | 1.5e-01   | 1.7e-01   | 2.0e-01   | 2.2e-01   | 2.4e-01   
    24         | 1.3e-01   | 1.6e-01   | 1.8e-01   | 2.0e-01   | 2.2e-01   
    26         | 1.2e-01   | 1.4e-01   | 1.7e-01   | 1.9e-01   | 2.1e-01   
    28         | 1.0e-01   | 1.3e-01   | 1.6e-01   | 1.8e-01   | 2.0e-01   
    30         | 9.3e-02   | 1.2e-01   | 1.4e-01   | 1.7e-01   | 1.9e-01   
    32         | 8.4e-02   | 1.1e-01   | 1.3e-01   | 1.6e-01   | 1.9e-01   
    34         | 7.6e-02   | 1.0e-01   | 1.2e-01   | 1.6e-01   | 1.8e-01   
    36         | 6.9e-02   | 9.3e-02   | 1.1e-01   | 1.5e-01   | 1.7e-01   
    38         | 6.4e-02   | 8.5e-02   | 1.0e-01   | 1.4e-01   | 1.7e-01   
    40         | 5.8e-02   | 7.5e-02   | 9.2e-02   | 1.4e-01   | 1.6e-01   
>>> X_minres.shape
torch.Size([2, 6, 4, 100, 3])
>>> torch.allclose(X_minres,X_true)
False
>>> print_data_signatures(data)
    data['x'].shape = (2, 6, 4, 100, 3)
    data['iterrange'].shape = (41,)
    data['times'].shape = (41,)
    data['losses_quantiles']
        data['losses_quantiles']['0'].shape = (41,)
        data['losses_quantiles']['1'].shape = (41,)
        data['losses_quantiles']['5'].shape = (41,)
        data['losses_quantiles']['10'].shape = (41,)
        data['losses_quantiles']['25'].shape = (41,)
        data['losses_quantiles']['40'].shape = (41,)
        data['losses_quantiles']['50'].shape = (41,)
        data['losses_quantiles']['60'].shape = (41,)
        data['losses_quantiles']['75'].shape = (41,)
        data['losses_quantiles']['90'].shape = (41,)
        data['losses_quantiles']['95'].shape = (41,)
        data['losses_quantiles']['99'].shape = (41,)
        data['losses_quantiles']['100'].shape = (41,)
    data['xs'].shape = (41, 2, 6, 4, 100, 3)
    data['losses'].shape = (41, 2, 6, 4, 3)
Source code in agsutil/algos.py 
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
def minres_qlp_cs(
        A,
        B,
        X0 = None,
        iters = None,
        residtol = None,
        verbose = False, 
        verbose_indent = 4,
        quantiles_losses = [0,1,5,10,25,40,50,60,75,90,95,99,100],
        verbose_quantiles_losses = [5,25,50,75,90],
        verbose_times = True, 
        warn = True,
        store_data_iters = False, 
        store_all_data = False,
        ):
    # https://github.com/schoi32/sci498rms/blob/master/Sorokin_MinresQLP_Python_Workspace/MinresQLP/Algorithms/cs_mqlp.py
    r"""
    MINRES QLP algorith for complex-symmetric matrices. 

    A translation of the [MATLAB version of MINRESQLP](http://www.stanford.edu/group/SOL/software.html).

    References 

    1.  S.-C. Choi, C. C. Paige, and M. A. Saunders,
        MINRES-QLP: A Krylov subspace method for indefinite or singular symmetric systems,
        SIAM Journal of Scientific Computing, submitted on March 7, 2010.

    2.  S.-C. Choi's PhD Dissertation, Stanford University, 2006: 
        http://www.stanford.edu/group/SOL/dissertations.html

    Args:
        A (Union[torch.Tensor,callable]): Symmetric matrix `A` with shape `(...,n,n)`, or  
            `callable(A)` where `A(X)` should return the batch matrix multiplication of `A` and `X`,  
        B (torch.Tensor): Right hand side tensor $B$ with shape `(...,n,k)`
        X0 (torch.Tensor): Initial guess for $X$ with shape `(...,n,k)`, defaults to zeros. 
        iters (int): number of minres iterations, defaults to `5n`. 
        residtol (float): Non-negative tolerance on the maximum residual for early stopping, defaults to `1e-12` for `torch.float64` and `2.5e-4` for `torch.float32`.
        verbose (int): Controls logging verbosity

            - If `True`, perform logging. 
            - If a positive int, only log every `verbose` iterations. 
            - If `None`, set to a reasonable positive int based on the maximum number of iterations
            - If `False`, don't log. 

        verbose_indent (int): Non-negative number of indentation spaces for logging.
        quantiles_losses (list): Loss quantiles to record.
        verbose_quantiles_losses (list): Loss quantiles to show in verbose log.
        verbose_times (bool): If `False`, do not show the times in the verbose log. This is mostly for testing where timing is not reproducible. 
        warn (bool): If `False`, then suppress warnings.
        store_data_iters (int): Controls storage iterations with the same options as verbose. If `store_data_iters==0`, then the data is not collected or returned. 

            - If `True`, store every iteration. 
            - If a positive int, only store every `store_data_iters` iterations. 
            - If `None`, set to a reasonable positive int based on the maximum number of iterations
            - If `False`, don't store data, and do not return data 

        store_all_data (bool): If `True`, store the `x` values as well as the metrics. 

    Returns:
        x (torch.Tensor): Optimized $X$.
        data (dict): Iteration data, only returned when `store_data_iters>0`

    Examples:

        >>> torch.set_default_dtype(torch.float64)
        >>> rng = torch.Generator().manual_seed(7)

    Column vector $b$ 

        >>> n = 5
        >>> A = torch.randn(n,n,dtype=torch.complex128,generator=rng)
        >>> A = (A+A.T)/2
        >>> b = torch.rand(n,dtype=torch.complex128,generator=rng)
        >>> x_true = torch.linalg.solve(A,b[...,None])[...,0]
        >>> x_true
        tensor([-0.6207-0.4121j,  0.5221+0.3249j, -1.0952+0.8594j,  0.9080-1.2110j,
                -0.9799+0.7372j])
        >>> torch.allclose(A@x_true-b,torch.zeros_like(b))
        True
        >>> x_minres = minres_qlp_cs(A,b[...,None],verbose=None,verbose_times=False)[...,0]
            iter i     | losses_quantiles                                          
                       | 5         | 25        | 50        | 75        | 90        
            ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
            1          | 9.6e-01   | 9.6e-01   | 9.6e-01   | 9.6e-01   | 9.6e-01   
            2          | 5.8e-01   | 5.8e-01   | 5.8e-01   | 5.8e-01   | 5.8e-01   
            3          | 4.5e-01   | 4.5e-01   | 4.5e-01   | 4.5e-01   | 4.5e-01   
            4          | 2.8e-01   | 2.8e-01   | 2.8e-01   | 2.8e-01   | 2.8e-01   
            5          | 1.7e-14   | 1.7e-14   | 1.7e-14   | 1.7e-14   | 1.7e-14   
        >>> torch.allclose(x_minres,x_true)
        True

    Matrix $B$

        >>> n = 5
        >>> k = 3
        >>> A = torch.randn(n,n,dtype=torch.complex128,generator=rng)
        >>> A = (A+A.T)/2
        >>> B = torch.rand(n,k,dtype=torch.complex128,generator=rng)
        >>> X_true = torch.linalg.solve(A,B)
        >>> X_true
        tensor([[ 0.0142+0.7190j,  0.0097+0.9734j, -0.3620+0.6413j],
                [ 0.4527+0.7455j,  0.4270+0.4941j,  0.9255+0.5685j],
                [-0.7182-1.0284j,  0.3193-0.4463j, -0.3421-0.8780j],
                [-0.5973+0.4910j, -0.8147+0.8256j, -0.8821+0.7147j],
                [-0.0926+0.5500j, -0.5192+0.2817j, -0.8496+0.6976j]])
        >>> torch.allclose(A@X_true-B,torch.zeros_like(B))
        True
        >>> X_minres = minres_qlp_cs(A,B,verbose=None,verbose_times=False)
            iter i     | losses_quantiles                                          
                       | 5         | 25        | 50        | 75        | 90        
            ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
            1          | 9.0e-01   | 9.1e-01   | 9.2e-01   | 9.5e-01   | 9.7e-01   
            2          | 2.2e-01   | 3.1e-01   | 4.3e-01   | 5.0e-01   | 5.5e-01   
            3          | 2.0e-01   | 2.2e-01   | 2.3e-01   | 2.9e-01   | 3.3e-01   
            4          | 3.4e-02   | 5.0e-02   | 7.1e-02   | 8.1e-02   | 8.7e-02   
            5          | 1.9e-15   | 4.2e-15   | 7.0e-15   | 1.5e-14   | 2.0e-14   
        >>> torch.allclose(X_minres,X_true)
        True

    Tri-diagonal $A$ with storage-saving multiplication function 

        >>> n = 4
        >>> k = 3
        >>> A_diag = torch.randn(n,dtype=torch.complex128,generator=rng)
        >>> A_off_diag = torch.randn(n-1,dtype=torch.complex128,generator=rng) 
        >>> A = torch.zeros(n,n,dtype=torch.complex128)
        >>> A[torch.arange(n),torch.arange(n)] = A_diag 
        >>> A[torch.arange(n-1),torch.arange(1,n)] = A_off_diag
        >>> A[torch.arange(1,n),torch.arange(n-1)] = A_off_diag
        >>> A
        tensor([[ 0.4070+0.4993j, -0.3137-0.5164j,  0.0000+0.0000j,  0.0000+0.0000j],
                [-0.3137-0.5164j, -0.2736+0.1860j, -0.2956-0.1092j,  0.0000+0.0000j],
                [ 0.0000+0.0000j, -0.2956-0.1092j,  0.4033-0.3862j, -0.0085+0.1981j],
                [ 0.0000+0.0000j,  0.0000+0.0000j, -0.0085+0.1981j, -0.1929-0.0194j]])
        >>> B = torch.rand(n,k,dtype=torch.complex128,generator=rng)
        >>> X_true = torch.linalg.solve(A,B)
        >>> X_true
        tensor([[-0.0284-1.1816j, -0.6400-0.1975j,  0.7656-1.1306j],
                [-1.6586-0.3570j, -2.4720-0.1880j, -1.1169-0.6640j],
                [-2.4585-1.2378j, -1.0485-2.0358j, -3.1323-0.3632j],
                [ 0.1591-7.4823j,  1.8120-2.6001j, -2.3910-7.6943j]])
        >>> torch.allclose(A@X_true-B,torch.zeros_like(B))
        True
        >>> def A_mult(x):
        ...     y = x*A_diag[:,None]
        ...     y[1:,:] += x[:-1,:]*A_off_diag[:,None]
        ...     y[:-1,:] += x[1:,:]*A_off_diag[:,None]
        ...     return y
        >>> torch.allclose(A_mult(X_true),A@X_true)
        True
        >>> X_minres = minres_qlp_cs(A_mult,B,verbose=None,verbose_times=False)
            iter i     | losses_quantiles                                          
                       | 5         | 25        | 50        | 75        | 90        
            ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
            1          | 7.8e-01   | 8.3e-01   | 8.9e-01   | 9.4e-01   | 9.7e-01   
            2          | 5.1e-01   | 5.2e-01   | 5.2e-01   | 5.4e-01   | 5.5e-01   
            3          | 3.2e-01   | 3.2e-01   | 3.2e-01   | 3.2e-01   | 3.3e-01   
            4          | 9.1e-16   | 1.1e-15   | 1.3e-15   | 2.1e-15   | 2.6e-15   
        >>> torch.allclose(X_minres,X_true)
        True

    Batched tri-diagonal $A$ with storage-saving multiplication function 

        >>> n = 100
        >>> k = 3
        >>> A_diag = torch.randn(2,1,4,n,dtype=torch.complex128,generator=rng)
        >>> A_off_diag = torch.randn(2,1,4,n-1,dtype=torch.complex128,generator=rng) 
        >>> A = torch.zeros(2,1,4,n,n,dtype=torch.complex128)
        >>> A[...,torch.arange(n),torch.arange(n)] = A_diag 
        >>> A[...,torch.arange(n-1),torch.arange(1,n)] = A_off_diag
        >>> A[...,torch.arange(1,n),torch.arange(n-1)] = A_off_diag
        >>> B = torch.rand(2,6,1,n,k,dtype=torch.complex128,generator=rng)
        >>> X_true = torch.linalg.solve(A,B)
        >>> torch.allclose(torch.einsum("...ij,...jk->...ik",A,X_true)-B,torch.zeros_like(B))
        True
        >>> def A_mult(x):
        ...     y = x*A_diag[...,:,None]
        ...     y[...,1:,:] += x[...,:-1,:]*A_off_diag[...,:,None]
        ...     y[...,:-1,:] += x[...,1:,:]*A_off_diag[...,:,None]
        ...     return y
        >>> torch.allclose(A_mult(X_true),torch.einsum("...ij,...jk->...ik",A,X_true))
        True
        >>> X_minres,data = minres_qlp_cs(A_mult,B,verbose=None,verbose_times=False,store_data_iters=None,store_all_data=True,iters=40)
            iter i     | losses_quantiles                                          
                       | 5         | 25        | 50        | 75        | 90        
            ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
            0          | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   | 1.0e+00   
            2          | 6.6e-01   | 6.8e-01   | 7.0e-01   | 7.2e-01   | 7.3e-01   
            4          | 5.1e-01   | 5.4e-01   | 5.6e-01   | 5.8e-01   | 5.9e-01   
            6          | 4.2e-01   | 4.5e-01   | 4.6e-01   | 4.9e-01   | 5.0e-01   
            8          | 3.5e-01   | 3.8e-01   | 4.0e-01   | 4.2e-01   | 4.4e-01   
            10         | 3.0e-01   | 3.3e-01   | 3.6e-01   | 3.8e-01   | 4.0e-01   
            12         | 2.6e-01   | 2.9e-01   | 3.2e-01   | 3.4e-01   | 3.6e-01   
            14         | 2.3e-01   | 2.6e-01   | 2.9e-01   | 3.1e-01   | 3.3e-01   
            16         | 2.1e-01   | 2.3e-01   | 2.6e-01   | 2.8e-01   | 3.0e-01   
            18         | 1.9e-01   | 2.1e-01   | 2.4e-01   | 2.6e-01   | 2.8e-01   
            20         | 1.7e-01   | 1.9e-01   | 2.2e-01   | 2.3e-01   | 2.6e-01   
            22         | 1.5e-01   | 1.7e-01   | 2.0e-01   | 2.2e-01   | 2.4e-01   
            24         | 1.3e-01   | 1.6e-01   | 1.8e-01   | 2.0e-01   | 2.2e-01   
            26         | 1.2e-01   | 1.4e-01   | 1.7e-01   | 1.9e-01   | 2.1e-01   
            28         | 1.0e-01   | 1.3e-01   | 1.6e-01   | 1.8e-01   | 2.0e-01   
            30         | 9.3e-02   | 1.2e-01   | 1.4e-01   | 1.7e-01   | 1.9e-01   
            32         | 8.4e-02   | 1.1e-01   | 1.3e-01   | 1.6e-01   | 1.9e-01   
            34         | 7.6e-02   | 1.0e-01   | 1.2e-01   | 1.6e-01   | 1.8e-01   
            36         | 6.9e-02   | 9.3e-02   | 1.1e-01   | 1.5e-01   | 1.7e-01   
            38         | 6.4e-02   | 8.5e-02   | 1.0e-01   | 1.4e-01   | 1.7e-01   
            40         | 5.8e-02   | 7.5e-02   | 9.2e-02   | 1.4e-01   | 1.6e-01   
        >>> X_minres.shape
        torch.Size([2, 6, 4, 100, 3])
        >>> torch.allclose(X_minres,X_true)
        False
        >>> print_data_signatures(data)
            data['x'].shape = (2, 6, 4, 100, 3)
            data['iterrange'].shape = (41,)
            data['times'].shape = (41,)
            data['losses_quantiles']
                data['losses_quantiles']['0'].shape = (41,)
                data['losses_quantiles']['1'].shape = (41,)
                data['losses_quantiles']['5'].shape = (41,)
                data['losses_quantiles']['10'].shape = (41,)
                data['losses_quantiles']['25'].shape = (41,)
                data['losses_quantiles']['40'].shape = (41,)
                data['losses_quantiles']['50'].shape = (41,)
                data['losses_quantiles']['60'].shape = (41,)
                data['losses_quantiles']['75'].shape = (41,)
                data['losses_quantiles']['90'].shape = (41,)
                data['losses_quantiles']['95'].shape = (41,)
                data['losses_quantiles']['99'].shape = (41,)
                data['losses_quantiles']['100'].shape = (41,)
            data['xs'].shape = (41, 2, 6, 4, 100, 3)
            data['losses'].shape = (41, 2, 6, 4, 3)
    """
    if warn and (not torch.get_default_dtype()==torch.float64): warnings.warn('''
            torch.get_default_dtype() = %s, but minres_qlp_cs often requires high precision updates. We recommend using:
                torch.set_default_dtype(torch.float64)'''%str(torch.get_default_dtype()))
    assert torch.get_default_dtype() in [torch.float32,torch.float64]
    default_dtype = torch.get_default_dtype()
    device = str(B.device)
    default_device = str(torch.get_default_device())
    assert B.ndim>=2, "B should have shape (...,n,k)"
    n = B.size(-2)
    k = B.size(-1)
    assert B.dtype in [torch.complex64,torch.complex128]
    if X0 is None: 
        X0 = torch.zeros_like(B)
    if isinstance(A,torch.Tensor):
        assert A.shape[-2:]==(n,n)
        assert torch.allclose(A.T,A)
        matvec = lambda X: torch.einsum("...ij,...jk->...ik",A,X)
    else:
        assert callable(A)
        matvec = A
    if iters is None: 
        iters = 5*n 
    assert iters>=0
    assert iters%1==0
    if residtol is None: 
        if default_dtype==torch.float64:
            residtol = 1e-12
        elif default_dtype==torch.float32:
            residtol = 2.5e-4
        else:
            raise Exception("default_dtype = %s not parsed"%str(default_dtype))
    assert residtol>=0
    if verbose is None: 
        verbose = max(1,iters//20)
    assert verbose%1==0
    assert verbose>=0 
    if store_data_iters is None: 
        store_data_iters = max(1,iters//1000)
    assert store_data_iters%1==0
    assert store_data_iters>=0 
    assert isinstance(store_all_data,bool)
    assert isinstance(quantiles_losses,list)
    assert all(0<=qt<=100 for qt in quantiles_losses)
    assert isinstance(verbose_quantiles_losses,list)
    assert all(qt in quantiles_losses for qt in verbose_quantiles_losses)
    assert verbose_indent%1==0 
    assert verbose_indent>=0
    assert isinstance(verbose_times,bool)
    if verbose:
        _h_iter = "%-10s "%"iter i"
        _h_times = "| %-10s"%"times" if verbose_times else ""
        _s_losses_qt = ("| %-9s "*len(verbose_quantiles_losses))%tuple(str(qt) for qt in verbose_quantiles_losses)
        _h_losses_qt = "| losses_quantiles"+" "*(len(_s_losses_qt)-len("| losses_quantiles"))
        _h = _h_iter+_h_losses_qt+_h_times
        _s = " "*len(_h_iter)+_s_losses_qt+("|"+" "*(len(_h_times)-1) if verbose_times else " "*len(_h_times))
        print(" "*verbose_indent+_h)
        print(" "*verbose_indent+_s)
        print(" "*verbose_indent+"~"*len(_s))
    timer = Timer(device=device)
    timer.tic()
    psolve = lambda X: X # TODO: implement more involved preconditioned solver
    inner = lambda a,b: torch.einsum("...ij,...ij->...j",a.conj(),b)
    Anorm = 0
    x = X0 
    Ax = matvec(x)
    assert Ax.shape[-2:]==(n,k)
    batch_shape = tuple(Ax.shape[:-2])
    if store_data_iters:
        iterrange = []
        times = []
        losses = []
        losses_quantiles = {str(qt):[] for qt in quantiles_losses}
        if store_all_data:
            xs = []
    r2 = B  # (...,k)
    r3 = r2 # (...,k)
    r3 = psolve(r2)
    beta1 = torch.sqrt(inner(r3,r2))  # (...,k)
    bnorm = torch.linalg.norm(B,dim=-2) # (...,k)
    # TODO: Check if below variables are necessary    
    beta = torch.zeros_like(beta1)
    tau = torch.zeros((*batch_shape,k),dtype=B.dtype,device=device)
    taul = torch.zeros((*batch_shape,k),dtype=B.dtype,device=device)
    phi = beta1
    betan = beta1
    cs = -torch.ones((*batch_shape,k),dtype=B.dtype,device=device)
    sn = torch.zeros((*batch_shape,k),dtype=B.dtype,device=device)
    cr1 = torch.ones((*batch_shape,k),dtype=B.dtype,device=device)
    sr1 = torch.zeros((*batch_shape,k),dtype=B.dtype,device=device)
    cr2 = -torch.ones((*batch_shape,k),dtype=B.dtype,device=device)
    sr2 = torch.zeros((*batch_shape,k),dtype=B.dtype,device=device)
    dltan = torch.zeros((*batch_shape,k),dtype=B.dtype,device=device)
    eplnn = torch.zeros((*batch_shape,k),dtype=B.dtype,device=device)
    gama = torch.zeros((*batch_shape,k),dtype=B.dtype,device=device)
    gamal = torch.zeros((*batch_shape,k),dtype=B.dtype,device=device)
    gamal2 = torch.zeros((*batch_shape,k),dtype=B.dtype,device=device)
    eta = torch.zeros((*batch_shape,k),dtype=B.dtype,device=device)
    etal = torch.zeros((*batch_shape,k),dtype=B.dtype,device=device)
    etal2 = torch.zeros((*batch_shape,k),dtype=B.dtype,device=device)
    vepln = torch.zeros((*batch_shape,k),dtype=B.dtype,device=device)
    veplnl = torch.zeros((*batch_shape,k),dtype=B.dtype,device=device)
    veplnl2 = torch.zeros((*batch_shape,k),dtype=B.dtype,device=device)
    ul3 = torch.zeros((*batch_shape,k),dtype=B.dtype,device=device)
    ul2 = torch.zeros((*batch_shape,k),dtype=B.dtype,device=device)
    ul = torch.zeros((*batch_shape,k),dtype=B.dtype,device=device)
    u = torch.zeros((*batch_shape,k),dtype=B.dtype,device=device)
    w = torch.zeros_like(B)
    wl = torch.zeros_like(B)
    r1 = torch.zeros_like(B)
    xl2 = torch.zeros_like(B)
    alfa = torch.zeros_like(B)
    shift = 0 # TODO: If shift != 0 then the method solves (A - shift*I)x = b
    for i in range(iters+1):
        resid = matvec(x)-B 
        breakcond = i==iters or resid.abs().amax()<=residtol
        loss = torch.linalg.norm(resid,dim=-2)/bnorm
        times_i = timer.toc()
        losses_quantiles_i = {str(qt): loss.nanquantile(qt/100) for qt in quantiles_losses}
        if store_data_iters and (i%store_data_iters==0 or breakcond):
            iterrange.append(i)
            losses.append(loss.to(default_device))
            times.append(times_i)
            for qt in quantiles_losses:
                losses_quantiles[str(qt)].append(losses_quantiles_i[str(qt)].to(default_device))
            if store_all_data:
                xs.append(x.expand(resid.shape).to(default_device))
        if verbose and (i%verbose==0 or breakcond):
            _s_iter = "%-10d "%i
            _s_losses_qt = ("| %-9.1e "*len(verbose_quantiles_losses))%tuple(losses_quantiles_i[str(qt)] for qt in verbose_quantiles_losses)
            _s_times = "| %-10.1f "%(times_i) if verbose_times else ""
            print(" "*verbose_indent+_s_iter+_s_losses_qt+_s_times)
        if breakcond: break 
        betal = beta
        beta = betan
        v = r3/beta[...,None,:]    
        r3 = matvec(v.conj())-shift*v.conj()
        if i>0:
            r3 = r3-(beta/betal)[...,None,:]*r1
        alfa = inner(v,r3)
        r3 = r3-(alfa/beta)[...,None,:]*r2
        r1 = r2
        r2 = r3
        r3 = psolve(r2)
        betan = torch.sqrt(inner(r2,r3))
        dbar = dltan
        dlta = cs*dbar+sn*alfa
        epln = eplnn
        gbar = sn.conj()*dbar-cs*alfa
        eplnn = sn*betan
        dltan = -cs*betan
        dlta_QLP = dlta
        gamal3 = gamal2
        gamal2 = gamal
        gamal = gama
        (cs,sn,gama) = symOrtho(gbar,betan)
        gama_tmp = gama
        taul2  = taul
        taul = tau
        tau = cs*phi
        phi = sn.conj()*phi
        if i>1:
            veplnl2  = veplnl
            etal2 = etal
            etal = eta
            dlta_tmp = sr2*vepln - cr2*dlta
            veplnl = cr2*vepln+sr2.conj()*dlta
            dlta = dlta_tmp
            eta = sr2.conj()*gama
            gama = -cr2*gama
        if i>0:
            (cr1,sr1,gamal) = symOrtho(gamal.conj(),dlta.conj())
            gamal = gamal.conj()
            vepln = sr1.conj()*gama
            gama = -cr1*gama
        ul4 = ul3
        ul3 = ul2
        if i>1:
            ul2 = (taul2-etal2*ul4-veplnl2*ul3)/gamal2
        if i>0:
            ul = (taul-etal*ul3-veplnl*ul2)/gamal
        u = (tau - eta*ul2 - vepln*ul) / gama
        if i==0:
            wl2 = wl
            wl = v.conj()*sr1.conj()[...,None,:]
            w  = v.conj()*cr1[...,None,:]
        elif i==1:
            wl2 = wl
            wl = w*cr1[...,None,:]+v.conj()*sr1.conj()[...,None,:]
            w = w*sr1[...,None,:]-v.conj()*cr1[...,None,:]
        else:
            wl2 = wl
            wl = w
            w  = wl2*sr2[...,None,:]-v.conj()*cr2[...,None,:]
            wl2 = wl2*cr2[...,None,:]+v.conj()*sr2.conj()[...,None,:]
            v = wl*cr1[...,None,:]+w*sr1.conj()[...,None,:]
            w = wl*sr1[...,None,:]-w*cr1[...,None,:]
            wl = v
        xl2 = xl2+wl2*ul2[...,None,:]
        x = xl2+wl*ul[...,None,:]+w*u[...,None,:]
        pass
        (cr2,sr2,gamal) = symOrtho(gamal.conj(),eplnn.conj())
        gamal = gamal.conj()
    if store_data_iters==0:
        return x 
    else:
        data = {
            "x": x.to(default_device), 
            "iterrange": torch.tensor(iterrange,dtype=int), 
            "times": torch.tensor(times), 
            "losses_quantiles": {str(qt):torch.tensor(losses_quantiles[str(qt)]) for qt in quantiles_losses},
            }
        if store_all_data:
            data["xs"] = torch.stack(xs,dim=0)
            data["losses"] = torch.stack(losses,dim=0)
        return x,data

root_find_bisection 

root_find_bisection(f, a, b, tol=None, max_iter=np.inf)

Finds roots of a function using the bisection method with batch support.

Parameters:

Name Type Description Default
f callable

Function whose roots to find

 required
a Tensor

Lower bounds of the search intervals.

 required
b Tensor

Upper bounds of the search intervals.

 required
tol float

Positive tolerance on the difference between a and b, defaults to 1e-12 for torch.float64 and 2.5e-4 for torch.float32.

 None
max_iter int

Maximum number of iterations.

 inf

Returns:

Name Type Description
root Tensor

Estimated roots for each batch element.

Examples:

>>> torch.set_default_dtype(torch.float64)
>>> rng = torch.Generator().manual_seed(7)

Scalar

>>> root = root_find_bisection(
...     lambda x: x**2-1/2,
...     a = torch.tensor(0),
...     b = torch.tensor(2),
... )
>>> root
tensor(0.7071)
>>> torch.allclose(root,torch.sqrt(torch.tensor(1/2)))
True

Batched

>>> ys = torch.linspace(0,1,5)[1:-1]
>>> ys 
tensor([0.2500, 0.5000, 0.7500])
>>> root = root_find_bisection(
...     lambda x: x**2-ys,
...     a = torch.zeros_like(ys),
...     b = torch.ones_like(ys),
... )
>>> root
tensor([0.5000, 0.7071, 0.8660])
>>> torch.allclose(root,torch.sqrt(ys))
True
Source code in agsutil/algos.py 
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
def root_find_bisection(f, a, b, tol=None, max_iter=np.inf):
    """
    Finds roots of a function using the bisection method with batch support.

    Parameters:
        f (callable): Function whose roots to find
        a (torch.Tensor): Lower bounds of the search intervals.
        b (torch.Tensor): Upper bounds of the search intervals.
        tol (float): Positive tolerance on the difference between a and b, defaults to `1e-12` for `torch.float64` and `2.5e-4` for `torch.float32`.
        max_iter (int): Maximum number of iterations.

    Returns:
        root (torch.Tensor): Estimated roots for each batch element.

    Examples:

        >>> torch.set_default_dtype(torch.float64)
        >>> rng = torch.Generator().manual_seed(7)

        Scalar 

        >>> root = root_find_bisection(
        ...     lambda x: x**2-1/2,
        ...     a = torch.tensor(0),
        ...     b = torch.tensor(2),
        ... )
        >>> root
        tensor(0.7071)
        >>> torch.allclose(root,torch.sqrt(torch.tensor(1/2)))
        True

        Batched 

        >>> ys = torch.linspace(0,1,5)[1:-1]
        >>> ys 
        tensor([0.2500, 0.5000, 0.7500])
        >>> root = root_find_bisection(
        ...     lambda x: x**2-ys,
        ...     a = torch.zeros_like(ys),
        ...     b = torch.ones_like(ys),
        ... )
        >>> root
        tensor([0.5000, 0.7071, 0.8660])
        >>> torch.allclose(root,torch.sqrt(ys))
        True
    """
    default_dtype = torch.get_default_dtype()
    if tol is None: 
        if default_dtype==torch.float64:
            tol = 1e-12
        elif default_dtype==torch.float32:
            tol = 2.5e-4
    assert tol>0
    a = a.clone() 
    b = b.clone()
    fa = f(a)
    fb = f(b)
    assert a.shape==fa.shape
    assert b.shape==fb.shape
    assert ((torch.sign(fa)==0)|(torch.sign(fb)==0)|(torch.sign(fa)!=torch.sign(fb))).all(), "Initial f(a) must have opposite sign of initial f(b)"
    faeq0 = fa==0
    fbeq0 = fb==0
    b[faeq0] = a[faeq0]
    a[fbeq0] = b[fbeq0]
    assert (b>=a).all()
    i = 0 
    while i<max_iter:
        c = (a+b)/2
        fc = f(c)
        breakcond = (torch.abs(b-a)<=tol).all()
        if breakcond: break
        same_sign_ac = fa*fc>0
        a = torch.where(same_sign_ac,c,a)
        fa = torch.where(same_sign_ac,fc,fa)
        b = torch.where(~same_sign_ac,c,b)
        fb = torch.where(~same_sign_ac,fc,fb)
        i += 1
    return (a+b)/2

torch autograd utils

gradb 

gradb(f, x, bkwargs={}, bdims=0, chunk_size=None)

Batched torch.func.grad and function evaluation

Parameters:

Name Type Description Default
f callable

Function to compute torch.func.grad of.

required
x Union[Tensor, Tuple]

(batched) Torch.Tensor items whose gradients will be computed.

 required
bkwargs dict

(batched) Torch.Tensor items whose gradients will not be computed.

 {}
bdims int

number of batch dimensions

0
chunk_size int

to be passed into torch.func.vmap.

None

Returns:

Name Type Description
gradys tuple

torch.Tensor gradients, one with respect to each item in x.

y Tensor

Function evaluations.

Examples:

>>> torch.set_default_dtype(torch.float64)
>>> rng = torch.Generator().manual_seed(7)

>>> f = lambda x: (x**2).sum(-1)
>>> x = torch.rand(5,generator=rng) 
>>> (grady,),y = gradb(f,x)
>>> grady.shape 
torch.Size([5])
>>> y.shape 
torch.Size([])
>>> torch.allclose(y,f(x))
True
>>> torch.allclose(grady,2*x)
True

>>> f = lambda x,z: (x**2*z**2).sum(-1)
>>> x = torch.rand(5,generator=rng) 
>>> z = torch.rand(5,generator=rng) 
>>> (grady_x,grady_z),y = gradb(f,(x,z))
>>> grady_x.shape 
torch.Size([5])
>>> grady_z.shape 
torch.Size([5])
>>> y.shape
torch.Size([])
>>> torch.allclose(y,f(x,z))
True
>>> torch.allclose(grady_x,2*x*z**2)
True
>>> torch.allclose(grady_z,x**2*2*z)
True

>>> f = lambda x: (x**2).sum(-1)
>>> x = torch.rand(3,4,5,generator=rng) 
>>> (grady,),y = gradb(f,x,bdims=2)
>>> grady.shape 
torch.Size([3, 4, 5])
>>> y.shape 
torch.Size([3, 4])
>>> torch.allclose(y,f(x))
True
>>> torch.allclose(grady,2*x)
True

>>> f = lambda x,z: (x**2*z**2).sum(-1)
>>> x = torch.rand(3,4,5,generator=rng) 
>>> z = torch.rand(3,4,5,generator=rng) 
>>> (grady_x,grady_z),y = gradb(f,(x,z),bdims=2)
>>> grady_x.shape 
torch.Size([3, 4, 5])
>>> grady_z.shape 
torch.Size([3, 4, 5])
>>> y.shape
torch.Size([3, 4])
>>> torch.allclose(y,f(x,z))
True
>>> torch.allclose(grady_x,2*x*z**2)
True
>>> torch.allclose(grady_z,x**2*2*z)
True

>>> f = lambda x,z: (x**2*z**2).sum((-2,-1))
>>> x = torch.rand(3,4,5,generator=rng) 
>>> z = torch.rand(3,4,5,generator=rng) 
>>> (grady_x,),y = gradb(f,x,bkwargs={"z":z},bdims=1)
>>> grady_x.shape 
torch.Size([3, 4, 5])
>>> y.shape
torch.Size([3])
>>> torch.allclose(y,f(x,z))
True
>>> torch.allclose(grady_x,2*x*z**2)
True
Source code in agsutil/autograd.py 
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
def gradb(f, x, bkwargs={}, bdims=0, chunk_size=None):
    r"""
    Batched `torch.func.grad` and function evaluation 

    Args:
        f (callable): Function to compute `torch.func.grad` of. 
        x (Union[torch.Tensor,Tuple]): (batched) `Torch.Tensor` items whose gradients will be computed.
        bkwargs (dict): (batched) `Torch.Tensor` items whose gradients will not be computed.
        bdims (int): number of batch dimensions 
        chunk_size (int): to be passed into `torch.func.vmap`. 

    Returns: 
        gradys (tuple): `torch.Tensor` gradients, one with respect to each item in `x`. 
        y (torch.Tensor): Function evaluations. 

    Examples: 
        >>> torch.set_default_dtype(torch.float64)
        >>> rng = torch.Generator().manual_seed(7)

        >>> f = lambda x: (x**2).sum(-1)
        >>> x = torch.rand(5,generator=rng) 
        >>> (grady,),y = gradb(f,x)
        >>> grady.shape 
        torch.Size([5])
        >>> y.shape 
        torch.Size([])
        >>> torch.allclose(y,f(x))
        True
        >>> torch.allclose(grady,2*x)
        True

        >>> f = lambda x,z: (x**2*z**2).sum(-1)
        >>> x = torch.rand(5,generator=rng) 
        >>> z = torch.rand(5,generator=rng) 
        >>> (grady_x,grady_z),y = gradb(f,(x,z))
        >>> grady_x.shape 
        torch.Size([5])
        >>> grady_z.shape 
        torch.Size([5])
        >>> y.shape
        torch.Size([])
        >>> torch.allclose(y,f(x,z))
        True
        >>> torch.allclose(grady_x,2*x*z**2)
        True
        >>> torch.allclose(grady_z,x**2*2*z)
        True

        >>> f = lambda x: (x**2).sum(-1)
        >>> x = torch.rand(3,4,5,generator=rng) 
        >>> (grady,),y = gradb(f,x,bdims=2)
        >>> grady.shape 
        torch.Size([3, 4, 5])
        >>> y.shape 
        torch.Size([3, 4])
        >>> torch.allclose(y,f(x))
        True
        >>> torch.allclose(grady,2*x)
        True

        >>> f = lambda x,z: (x**2*z**2).sum(-1)
        >>> x = torch.rand(3,4,5,generator=rng) 
        >>> z = torch.rand(3,4,5,generator=rng) 
        >>> (grady_x,grady_z),y = gradb(f,(x,z),bdims=2)
        >>> grady_x.shape 
        torch.Size([3, 4, 5])
        >>> grady_z.shape 
        torch.Size([3, 4, 5])
        >>> y.shape
        torch.Size([3, 4])
        >>> torch.allclose(y,f(x,z))
        True
        >>> torch.allclose(grady_x,2*x*z**2)
        True
        >>> torch.allclose(grady_z,x**2*2*z)
        True

        >>> f = lambda x,z: (x**2*z**2).sum((-2,-1))
        >>> x = torch.rand(3,4,5,generator=rng) 
        >>> z = torch.rand(3,4,5,generator=rng) 
        >>> (grady_x,),y = gradb(f,x,bkwargs={"z":z},bdims=1)
        >>> grady_x.shape 
        torch.Size([3, 4, 5])
        >>> y.shape
        torch.Size([3])
        >>> torch.allclose(y,f(x,z))
        True
        >>> torch.allclose(grady_x,2*x*z**2)
        True
    """
    if isinstance(x,torch.Tensor): x = (x,)
    lenx = len(x)
    assert len(x)>=1
    batch_shape = list(x[0].shape[:bdims])
    for i,xi in enumerate(x): assert list(xi.shape[:bdims])==batch_shape, "input %d has shape = %s, but expected first dims to match batch_shape = %s"%(i,list(xi.shape),list(batch_shape))
    for k,v in bkwargs.items(): assert list(v.shape[:bdims])==batch_shape, "bkwargs['%s'].shape = %s, but expected first dims to match batch_shape = %s"%(k,list(v.shape),list(batch_shape))
    len_bkwargs = len(bkwargs) 
    bkwargs_keys = list(bkwargs.keys())
    def fwrap(*inputs):
        x = inputs[:lenx]
        bkwargs_vals = inputs[lenx:(lenx+len_bkwargs)]
        bkwargs = {bkwargs_keys[l]:bkwargs_vals[l] for l in range(len_bkwargs)}
        y = f(*x,**bkwargs)
        return y,y
    gradfwrap = torch.func.grad(fwrap,argnums=tuple(i for i in range(len(x))),has_aux=True)
    gradfwrapvec = torch.vmap(gradfwrap,in_dims=(0,)*(lenx+len_bkwargs),chunk_size=chunk_size)
    x_input = [xi.flatten(end_dim=bdims-1) if bdims>0 else xi[None,...] for xi in x]
    bkwargs_vals_input = [bkwargs[key].flatten(end_dim=bdims-1) if bdims>0 else bkwargs[key][None,...] for key in bkwargs_keys]
    gradys,y = gradfwrapvec(*x_input,*bkwargs_vals_input)
    if bdims==0:
        return tuple(grady[0] for grady in gradys),y[0]
    else:
        return tuple(grady.reshape(batch_shape+list(grady.shape[1:])) for grady in gradys),y.reshape(batch_shape+list(y.shape[1:]))

jacfwdb 

jacfwdb(f, x, bkwargs={}, bdims=0, chunk_size=None)

Batched torch.func.jacfwd with function evaluation

Parameters:

Name Type Description Default
f callable

Function to compute torch.func.jacfwd of.

required
x Union[Tensor, Tuple]

(batched) Torch.Tensor items whose jacobians will be computed.

 required
bkwargs dict

(batched) Torch.Tensor items whose jacobians will not be computed.

 {}
bdims int

number of batch dimensions

0
chunk_size int

to be passed into torch.func.vmap.

None

Returns:

Name Type Description
jacys tuple

torch.Tensor jacobians, one with respect to each item in x.

y Tensor

Function evaluations.

Examples:

>>> torch.set_default_dtype(torch.float64)
>>> rng = torch.Generator().manual_seed(7)

>>> f = lambda x: (x[...,None]**torch.arange(2,5)).sum(-2)
>>> x = torch.rand(5,generator=rng) 
>>> (jacy,),y = jacfwdb(f,x)
>>> jacy.shape 
torch.Size([3, 5])
>>> y.shape 
torch.Size([3])
>>> torch.allclose(y,f(x))
True
>>> torch.allclose(jacy,torch.arange(2,5)[:,None]*x**torch.arange(1,4)[:,None])
True

>>> f = lambda x,z: (x[...,None,None]**torch.arange(2,4)[:,None]*z[...,None,None]**torch.arange(3,5)[None,:]).sum(-3)
>>> x = torch.rand(5,generator=rng) 
>>> z = torch.rand(5,generator=rng) 
>>> (jacy_x,jacy_z),y = jacfwdb(f,(x,z))
>>> jacy_x.shape 
torch.Size([2, 2, 5])
>>> jacy_z.shape 
torch.Size([2, 2, 5])
>>> y.shape
torch.Size([2, 2])
>>> torch.allclose(y,f(x,z))
True
>>> torch.allclose(jacy_x,torch.arange(2,4)[:,None,None]*x**torch.arange(1,3)[:,None,None]*z**torch.arange(3,5)[None,:,None])
True
>>> torch.allclose(jacy_z,x**torch.arange(2,4)[:,None,None]*torch.arange(3,5)[None,:,None]*z**torch.arange(2,4)[None,:,None])
True

>>> f = lambda x: (x[...,None]**torch.arange(2,5)).sum(-2)
>>> x = torch.rand(3,4,5,generator=rng) 
>>> (jacy,),y = jacfwdb(f,x,bdims=2)
>>> jacy.shape 
torch.Size([3, 4, 3, 5])
>>> y.shape 
torch.Size([3, 4, 3])
>>> torch.allclose(y,f(x))
True
>>> torch.allclose(jacy,torch.arange(2,5)[:,None]*x[...,None,:]**torch.arange(1,4)[:,None])
True

>>> f = lambda x,z: (x[...,None,None]**torch.arange(2,4)[:,None]*z[...,None,None]**torch.arange(3,5)[None,:]).sum(-3)
>>> x = torch.rand(3,4,5,generator=rng) 
>>> z = torch.rand(3,4,5,generator=rng) 
>>> (jacy_x,jacy_z),y = jacfwdb(f,(x,z),bdims=2)
>>> jacy_x.shape 
torch.Size([3, 4, 2, 2, 5])
>>> jacy_z.shape 
torch.Size([3, 4, 2, 2, 5])
>>> y.shape
torch.Size([3, 4, 2, 2])
>>> torch.allclose(y,f(x,z))
True
>>> torch.allclose(jacy_x,torch.arange(2,4)[:,None,None]*x[...,None,None,:]**torch.arange(1,3)[:,None,None]*z[...,None,None,:]**torch.arange(3,5)[None,:,None])
True
>>> torch.allclose(jacy_z,x[...,None,None,:]**torch.arange(2,4)[:,None,None]*torch.arange(3,5)[None,:,None]*z[...,None,None,:]**torch.arange(2,4)[None,:,None])
True

>>> f = lambda x,z: (x[...,None,None]**torch.arange(2,4)[:,None]*z[...,None,None]**torch.arange(3,5)[None,:]).sum((-4,-3))
>>> x = torch.rand(3,4,5,6,generator=rng) 
>>> z = torch.rand(3,4,5,6,generator=rng) 
>>> (jacy_x,),y = jacfwdb(f,x,bkwargs={"z":z},bdims=2)
>>> jacy_x.shape 
torch.Size([3, 4, 2, 2, 5, 6])
>>> y.shape
torch.Size([3, 4, 2, 2])
>>> torch.allclose(y,f(x,z))
True
>>> torch.allclose(jacy_x,torch.arange(2,4)[:,None,None,None]*x[...,None,None,:,:]**torch.arange(1,3)[:,None,None,None]*z[...,None,None,:,:]**torch.arange(3,5)[None,:,None,None])
True

>>> def f(x):
...     y = (x[...,None]**torch.arange(2,4)).sum(-2)
...     u = (x[...,None]**torch.arange(3,5)).sum(-2)
...     v = (x[...,None]**torch.arange(4,6)).sum(-2)
...     return y,u,v
>>> x = torch.rand(5,generator=rng) 
>>> ((jacy_x,),(jacu_x,),(jacv_x,)),(y,u,v) = jacfwdb(f,x)
>>> jacy_x.shape
torch.Size([2, 5])
>>> jacu_x.shape
torch.Size([2, 5])
>>> jacv_x.shape
torch.Size([2, 5])
>>> y.shape 
torch.Size([2])
>>> u.shape 
torch.Size([2])
>>> v.shape
torch.Size([2])

>>> def f(x,z):
...     y = (x[...,None,None]**torch.arange(2,4)[:,None]*z[...,None,None]**torch.arange(3,5)[None,:]).sum(-3)
...     u = (x[...,None,None]**torch.arange(3,5)[:,None]*z[...,None,None]**torch.arange(4,6)[None,:]).sum(-3)
...     v = (x[...,None,None]**torch.arange(4,6)[:,None]*z[...,None,None]**torch.arange(5,7)[None,:]).sum(-3)
...     return y,u,v
>>> x = torch.rand(3,4,5,generator=rng) 
>>> z = torch.rand(3,4,5,generator=rng) 
>>> ((jacy_x,jacy_z),(jacu_x,jacu_z),(jacv_x,jacv_z)),(y,u,v) = jacfwdb(f,(x,z),bdims=2)
>>> jacy_x.shape
torch.Size([3, 4, 2, 2, 5])
>>> jacy_z.shape
torch.Size([3, 4, 2, 2, 5])
>>> jacu_x.shape
torch.Size([3, 4, 2, 2, 5])
>>> jacu_z.shape
torch.Size([3, 4, 2, 2, 5])
>>> jacv_x.shape
torch.Size([3, 4, 2, 2, 5])
>>> jacv_z.shape
torch.Size([3, 4, 2, 2, 5])
>>> y.shape 
torch.Size([3, 4, 2, 2])
>>> u.shape 
torch.Size([3, 4, 2, 2])
>>> v.shape
torch.Size([3, 4, 2, 2])
Source code in agsutil/autograd.py 
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
def jacfwdb(f, x, bkwargs={}, bdims=0, chunk_size=None):
    r"""
    Batched `torch.func.jacfwd` with function evaluation 

    Args:
        f (callable): Function to compute `torch.func.jacfwd` of. 
        x (Union[torch.Tensor,Tuple]): (batched) `Torch.Tensor` items whose jacobians will be computed.
        bkwargs (dict): (batched) `Torch.Tensor` items whose jacobians will not be computed.
        bdims (int): number of batch dimensions 
        chunk_size (int): to be passed into `torch.func.vmap`. 

    Returns: 
        jacys (tuple): `torch.Tensor` jacobians, one with respect to each item in `x`. 
        y (torch.Tensor): Function evaluations. 

    Examples: 
        >>> torch.set_default_dtype(torch.float64)
        >>> rng = torch.Generator().manual_seed(7)

        >>> f = lambda x: (x[...,None]**torch.arange(2,5)).sum(-2)
        >>> x = torch.rand(5,generator=rng) 
        >>> (jacy,),y = jacfwdb(f,x)
        >>> jacy.shape 
        torch.Size([3, 5])
        >>> y.shape 
        torch.Size([3])
        >>> torch.allclose(y,f(x))
        True
        >>> torch.allclose(jacy,torch.arange(2,5)[:,None]*x**torch.arange(1,4)[:,None])
        True

        >>> f = lambda x,z: (x[...,None,None]**torch.arange(2,4)[:,None]*z[...,None,None]**torch.arange(3,5)[None,:]).sum(-3)
        >>> x = torch.rand(5,generator=rng) 
        >>> z = torch.rand(5,generator=rng) 
        >>> (jacy_x,jacy_z),y = jacfwdb(f,(x,z))
        >>> jacy_x.shape 
        torch.Size([2, 2, 5])
        >>> jacy_z.shape 
        torch.Size([2, 2, 5])
        >>> y.shape
        torch.Size([2, 2])
        >>> torch.allclose(y,f(x,z))
        True
        >>> torch.allclose(jacy_x,torch.arange(2,4)[:,None,None]*x**torch.arange(1,3)[:,None,None]*z**torch.arange(3,5)[None,:,None])
        True
        >>> torch.allclose(jacy_z,x**torch.arange(2,4)[:,None,None]*torch.arange(3,5)[None,:,None]*z**torch.arange(2,4)[None,:,None])
        True

        >>> f = lambda x: (x[...,None]**torch.arange(2,5)).sum(-2)
        >>> x = torch.rand(3,4,5,generator=rng) 
        >>> (jacy,),y = jacfwdb(f,x,bdims=2)
        >>> jacy.shape 
        torch.Size([3, 4, 3, 5])
        >>> y.shape 
        torch.Size([3, 4, 3])
        >>> torch.allclose(y,f(x))
        True
        >>> torch.allclose(jacy,torch.arange(2,5)[:,None]*x[...,None,:]**torch.arange(1,4)[:,None])
        True

        >>> f = lambda x,z: (x[...,None,None]**torch.arange(2,4)[:,None]*z[...,None,None]**torch.arange(3,5)[None,:]).sum(-3)
        >>> x = torch.rand(3,4,5,generator=rng) 
        >>> z = torch.rand(3,4,5,generator=rng) 
        >>> (jacy_x,jacy_z),y = jacfwdb(f,(x,z),bdims=2)
        >>> jacy_x.shape 
        torch.Size([3, 4, 2, 2, 5])
        >>> jacy_z.shape 
        torch.Size([3, 4, 2, 2, 5])
        >>> y.shape
        torch.Size([3, 4, 2, 2])
        >>> torch.allclose(y,f(x,z))
        True
        >>> torch.allclose(jacy_x,torch.arange(2,4)[:,None,None]*x[...,None,None,:]**torch.arange(1,3)[:,None,None]*z[...,None,None,:]**torch.arange(3,5)[None,:,None])
        True
        >>> torch.allclose(jacy_z,x[...,None,None,:]**torch.arange(2,4)[:,None,None]*torch.arange(3,5)[None,:,None]*z[...,None,None,:]**torch.arange(2,4)[None,:,None])
        True

        >>> f = lambda x,z: (x[...,None,None]**torch.arange(2,4)[:,None]*z[...,None,None]**torch.arange(3,5)[None,:]).sum((-4,-3))
        >>> x = torch.rand(3,4,5,6,generator=rng) 
        >>> z = torch.rand(3,4,5,6,generator=rng) 
        >>> (jacy_x,),y = jacfwdb(f,x,bkwargs={"z":z},bdims=2)
        >>> jacy_x.shape 
        torch.Size([3, 4, 2, 2, 5, 6])
        >>> y.shape
        torch.Size([3, 4, 2, 2])
        >>> torch.allclose(y,f(x,z))
        True
        >>> torch.allclose(jacy_x,torch.arange(2,4)[:,None,None,None]*x[...,None,None,:,:]**torch.arange(1,3)[:,None,None,None]*z[...,None,None,:,:]**torch.arange(3,5)[None,:,None,None])
        True

        >>> def f(x):
        ...     y = (x[...,None]**torch.arange(2,4)).sum(-2)
        ...     u = (x[...,None]**torch.arange(3,5)).sum(-2)
        ...     v = (x[...,None]**torch.arange(4,6)).sum(-2)
        ...     return y,u,v
        >>> x = torch.rand(5,generator=rng) 
        >>> ((jacy_x,),(jacu_x,),(jacv_x,)),(y,u,v) = jacfwdb(f,x)
        >>> jacy_x.shape
        torch.Size([2, 5])
        >>> jacu_x.shape
        torch.Size([2, 5])
        >>> jacv_x.shape
        torch.Size([2, 5])
        >>> y.shape 
        torch.Size([2])
        >>> u.shape 
        torch.Size([2])
        >>> v.shape
        torch.Size([2])

        >>> def f(x,z):
        ...     y = (x[...,None,None]**torch.arange(2,4)[:,None]*z[...,None,None]**torch.arange(3,5)[None,:]).sum(-3)
        ...     u = (x[...,None,None]**torch.arange(3,5)[:,None]*z[...,None,None]**torch.arange(4,6)[None,:]).sum(-3)
        ...     v = (x[...,None,None]**torch.arange(4,6)[:,None]*z[...,None,None]**torch.arange(5,7)[None,:]).sum(-3)
        ...     return y,u,v
        >>> x = torch.rand(3,4,5,generator=rng) 
        >>> z = torch.rand(3,4,5,generator=rng) 
        >>> ((jacy_x,jacy_z),(jacu_x,jacu_z),(jacv_x,jacv_z)),(y,u,v) = jacfwdb(f,(x,z),bdims=2)
        >>> jacy_x.shape
        torch.Size([3, 4, 2, 2, 5])
        >>> jacy_z.shape
        torch.Size([3, 4, 2, 2, 5])
        >>> jacu_x.shape
        torch.Size([3, 4, 2, 2, 5])
        >>> jacu_z.shape
        torch.Size([3, 4, 2, 2, 5])
        >>> jacv_x.shape
        torch.Size([3, 4, 2, 2, 5])
        >>> jacv_z.shape
        torch.Size([3, 4, 2, 2, 5])
        >>> y.shape 
        torch.Size([3, 4, 2, 2])
        >>> u.shape 
        torch.Size([3, 4, 2, 2])
        >>> v.shape
        torch.Size([3, 4, 2, 2])
    """
    if isinstance(x,torch.Tensor): x = (x,)
    lenx = len(x) 
    assert len(x)>=1
    batch_shape = list(x[0].shape[:bdims])
    for i,xi in enumerate(x): assert list(xi.shape[:bdims])==batch_shape, "input %d has shape = %s, but expected first dims to match batch_shape = %s"%(i,list(xi.shape),list(batch_shape))
    for k,v in bkwargs.items(): assert list(v.shape[:bdims])==batch_shape, "bkwargs['%s'].shape = %s, but expected first dims to match batch_shape = %s"%(k,list(v.shape),list(batch_shape))
    len_bkwargs = len(bkwargs) 
    bkwargs_keys = list(bkwargs.keys())
    def fwrap(*inputs):
        x = inputs[:lenx]
        bkwargs_vals = inputs[lenx:(lenx+len_bkwargs)]
        bkwargs = {bkwargs_keys[l]:bkwargs_vals[l] for l in range(len_bkwargs)}
        y = f(*x,**bkwargs)
        return y,y
    jacfwrap = torch.func.jacfwd(fwrap,argnums=tuple(i for i in range(len(x))),has_aux=True)
    jacfwrapvec = torch.vmap(jacfwrap,in_dims=(0,)*(lenx+len_bkwargs),chunk_size=chunk_size)
    x_input = [xi.flatten(end_dim=bdims-1) if bdims>0 else xi[None,...] for xi in x]
    bkwargs_vals_input = [bkwargs[key].flatten(end_dim=bdims-1) if bdims>0 else bkwargs[key][None,...] for key in bkwargs_keys]
    jacys,y = jacfwrapvec(*x_input,*bkwargs_vals_input)
    if isinstance(y,torch.Tensor):
        if bdims==0:
            return tuple(jacy[0] for jacy in jacys),y[0]
        else:
            return tuple(jacy.reshape(batch_shape+list(jacy.shape[1:])) for jacy in jacys),y.reshape(batch_shape+list(y.shape[1:]))
    else:
        if bdims==0:
            return tuple(tuple(jacyk[0] for jacyk in jacy) for jacy in jacys),tuple(yk[0] for yk in y)
        else:
            return tuple(tuple(jacyk.reshape(batch_shape+list(jacyk.shape[1:])) for jacyk in jacy) for jacy in jacys),tuple(yk.reshape(batch_shape+list(yk.shape[1:])) for yk in y)

jacrevb 

jacrevb(f, x, bkwargs={}, bdims=0, chunk_size=None)

Batched torch.func.jacrev with function evaluation

Parameters:

Name Type Description Default
f callable

Function to compute torch.func.jacrev of.

required
x Tuple

(Union[torch.Tensor,Tuple]) Torch.Tensor items whose jacobians will be computed.

 required
bkwargs dict

(batched) Torch.Tensor items whose jacobians will not be computed.

 {}
bdims int

number of batch dimensions

0
chunk_size int

to be passed into torch.func.vmap.

None

Returns:

Name Type Description
jacys tuple

torch.Tensor jacobians, one with respect to each item in x.

y Tensor

Function evaluations.

Examples:

>>> torch.set_default_dtype(torch.float64)
>>> rng = torch.Generator().manual_seed(7)

>>> f = lambda x: (x[...,None]**torch.arange(2,5)).sum(-2)
>>> x = torch.rand(5,generator=rng) 
>>> (jacy,),y = jacrevb(f,x)
>>> jacy.shape 
torch.Size([3, 5])
>>> y.shape 
torch.Size([3])
>>> torch.allclose(y,f(x))
True
>>> torch.allclose(jacy,torch.arange(2,5)[:,None]*x**torch.arange(1,4)[:,None])
True

>>> f = lambda x,z: (x[...,None,None]**torch.arange(2,4)[:,None]*z[...,None,None]**torch.arange(3,5)[None,:]).sum(-3)
>>> x = torch.rand(5,generator=rng) 
>>> z = torch.rand(5,generator=rng) 
>>> (jacy_x,jacy_z),y = jacrevb(f,(x,z))
>>> jacy_x.shape 
torch.Size([2, 2, 5])
>>> jacy_z.shape 
torch.Size([2, 2, 5])
>>> y.shape
torch.Size([2, 2])
>>> torch.allclose(y,f(x,z))
True
>>> torch.allclose(jacy_x,torch.arange(2,4)[:,None,None]*x**torch.arange(1,3)[:,None,None]*z**torch.arange(3,5)[None,:,None])
True
>>> torch.allclose(jacy_z,x**torch.arange(2,4)[:,None,None]*torch.arange(3,5)[None,:,None]*z**torch.arange(2,4)[None,:,None])
True

>>> f = lambda x: (x[...,None]**torch.arange(2,5)).sum(-2)
>>> x = torch.rand(3,4,5,generator=rng) 
>>> (jacy,),y = jacrevb(f,x,bdims=2)
>>> jacy.shape 
torch.Size([3, 4, 3, 5])
>>> y.shape 
torch.Size([3, 4, 3])
>>> torch.allclose(y,f(x))
True
>>> torch.allclose(jacy,torch.arange(2,5)[:,None]*x[...,None,:]**torch.arange(1,4)[:,None])
True

>>> f = lambda x,z: (x[...,None,None]**torch.arange(2,4)[:,None]*z[...,None,None]**torch.arange(3,5)[None,:]).sum(-3)
>>> x = torch.rand(3,4,5,generator=rng) 
>>> z = torch.rand(3,4,5,generator=rng) 
>>> (jacy_x,jacy_z),y = jacrevb(f,(x,z),bdims=2)
>>> jacy_x.shape 
torch.Size([3, 4, 2, 2, 5])
>>> jacy_z.shape 
torch.Size([3, 4, 2, 2, 5])
>>> y.shape
torch.Size([3, 4, 2, 2])
>>> torch.allclose(y,f(x,z))
True
>>> torch.allclose(jacy_x,torch.arange(2,4)[:,None,None]*x[...,None,None,:]**torch.arange(1,3)[:,None,None]*z[...,None,None,:]**torch.arange(3,5)[None,:,None])
True
>>> torch.allclose(jacy_z,x[...,None,None,:]**torch.arange(2,4)[:,None,None]*torch.arange(3,5)[None,:,None]*z[...,None,None,:]**torch.arange(2,4)[None,:,None])
True

>>> f = lambda x,z: (x[...,None,None]**torch.arange(2,4)[:,None]*z[...,None,None]**torch.arange(3,5)[None,:]).sum((-4,-3))
>>> x = torch.rand(3,4,5,6,generator=rng) 
>>> z = torch.rand(3,4,5,6,generator=rng) 
>>> (jacy_x,),y = jacrevb(f,x,bkwargs={"z":z},bdims=2)
>>> jacy_x.shape 
torch.Size([3, 4, 2, 2, 5, 6])
>>> y.shape
torch.Size([3, 4, 2, 2])
>>> torch.allclose(y,f(x,z))
True
>>> torch.allclose(jacy_x,torch.arange(2,4)[:,None,None,None]*x[...,None,None,:,:]**torch.arange(1,3)[:,None,None,None]*z[...,None,None,:,:]**torch.arange(3,5)[None,:,None,None])
True

>>> def f(x):
...     y = (x[...,None]**torch.arange(2,4)).sum(-2)
...     u = (x[...,None]**torch.arange(3,5)).sum(-2)
...     v = (x[...,None]**torch.arange(4,6)).sum(-2)
...     return y,u,v
>>> x = torch.rand(5,generator=rng) 
>>> ((jacy_x,),(jacu_x,),(jacv_x,)),(y,u,v) = jacrevb(f,x)
>>> jacy_x.shape
torch.Size([2, 5])
>>> jacu_x.shape
torch.Size([2, 5])
>>> jacv_x.shape
torch.Size([2, 5])
>>> y.shape 
torch.Size([2])
>>> u.shape 
torch.Size([2])
>>> v.shape
torch.Size([2])

>>> def f(x,z):
...     y = (x[...,None,None]**torch.arange(2,4)[:,None]*z[...,None,None]**torch.arange(3,5)[None,:]).sum(-3)
...     u = (x[...,None,None]**torch.arange(3,5)[:,None]*z[...,None,None]**torch.arange(4,6)[None,:]).sum(-3)
...     v = (x[...,None,None]**torch.arange(4,6)[:,None]*z[...,None,None]**torch.arange(5,7)[None,:]).sum(-3)
...     return y,u,v
>>> x = torch.rand(3,4,5,generator=rng) 
>>> z = torch.rand(3,4,5,generator=rng) 
>>> ((jacy_x,jacy_z),(jacu_x,jacu_z),(jacv_x,jacv_z)),(y,u,v) = jacrevb(f,(x,z),bdims=2)
>>> jacy_x.shape
torch.Size([3, 4, 2, 2, 5])
>>> jacy_z.shape
torch.Size([3, 4, 2, 2, 5])
>>> jacu_x.shape
torch.Size([3, 4, 2, 2, 5])
>>> jacu_z.shape
torch.Size([3, 4, 2, 2, 5])
>>> jacv_x.shape
torch.Size([3, 4, 2, 2, 5])
>>> jacv_z.shape
torch.Size([3, 4, 2, 2, 5])
>>> y.shape 
torch.Size([3, 4, 2, 2])
>>> u.shape 
torch.Size([3, 4, 2, 2])
>>> v.shape
torch.Size([3, 4, 2, 2])
Source code in agsutil/autograd.py 
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
def jacrevb(f, x, bkwargs={}, bdims=0, chunk_size=None):
    r"""
    Batched `torch.func.jacrev` with function evaluation 

    Args:
        f (callable): Function to compute `torch.func.jacrev` of. 
        x (Tuple): (Union[torch.Tensor,Tuple]) `Torch.Tensor` items whose jacobians will be computed.
        bkwargs (dict): (batched) `Torch.Tensor` items whose jacobians will not be computed.
        bdims (int): number of batch dimensions 
        chunk_size (int): to be passed into `torch.func.vmap`. 

    Returns: 
        jacys (tuple): `torch.Tensor` jacobians, one with respect to each item in `x`. 
        y (torch.Tensor): Function evaluations. 

    Examples: 
        >>> torch.set_default_dtype(torch.float64)
        >>> rng = torch.Generator().manual_seed(7)

        >>> f = lambda x: (x[...,None]**torch.arange(2,5)).sum(-2)
        >>> x = torch.rand(5,generator=rng) 
        >>> (jacy,),y = jacrevb(f,x)
        >>> jacy.shape 
        torch.Size([3, 5])
        >>> y.shape 
        torch.Size([3])
        >>> torch.allclose(y,f(x))
        True
        >>> torch.allclose(jacy,torch.arange(2,5)[:,None]*x**torch.arange(1,4)[:,None])
        True

        >>> f = lambda x,z: (x[...,None,None]**torch.arange(2,4)[:,None]*z[...,None,None]**torch.arange(3,5)[None,:]).sum(-3)
        >>> x = torch.rand(5,generator=rng) 
        >>> z = torch.rand(5,generator=rng) 
        >>> (jacy_x,jacy_z),y = jacrevb(f,(x,z))
        >>> jacy_x.shape 
        torch.Size([2, 2, 5])
        >>> jacy_z.shape 
        torch.Size([2, 2, 5])
        >>> y.shape
        torch.Size([2, 2])
        >>> torch.allclose(y,f(x,z))
        True
        >>> torch.allclose(jacy_x,torch.arange(2,4)[:,None,None]*x**torch.arange(1,3)[:,None,None]*z**torch.arange(3,5)[None,:,None])
        True
        >>> torch.allclose(jacy_z,x**torch.arange(2,4)[:,None,None]*torch.arange(3,5)[None,:,None]*z**torch.arange(2,4)[None,:,None])
        True

        >>> f = lambda x: (x[...,None]**torch.arange(2,5)).sum(-2)
        >>> x = torch.rand(3,4,5,generator=rng) 
        >>> (jacy,),y = jacrevb(f,x,bdims=2)
        >>> jacy.shape 
        torch.Size([3, 4, 3, 5])
        >>> y.shape 
        torch.Size([3, 4, 3])
        >>> torch.allclose(y,f(x))
        True
        >>> torch.allclose(jacy,torch.arange(2,5)[:,None]*x[...,None,:]**torch.arange(1,4)[:,None])
        True

        >>> f = lambda x,z: (x[...,None,None]**torch.arange(2,4)[:,None]*z[...,None,None]**torch.arange(3,5)[None,:]).sum(-3)
        >>> x = torch.rand(3,4,5,generator=rng) 
        >>> z = torch.rand(3,4,5,generator=rng) 
        >>> (jacy_x,jacy_z),y = jacrevb(f,(x,z),bdims=2)
        >>> jacy_x.shape 
        torch.Size([3, 4, 2, 2, 5])
        >>> jacy_z.shape 
        torch.Size([3, 4, 2, 2, 5])
        >>> y.shape
        torch.Size([3, 4, 2, 2])
        >>> torch.allclose(y,f(x,z))
        True
        >>> torch.allclose(jacy_x,torch.arange(2,4)[:,None,None]*x[...,None,None,:]**torch.arange(1,3)[:,None,None]*z[...,None,None,:]**torch.arange(3,5)[None,:,None])
        True
        >>> torch.allclose(jacy_z,x[...,None,None,:]**torch.arange(2,4)[:,None,None]*torch.arange(3,5)[None,:,None]*z[...,None,None,:]**torch.arange(2,4)[None,:,None])
        True

        >>> f = lambda x,z: (x[...,None,None]**torch.arange(2,4)[:,None]*z[...,None,None]**torch.arange(3,5)[None,:]).sum((-4,-3))
        >>> x = torch.rand(3,4,5,6,generator=rng) 
        >>> z = torch.rand(3,4,5,6,generator=rng) 
        >>> (jacy_x,),y = jacrevb(f,x,bkwargs={"z":z},bdims=2)
        >>> jacy_x.shape 
        torch.Size([3, 4, 2, 2, 5, 6])
        >>> y.shape
        torch.Size([3, 4, 2, 2])
        >>> torch.allclose(y,f(x,z))
        True
        >>> torch.allclose(jacy_x,torch.arange(2,4)[:,None,None,None]*x[...,None,None,:,:]**torch.arange(1,3)[:,None,None,None]*z[...,None,None,:,:]**torch.arange(3,5)[None,:,None,None])
        True

        >>> def f(x):
        ...     y = (x[...,None]**torch.arange(2,4)).sum(-2)
        ...     u = (x[...,None]**torch.arange(3,5)).sum(-2)
        ...     v = (x[...,None]**torch.arange(4,6)).sum(-2)
        ...     return y,u,v
        >>> x = torch.rand(5,generator=rng) 
        >>> ((jacy_x,),(jacu_x,),(jacv_x,)),(y,u,v) = jacrevb(f,x)
        >>> jacy_x.shape
        torch.Size([2, 5])
        >>> jacu_x.shape
        torch.Size([2, 5])
        >>> jacv_x.shape
        torch.Size([2, 5])
        >>> y.shape 
        torch.Size([2])
        >>> u.shape 
        torch.Size([2])
        >>> v.shape
        torch.Size([2])

        >>> def f(x,z):
        ...     y = (x[...,None,None]**torch.arange(2,4)[:,None]*z[...,None,None]**torch.arange(3,5)[None,:]).sum(-3)
        ...     u = (x[...,None,None]**torch.arange(3,5)[:,None]*z[...,None,None]**torch.arange(4,6)[None,:]).sum(-3)
        ...     v = (x[...,None,None]**torch.arange(4,6)[:,None]*z[...,None,None]**torch.arange(5,7)[None,:]).sum(-3)
        ...     return y,u,v
        >>> x = torch.rand(3,4,5,generator=rng) 
        >>> z = torch.rand(3,4,5,generator=rng) 
        >>> ((jacy_x,jacy_z),(jacu_x,jacu_z),(jacv_x,jacv_z)),(y,u,v) = jacrevb(f,(x,z),bdims=2)
        >>> jacy_x.shape
        torch.Size([3, 4, 2, 2, 5])
        >>> jacy_z.shape
        torch.Size([3, 4, 2, 2, 5])
        >>> jacu_x.shape
        torch.Size([3, 4, 2, 2, 5])
        >>> jacu_z.shape
        torch.Size([3, 4, 2, 2, 5])
        >>> jacv_x.shape
        torch.Size([3, 4, 2, 2, 5])
        >>> jacv_z.shape
        torch.Size([3, 4, 2, 2, 5])
        >>> y.shape 
        torch.Size([3, 4, 2, 2])
        >>> u.shape 
        torch.Size([3, 4, 2, 2])
        >>> v.shape
        torch.Size([3, 4, 2, 2])
    """
    if isinstance(x,torch.Tensor): x = (x,)
    lenx = len(x) 
    assert len(x)>=1
    batch_shape = list(x[0].shape[:bdims])
    for i,xi in enumerate(x): assert list(xi.shape[:bdims])==batch_shape, "input %d has shape = %s, but expected first dims to match batch_shape = %s"%(i,list(xi.shape),list(batch_shape))
    for k,v in bkwargs.items(): assert list(v.shape[:bdims])==batch_shape, "bkwargs['%s'].shape = %s, but expected first dims to match batch_shape = %s"%(k,list(v.shape),list(batch_shape))
    len_bkwargs = len(bkwargs) 
    bkwargs_keys = list(bkwargs.keys())
    def fwrap(*inputs):
        x = inputs[:lenx]
        bkwargs_vals = inputs[lenx:(lenx+len_bkwargs)]
        bkwargs = {bkwargs_keys[l]:bkwargs_vals[l] for l in range(len_bkwargs)}
        y = f(*x,**bkwargs)
        return y,y
    jacfwrap = torch.func.jacrev(fwrap,argnums=tuple(i for i in range(len(x))),has_aux=True)
    jacfwrapvec = torch.vmap(jacfwrap,in_dims=(0,)*(lenx+len_bkwargs),chunk_size=chunk_size)
    x_input = [xi.flatten(end_dim=bdims-1) if bdims>0 else xi[None,...] for xi in x]
    bkwargs_vals_input = [bkwargs[key].flatten(end_dim=bdims-1) if bdims>0 else bkwargs[key][None,...] for key in bkwargs_keys]
    jacys,y = jacfwrapvec(*x_input,*bkwargs_vals_input)
    if isinstance(y,torch.Tensor):
        if bdims==0:
            return tuple(jacy[0] for jacy in jacys),y[0]
        else:
            return tuple(jacy.reshape(batch_shape+list(jacy.shape[1:])) for jacy in jacys),y.reshape(batch_shape+list(y.shape[1:]))
    else:
        if bdims==0:
            return tuple(tuple(jacyk[0] for jacyk in jacy) for jacy in jacys),tuple(yk[0] for yk in y)
        else:
            return tuple(tuple(jacyk.reshape(batch_shape+list(jacyk.shape[1:])) for jacyk in jacy) for jacy in jacys),tuple(yk.reshape(batch_shape+list(yk.shape[1:])) for yk in y)

jvpb 

jvpb(f, x, p, bkwargs={}, bdims=0, chunk_size=None)

Batched torch.func.jvp with function evaluation (forward-mode auto-diff)

Parameters:

Name Type Description Default
f callable

Function to compute torch.func.jvp of.

 required
x Union[Tensor, Tuple]

(batched) Torch.Tensor primals.

 required
p Union[Tensor, Tuple]

(batched) Torch.Tensor tangents.

 required
bkwargs dict

(batched) Torch.Tensor items whose jacobians will not be computed.

 {}
bdims int

number of batch dimensions

0
chunk_size int

to be passed into torch.func.vmap.

None

Returns:

Name Type Description
jvp tuple

torch.Tensor jacobian vector products, one with respect to each item in x.

y Tensor

Function evaluations.

Examples:

>>> torch.set_default_dtype(torch.float64)
>>> rng = torch.Generator().manual_seed(7)

>>> f = lambda x: (x[...,None]**torch.arange(2,5)).sum(-2)
>>> x = torch.rand(5,generator=rng)
>>> p = torch.rand(5,generator=rng)
>>> (jvpy,),y = jvpb(f,x,p)
>>> jvpy.shape 
torch.Size([3])
>>> y.shape 
torch.Size([3])
>>> torch.allclose(y,f(x))
True
>>> (jac_x,),_ = jacfwdb(f,x)
>>> torch.allclose(jvpy,jac_x@p)
True

>>> f = lambda x,z: (x[...,:,None,None]**torch.arange(2,5)*z[...,None,:,None]**torch.arange(1,4)).sum((-3,-2))
>>> x = torch.rand(5,generator=rng)
>>> z = torch.rand(4,generator=rng)
>>> p = torch.rand(5,generator=rng)
>>> q = torch.rand(4,generator=rng)
>>> (jvpy,),y = jvpb(f,(x,z),(p,q))
>>> jvpy.shape
torch.Size([3])
>>> y.shape
torch.Size([3])
>>> torch.allclose(y,f(x,z))
True
>>> (jac_x,jac_z),_ = jacfwdb(f,(x,z))
>>> torch.allclose(jvpy,jac_x@p+jac_z@q)
True

>>> f = lambda x: (x[...,None]**torch.arange(2,5)).sum(-2)
>>> x = torch.rand(6,4,5,generator=rng)
>>> p = torch.rand(6,4,5,generator=rng)
>>> (jvpy,),y = jvpb(f,x,p,bdims=2)
>>> jvpy.shape 
torch.Size([6, 4, 3])
>>> y.shape 
torch.Size([6, 4, 3])
>>> torch.allclose(y,f(x))
True
>>> (jac_x,),_ = jacfwdb(f,x,bdims=2)
>>> torch.allclose(jvpy,(jac_x*p[...,None,:]).sum(-1))
True

>>> f = lambda x,z: (x[...,:,None,None]**torch.arange(2,5)*z[...,None,:,None]**torch.arange(1,4)).sum((-3,-2))
>>> x = torch.rand(6,7,5,generator=rng)
>>> z = torch.rand(6,7,4,generator=rng)
>>> p = torch.rand(6,7,5,generator=rng)
>>> q = torch.rand(6,7,4,generator=rng)
>>> (jvpy,),y = jvpb(f,(x,z),(p,q),bdims=2)
>>> jvpy.shape
torch.Size([6, 7, 3])
>>> y.shape
torch.Size([6, 7, 3])
>>> torch.allclose(y,f(x,z))
True
>>> (jac_x,jac_z),_ = jacfwdb(f,(x,z),bdims=2)
>>> torch.allclose(jvpy,(jac_x*p[...,None,:]).sum(-1)+(jac_z*q[...,None,:]).sum(-1))
True

>>> f = lambda x,z: (x[...,:,None,None]**torch.arange(2,5)*z[...,None,:,None]**torch.arange(1,4)).sum((-3,-2))
>>> x = torch.rand(6,7,5,generator=rng)
>>> z = torch.rand(6,7,4,generator=rng)
>>> p = torch.rand(6,7,5,generator=rng)
>>> (jvpy,),y = jvpb(f,x,p,bkwargs={"z":z},bdims=2)
>>> jvpy.shape
torch.Size([6, 7, 3])
>>> y.shape
torch.Size([6, 7, 3])
>>> torch.allclose(y,f(x,z))
True
>>> (jac_x,jac_z),_ = jacfwdb(f,(x,z),bdims=2)
>>> torch.allclose(jvpy,(jac_x*p[...,None,:]).sum(-1))
True

>>> def f(x):
...     y = (x[...,:,None]**torch.arange(2,5)).sum(-2)
...     u = (x[...,:,None]**torch.arange(3,6)).sum(-2)
...     v = (x[...,:,None]**torch.arange(4,7)).sum(-2)
...     return y,u,v
>>> x = torch.rand(5,generator=rng)
>>> p = torch.rand(5,generator=rng)
>>> (jvpy,jvpu,jvpv),(y,u,v) = jvpb(f,x,p)
>>> jvpy.shape
torch.Size([3])
>>> jvpu.shape
torch.Size([3])
>>> jvpv.shape
torch.Size([3])
>>> y.shape
torch.Size([3])
>>> u.shape
torch.Size([3])
>>> v.shape
torch.Size([3])
>>> torch.allclose(y,f(x)[0])
True
>>> torch.allclose(u,f(x)[1])
True
>>> torch.allclose(v,f(x)[2])
True
>>> ((jacy_x,),(jacu_x,),(jacv_x,)),_ = jacfwdb(f,x)
>>> torch.allclose(jvpy,(jacy_x*p[...,None,:]).sum(-1))
True
>>> torch.allclose(jvpu,(jacu_x*p[...,None,:]).sum(-1))
True
>>> torch.allclose(jvpv,(jacv_x*p[...,None,:]).sum(-1))
True

>>> def f(x,z):
...     y = (x[...,:,None,None]**torch.arange(2,5)*z[...,None,:,None]**torch.arange(1,4)).sum((-3,-2))
...     u = (x[...,:,None,None]**torch.arange(3,6)*z[...,None,:,None]**torch.arange(2,5)).sum((-3,-2))
...     v = (x[...,:,None,None]**torch.arange(4,7)*z[...,None,:,None]**torch.arange(3,6)).sum((-3,-2))
...     return y,u,v
>>> x = torch.rand(6,7,5,generator=rng)
>>> z = torch.rand(6,7,4,generator=rng)
>>> p = torch.rand(6,7,5,generator=rng)
>>> q = torch.rand(6,7,4,generator=rng)
>>> (jvpy,jvpu,jvpv),(y,u,v) = jvpb(f,(x,z),(p,q),bdims=2)
>>> jvpy.shape
torch.Size([6, 7, 3])
>>> jvpu.shape
torch.Size([6, 7, 3])
>>> jvpv.shape
torch.Size([6, 7, 3])
>>> y.shape
torch.Size([6, 7, 3])
>>> u.shape
torch.Size([6, 7, 3])
>>> v.shape
torch.Size([6, 7, 3])
>>> torch.allclose(y,f(x,z)[0])
True
>>> torch.allclose(u,f(x,z)[1])
True
>>> torch.allclose(v,f(x,z)[2])
True
>>> ((jacy_x,jacy_z),(jacu_x,jacu_z),(jacv_x,jacv_z)),_ = jacfwdb(f,(x,z),bdims=2)
>>> torch.allclose(jvpy,(jacy_x*p[...,None,:]).sum(-1)+(jacy_z*q[...,None,:]).sum(-1))
True
>>> torch.allclose(jvpu,(jacu_x*p[...,None,:]).sum(-1)+(jacu_z*q[...,None,:]).sum(-1))
True
>>> torch.allclose(jvpv,(jacv_x*p[...,None,:]).sum(-1)+(jacv_z*q[...,None,:]).sum(-1))
True
Source code in agsutil/autograd.py 
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
def jvpb(f, x, p, bkwargs={}, bdims=0, chunk_size=None):
    r"""
    Batched `torch.func.jvp` with function evaluation (forward-mode auto-diff)

    Args:
        f (callable): Function to compute `torch.func.jvp` of.
        x (Union[torch.Tensor,Tuple]): (batched) `Torch.Tensor` primals.
        p (Union[torch.Tensor,Tuple]): (batched) `Torch.Tensor` tangents.
        bkwargs (dict): (batched) `Torch.Tensor` items whose jacobians will not be computed.
        bdims (int): number of batch dimensions 
        chunk_size (int): to be passed into `torch.func.vmap`. 

    Returns: 
        jvp (tuple): `torch.Tensor` jacobian vector products, one with respect to each item in `x`. 
        y (torch.Tensor): Function evaluations. 

    Examples: 
        >>> torch.set_default_dtype(torch.float64)
        >>> rng = torch.Generator().manual_seed(7)

        >>> f = lambda x: (x[...,None]**torch.arange(2,5)).sum(-2)
        >>> x = torch.rand(5,generator=rng)
        >>> p = torch.rand(5,generator=rng)
        >>> (jvpy,),y = jvpb(f,x,p)
        >>> jvpy.shape 
        torch.Size([3])
        >>> y.shape 
        torch.Size([3])
        >>> torch.allclose(y,f(x))
        True
        >>> (jac_x,),_ = jacfwdb(f,x)
        >>> torch.allclose(jvpy,jac_x@p)
        True

        >>> f = lambda x,z: (x[...,:,None,None]**torch.arange(2,5)*z[...,None,:,None]**torch.arange(1,4)).sum((-3,-2))
        >>> x = torch.rand(5,generator=rng)
        >>> z = torch.rand(4,generator=rng)
        >>> p = torch.rand(5,generator=rng)
        >>> q = torch.rand(4,generator=rng)
        >>> (jvpy,),y = jvpb(f,(x,z),(p,q))
        >>> jvpy.shape
        torch.Size([3])
        >>> y.shape
        torch.Size([3])
        >>> torch.allclose(y,f(x,z))
        True
        >>> (jac_x,jac_z),_ = jacfwdb(f,(x,z))
        >>> torch.allclose(jvpy,jac_x@p+jac_z@q)
        True

        >>> f = lambda x: (x[...,None]**torch.arange(2,5)).sum(-2)
        >>> x = torch.rand(6,4,5,generator=rng)
        >>> p = torch.rand(6,4,5,generator=rng)
        >>> (jvpy,),y = jvpb(f,x,p,bdims=2)
        >>> jvpy.shape 
        torch.Size([6, 4, 3])
        >>> y.shape 
        torch.Size([6, 4, 3])
        >>> torch.allclose(y,f(x))
        True
        >>> (jac_x,),_ = jacfwdb(f,x,bdims=2)
        >>> torch.allclose(jvpy,(jac_x*p[...,None,:]).sum(-1))
        True

        >>> f = lambda x,z: (x[...,:,None,None]**torch.arange(2,5)*z[...,None,:,None]**torch.arange(1,4)).sum((-3,-2))
        >>> x = torch.rand(6,7,5,generator=rng)
        >>> z = torch.rand(6,7,4,generator=rng)
        >>> p = torch.rand(6,7,5,generator=rng)
        >>> q = torch.rand(6,7,4,generator=rng)
        >>> (jvpy,),y = jvpb(f,(x,z),(p,q),bdims=2)
        >>> jvpy.shape
        torch.Size([6, 7, 3])
        >>> y.shape
        torch.Size([6, 7, 3])
        >>> torch.allclose(y,f(x,z))
        True
        >>> (jac_x,jac_z),_ = jacfwdb(f,(x,z),bdims=2)
        >>> torch.allclose(jvpy,(jac_x*p[...,None,:]).sum(-1)+(jac_z*q[...,None,:]).sum(-1))
        True

        >>> f = lambda x,z: (x[...,:,None,None]**torch.arange(2,5)*z[...,None,:,None]**torch.arange(1,4)).sum((-3,-2))
        >>> x = torch.rand(6,7,5,generator=rng)
        >>> z = torch.rand(6,7,4,generator=rng)
        >>> p = torch.rand(6,7,5,generator=rng)
        >>> (jvpy,),y = jvpb(f,x,p,bkwargs={"z":z},bdims=2)
        >>> jvpy.shape
        torch.Size([6, 7, 3])
        >>> y.shape
        torch.Size([6, 7, 3])
        >>> torch.allclose(y,f(x,z))
        True
        >>> (jac_x,jac_z),_ = jacfwdb(f,(x,z),bdims=2)
        >>> torch.allclose(jvpy,(jac_x*p[...,None,:]).sum(-1))
        True

        >>> def f(x):
        ...     y = (x[...,:,None]**torch.arange(2,5)).sum(-2)
        ...     u = (x[...,:,None]**torch.arange(3,6)).sum(-2)
        ...     v = (x[...,:,None]**torch.arange(4,7)).sum(-2)
        ...     return y,u,v
        >>> x = torch.rand(5,generator=rng)
        >>> p = torch.rand(5,generator=rng)
        >>> (jvpy,jvpu,jvpv),(y,u,v) = jvpb(f,x,p)
        >>> jvpy.shape
        torch.Size([3])
        >>> jvpu.shape
        torch.Size([3])
        >>> jvpv.shape
        torch.Size([3])
        >>> y.shape
        torch.Size([3])
        >>> u.shape
        torch.Size([3])
        >>> v.shape
        torch.Size([3])
        >>> torch.allclose(y,f(x)[0])
        True
        >>> torch.allclose(u,f(x)[1])
        True
        >>> torch.allclose(v,f(x)[2])
        True
        >>> ((jacy_x,),(jacu_x,),(jacv_x,)),_ = jacfwdb(f,x)
        >>> torch.allclose(jvpy,(jacy_x*p[...,None,:]).sum(-1))
        True
        >>> torch.allclose(jvpu,(jacu_x*p[...,None,:]).sum(-1))
        True
        >>> torch.allclose(jvpv,(jacv_x*p[...,None,:]).sum(-1))
        True

        >>> def f(x,z):
        ...     y = (x[...,:,None,None]**torch.arange(2,5)*z[...,None,:,None]**torch.arange(1,4)).sum((-3,-2))
        ...     u = (x[...,:,None,None]**torch.arange(3,6)*z[...,None,:,None]**torch.arange(2,5)).sum((-3,-2))
        ...     v = (x[...,:,None,None]**torch.arange(4,7)*z[...,None,:,None]**torch.arange(3,6)).sum((-3,-2))
        ...     return y,u,v
        >>> x = torch.rand(6,7,5,generator=rng)
        >>> z = torch.rand(6,7,4,generator=rng)
        >>> p = torch.rand(6,7,5,generator=rng)
        >>> q = torch.rand(6,7,4,generator=rng)
        >>> (jvpy,jvpu,jvpv),(y,u,v) = jvpb(f,(x,z),(p,q),bdims=2)
        >>> jvpy.shape
        torch.Size([6, 7, 3])
        >>> jvpu.shape
        torch.Size([6, 7, 3])
        >>> jvpv.shape
        torch.Size([6, 7, 3])
        >>> y.shape
        torch.Size([6, 7, 3])
        >>> u.shape
        torch.Size([6, 7, 3])
        >>> v.shape
        torch.Size([6, 7, 3])
        >>> torch.allclose(y,f(x,z)[0])
        True
        >>> torch.allclose(u,f(x,z)[1])
        True
        >>> torch.allclose(v,f(x,z)[2])
        True
        >>> ((jacy_x,jacy_z),(jacu_x,jacu_z),(jacv_x,jacv_z)),_ = jacfwdb(f,(x,z),bdims=2)
        >>> torch.allclose(jvpy,(jacy_x*p[...,None,:]).sum(-1)+(jacy_z*q[...,None,:]).sum(-1))
        True
        >>> torch.allclose(jvpu,(jacu_x*p[...,None,:]).sum(-1)+(jacu_z*q[...,None,:]).sum(-1))
        True
        >>> torch.allclose(jvpv,(jacv_x*p[...,None,:]).sum(-1)+(jacv_z*q[...,None,:]).sum(-1))
        True
    """
    if isinstance(x,torch.Tensor): x = (x,)
    if isinstance(p,torch.Tensor): p = (p,)
    lenx = len(x) 
    assert len(x)>=1
    assert len(p)==lenx
    batch_shape = list(x[0].shape[:bdims])
    for i,xi in enumerate(x): assert list(xi.shape[:bdims])==batch_shape, "x input %d has shape = %s, but expected first dims to match batch_shape = %s"%(i,list(xi.shape),list(batch_shape))
    for i,pi in enumerate(p): assert list(pi.shape[:bdims])==batch_shape, "p input %d has shape = %s, but expected first dims to match batch_shape = %s"%(i,list(xi.shape),list(batch_shape))
    for k,v in bkwargs.items(): assert list(v.shape[:bdims])==batch_shape, "bkwargs['%s'].shape = %s, but expected first dims to match batch_shape = %s"%(k,list(v.shape),list(batch_shape))
    len_bkwargs = len(bkwargs) 
    bkwargs_keys = list(bkwargs.keys())
    def jvpfwrap(*inputs):
        x,p,bkwargs_vals = inputs[:lenx],inputs[lenx:(2*lenx)],inputs[(2*lenx):]
        return torch.func.jvp(lambda *x: f(*x,*bkwargs_vals),x,p)
    jvpfwrapvec = torch.vmap(jvpfwrap,in_dims=(0,)*(2*lenx+len_bkwargs),chunk_size=chunk_size)
    x_input = [xi.flatten(end_dim=bdims-1) if bdims>0 else xi[None,...] for xi in x]
    p_input = [pi.flatten(end_dim=bdims-1) if bdims>0 else pi[None,...] for pi in p]
    bkwargs_vals_input = [bkwargs[key].flatten(end_dim=bdims-1) if bdims>0 else bkwargs[key][None,...] for key in bkwargs_keys]
    y,jvpy = jvpfwrapvec(*x_input,*p_input,*bkwargs_vals_input)
    if isinstance(jvpy,torch.Tensor): jvpy = (jvpy,)
    if isinstance(y,torch.Tensor):
        if bdims==0:
            return tuple(jvpyk[0] for jvpyk in jvpy),y[0]
        else:
            return tuple(jvpyk.reshape(batch_shape+list(jvpyk.shape[1:])) for jvpyk in jvpy),y.reshape(batch_shape+list(y.shape[1:]))
    else:
        if bdims==0:
            return tuple(jvpyk[0] for jvpyk in jvpy),tuple(yk[0] for yk in y)
        else:
            return tuple(jvpyk.reshape(batch_shape+list(jvpyk.shape[1:])) for jvpyk in jvpy),tuple(yk.reshape(batch_shape+list(yk.shape[1:])) for yk in y)

vjpb 

vjpb(f, x, p, bkwargs={}, bdims=0, chunk_size=None)

Batched torch.func.vjp with function evaluation (forward-mode auto-diff)

Parameters:

Name Type Description Default
f callable

Function to compute torch.func.vjp of.

 required
x Union[Tensor, Tuple]

(batched) Torch.Tensor primals.

 required
p Union[Tensor, Tuple]

(batched) Torch.Tensor tangents.

 required
bkwargs dict

(batched) Torch.Tensor items whose jacobians will not be computed.

 {}
bdims int

number of batch dimensions

0
chunk_size int

to be passed into torch.func.vmap.

None

Returns:

Name Type Description
vjp tuple

torch.Tensor jacobian vector products, one with respect to each item in x.

y Tensor

Function evaluations.

Examples:

>>> torch.set_default_dtype(torch.float64)
>>> rng = torch.Generator().manual_seed(7)

>>> f = lambda x: (x[...,None]**torch.arange(2,5)).sum(-2)
>>> x = torch.rand(5,generator=rng)
>>> p = torch.rand(3,generator=rng)
>>> (vjpx,),y = vjpb(f,x,p)
>>> vjpx.shape 
torch.Size([5])
>>> y.shape 
torch.Size([3])
>>> torch.allclose(y,f(x))
True
>>> (jac_x,),_ = jacfwdb(f,x)
>>> torch.allclose(vjpx,p@jac_x)
True

>>> def f(x,z):
...     y = (x[...,:,None,None]**torch.arange(2,5)*z[...,None,:,None]**torch.arange(1,4)).sum((-3,-2))
...     u = (x[...,:,None,None]**torch.arange(3,6)*z[...,None,:,None]**torch.arange(2,5)).sum((-3,-2))
...     v = (x[...,:,None,None]**torch.arange(4,7)*z[...,None,:,None]**torch.arange(3,6)).sum((-3,-2))
...     return y,u,v
>>> x = torch.rand(5,generator=rng)
>>> z = torch.rand(6,generator=rng)
>>> p = torch.rand(3,generator=rng)
>>> q = torch.rand(3,generator=rng)
>>> r = torch.rand(3,generator=rng)
>>> (vjp_x,vjp_z),(y,u,v) = vjpb(f,(x,z),(p,q,r))
>>> vjp_x.shape
torch.Size([5])
>>> vjp_z.shape
torch.Size([6])
>>> y.shape
torch.Size([3])
>>> v.shape
torch.Size([3])
>>> torch.allclose(y,f(x,z)[0])
True
>>> torch.allclose(u,f(x,z)[1])
True
>>> torch.allclose(v,f(x,z)[1])
False
>>> ((jacy_x,jacy_z),(jacu_x,jacu_z),(jacv_x,jacv_z)),_ = jacfwdb(f,(x,z))
>>> torch.allclose(vjp_x,p@jacy_x+q@jacu_x+r@jacv_x)
True
>>> torch.allclose(vjp_z,p@jacy_z+q@jacu_z+r@jacv_z)
True

>>> f = lambda x: (x[...,None]**torch.arange(2,5)).sum(-2)
>>> x = torch.rand(7,5,generator=rng)
>>> p = torch.rand(7,3,generator=rng)
>>> (vjpx,),y = vjpb(f,x,p,bdims=1)
>>> vjpx.shape 
torch.Size([7, 5])
>>> y.shape 
torch.Size([7, 3])
>>> torch.allclose(y,f(x))
True
>>> (jac_x,),_ = jacfwdb(f,x,bdims=1)
>>> torch.allclose(vjpx,(p[...,None]*jac_x).sum(-2))
True

>>> def f(x,z):
...     y = (x[...,:,None,None]**torch.arange(2,5)*z[...,None,:,None]**torch.arange(1,4)).sum((-3,-2))
...     u = (x[...,:,None,None]**torch.arange(3,6)*z[...,None,:,None]**torch.arange(2,5)).sum((-3,-2))
...     v = (x[...,:,None,None]**torch.arange(4,7)*z[...,None,:,None]**torch.arange(3,6)).sum((-3,-2))
...     return y,u,v
>>> x = torch.rand(7,5,generator=rng)
>>> z = torch.rand(7,6,generator=rng)
>>> p = torch.rand(7,3,generator=rng)
>>> q = torch.rand(7,3,generator=rng)
>>> r = torch.rand(7,3,generator=rng)
>>> (vjp_x,vjp_z),(y,u,v) = vjpb(f,(x,z),(p,q,r),bdims=1)
>>> vjp_x.shape
torch.Size([7, 5])
>>> vjp_z.shape
torch.Size([7, 6])
>>> y.shape
torch.Size([7, 3])
>>> v.shape
torch.Size([7, 3])
>>> torch.allclose(y,f(x,z)[0])
True
>>> torch.allclose(u,f(x,z)[1])
True
>>> torch.allclose(v,f(x,z)[1])
False
>>> ((jacy_x,jacy_z),(jacu_x,jacu_z),(jacv_x,jacv_z)),_ = jacfwdb(f,(x,z),bdims=1)
>>> torch.allclose(vjp_x,(p[...,None]*jacy_x).sum(-2)+(q[...,None]*jacu_x).sum(-2)+(r[...,None]*jacv_x).sum(-2))
True
>>> torch.allclose(vjp_z,(p[...,None]*jacy_z).sum(-2)+(q[...,None]*jacu_z).sum(-2)+(r[...,None]*jacv_z).sum(-2))
True

>>> def f(x,z):
...     y = (x[...,:,None,None]**torch.arange(2,5)*z[...,None,:,None]**torch.arange(1,4)).sum((-3,-2))
...     u = (x[...,:,None,None]**torch.arange(3,6)*z[...,None,:,None]**torch.arange(2,5)).sum((-3,-2))
...     v = (x[...,:,None,None]**torch.arange(4,7)*z[...,None,:,None]**torch.arange(3,6)).sum((-3,-2))
...     return y,u,v
>>> x = torch.rand(7,5,generator=rng)
>>> z = torch.rand(7,6,generator=rng)
>>> p = torch.rand(7,3,generator=rng)
>>> q = torch.rand(7,3,generator=rng)
>>> r = torch.rand(7,3,generator=rng)
>>> (vjp_x,),(y,u,v) = vjpb(f,x,(p,q,r),bkwargs={"z":z},bdims=1)
>>> vjp_x.shape
torch.Size([7, 5])
>>> y.shape
torch.Size([7, 3])
>>> v.shape
torch.Size([7, 3])
>>> torch.allclose(y,f(x,z)[0])
True
>>> torch.allclose(u,f(x,z)[1])
True
>>> torch.allclose(v,f(x,z)[1])
False
>>> ((jacy_x,jacy_z),(jacu_x,jacu_z),(jacv_x,jacv_z)),_ = jacfwdb(f,(x,z),bdims=1)
>>> torch.allclose(vjp_x,(p[...,None]*jacy_x).sum(-2)+(q[...,None]*jacu_x).sum(-2)+(r[...,None]*jacv_x).sum(-2))
True
Source code in agsutil/autograd.py 
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
def vjpb(f, x, p, bkwargs={}, bdims=0, chunk_size=None):
    r"""
    Batched `torch.func.vjp` with function evaluation (forward-mode auto-diff)

    Args:
        f (callable): Function to compute `torch.func.vjp` of.
        x (Union[torch.Tensor,Tuple]): (batched) `Torch.Tensor` primals.
        p (Union[torch.Tensor,Tuple]): (batched) `Torch.Tensor` tangents.
        bkwargs (dict): (batched) `Torch.Tensor` items whose jacobians will not be computed.
        bdims (int): number of batch dimensions 
        chunk_size (int): to be passed into `torch.func.vmap`. 

    Returns: 
        vjp (tuple): `torch.Tensor` jacobian vector products, one with respect to each item in `x`. 
        y (torch.Tensor): Function evaluations. 

    Examples: 
        >>> torch.set_default_dtype(torch.float64)
        >>> rng = torch.Generator().manual_seed(7)

        >>> f = lambda x: (x[...,None]**torch.arange(2,5)).sum(-2)
        >>> x = torch.rand(5,generator=rng)
        >>> p = torch.rand(3,generator=rng)
        >>> (vjpx,),y = vjpb(f,x,p)
        >>> vjpx.shape 
        torch.Size([5])
        >>> y.shape 
        torch.Size([3])
        >>> torch.allclose(y,f(x))
        True
        >>> (jac_x,),_ = jacfwdb(f,x)
        >>> torch.allclose(vjpx,p@jac_x)
        True

        >>> def f(x,z):
        ...     y = (x[...,:,None,None]**torch.arange(2,5)*z[...,None,:,None]**torch.arange(1,4)).sum((-3,-2))
        ...     u = (x[...,:,None,None]**torch.arange(3,6)*z[...,None,:,None]**torch.arange(2,5)).sum((-3,-2))
        ...     v = (x[...,:,None,None]**torch.arange(4,7)*z[...,None,:,None]**torch.arange(3,6)).sum((-3,-2))
        ...     return y,u,v
        >>> x = torch.rand(5,generator=rng)
        >>> z = torch.rand(6,generator=rng)
        >>> p = torch.rand(3,generator=rng)
        >>> q = torch.rand(3,generator=rng)
        >>> r = torch.rand(3,generator=rng)
        >>> (vjp_x,vjp_z),(y,u,v) = vjpb(f,(x,z),(p,q,r))
        >>> vjp_x.shape
        torch.Size([5])
        >>> vjp_z.shape
        torch.Size([6])
        >>> y.shape
        torch.Size([3])
        >>> v.shape
        torch.Size([3])
        >>> torch.allclose(y,f(x,z)[0])
        True
        >>> torch.allclose(u,f(x,z)[1])
        True
        >>> torch.allclose(v,f(x,z)[1])
        False
        >>> ((jacy_x,jacy_z),(jacu_x,jacu_z),(jacv_x,jacv_z)),_ = jacfwdb(f,(x,z))
        >>> torch.allclose(vjp_x,p@jacy_x+q@jacu_x+r@jacv_x)
        True
        >>> torch.allclose(vjp_z,p@jacy_z+q@jacu_z+r@jacv_z)
        True

        >>> f = lambda x: (x[...,None]**torch.arange(2,5)).sum(-2)
        >>> x = torch.rand(7,5,generator=rng)
        >>> p = torch.rand(7,3,generator=rng)
        >>> (vjpx,),y = vjpb(f,x,p,bdims=1)
        >>> vjpx.shape 
        torch.Size([7, 5])
        >>> y.shape 
        torch.Size([7, 3])
        >>> torch.allclose(y,f(x))
        True
        >>> (jac_x,),_ = jacfwdb(f,x,bdims=1)
        >>> torch.allclose(vjpx,(p[...,None]*jac_x).sum(-2))
        True

        >>> def f(x,z):
        ...     y = (x[...,:,None,None]**torch.arange(2,5)*z[...,None,:,None]**torch.arange(1,4)).sum((-3,-2))
        ...     u = (x[...,:,None,None]**torch.arange(3,6)*z[...,None,:,None]**torch.arange(2,5)).sum((-3,-2))
        ...     v = (x[...,:,None,None]**torch.arange(4,7)*z[...,None,:,None]**torch.arange(3,6)).sum((-3,-2))
        ...     return y,u,v
        >>> x = torch.rand(7,5,generator=rng)
        >>> z = torch.rand(7,6,generator=rng)
        >>> p = torch.rand(7,3,generator=rng)
        >>> q = torch.rand(7,3,generator=rng)
        >>> r = torch.rand(7,3,generator=rng)
        >>> (vjp_x,vjp_z),(y,u,v) = vjpb(f,(x,z),(p,q,r),bdims=1)
        >>> vjp_x.shape
        torch.Size([7, 5])
        >>> vjp_z.shape
        torch.Size([7, 6])
        >>> y.shape
        torch.Size([7, 3])
        >>> v.shape
        torch.Size([7, 3])
        >>> torch.allclose(y,f(x,z)[0])
        True
        >>> torch.allclose(u,f(x,z)[1])
        True
        >>> torch.allclose(v,f(x,z)[1])
        False
        >>> ((jacy_x,jacy_z),(jacu_x,jacu_z),(jacv_x,jacv_z)),_ = jacfwdb(f,(x,z),bdims=1)
        >>> torch.allclose(vjp_x,(p[...,None]*jacy_x).sum(-2)+(q[...,None]*jacu_x).sum(-2)+(r[...,None]*jacv_x).sum(-2))
        True
        >>> torch.allclose(vjp_z,(p[...,None]*jacy_z).sum(-2)+(q[...,None]*jacu_z).sum(-2)+(r[...,None]*jacv_z).sum(-2))
        True

        >>> def f(x,z):
        ...     y = (x[...,:,None,None]**torch.arange(2,5)*z[...,None,:,None]**torch.arange(1,4)).sum((-3,-2))
        ...     u = (x[...,:,None,None]**torch.arange(3,6)*z[...,None,:,None]**torch.arange(2,5)).sum((-3,-2))
        ...     v = (x[...,:,None,None]**torch.arange(4,7)*z[...,None,:,None]**torch.arange(3,6)).sum((-3,-2))
        ...     return y,u,v
        >>> x = torch.rand(7,5,generator=rng)
        >>> z = torch.rand(7,6,generator=rng)
        >>> p = torch.rand(7,3,generator=rng)
        >>> q = torch.rand(7,3,generator=rng)
        >>> r = torch.rand(7,3,generator=rng)
        >>> (vjp_x,),(y,u,v) = vjpb(f,x,(p,q,r),bkwargs={"z":z},bdims=1)
        >>> vjp_x.shape
        torch.Size([7, 5])
        >>> y.shape
        torch.Size([7, 3])
        >>> v.shape
        torch.Size([7, 3])
        >>> torch.allclose(y,f(x,z)[0])
        True
        >>> torch.allclose(u,f(x,z)[1])
        True
        >>> torch.allclose(v,f(x,z)[1])
        False
        >>> ((jacy_x,jacy_z),(jacu_x,jacu_z),(jacv_x,jacv_z)),_ = jacfwdb(f,(x,z),bdims=1)
        >>> torch.allclose(vjp_x,(p[...,None]*jacy_x).sum(-2)+(q[...,None]*jacu_x).sum(-2)+(r[...,None]*jacv_x).sum(-2))
        True
    """
    if isinstance(x,torch.Tensor): x = (x,)
    if isinstance(p,torch.Tensor): p = (p,)
    lenx = len(x) 
    assert len(x)>=1
    lenp = len(p)
    assert len(p)>=1
    batch_shape = list(x[0].shape[:bdims])
    for i,xi in enumerate(x): assert list(xi.shape[:bdims])==batch_shape, "x input %d has shape = %s, but expected first dims to match batch_shape = %s"%(i,list(xi.shape),list(batch_shape))
    for i,pi in enumerate(p): assert list(pi.shape[:bdims])==batch_shape, "p input %d has shape = %s, but expected first dims to match batch_shape = %s"%(i,list(xi.shape),list(batch_shape))
    for k,v in bkwargs.items(): assert list(v.shape[:bdims])==batch_shape, "bkwargs['%s'].shape = %s, but expected first dims to match batch_shape = %s"%(k,list(v.shape),list(batch_shape))
    len_bkwargs = len(bkwargs) 
    bkwargs_keys = list(bkwargs.keys())
    def vjpfwrap(*inputs):
        x,p,bkwargs_vals = inputs[:lenx],inputs[lenx:(lenx+lenp)],inputs[(lenx+lenp):]
        y,vjpfwrap_inner = torch.func.vjp(lambda *x: f(*x,*bkwargs_vals),*x)
        vjpywrap = vjpfwrap_inner(p[0] if isinstance(y,torch.Tensor) else p)
        return y,vjpywrap
    vjpfwrapvec = torch.vmap(vjpfwrap,in_dims=(0,)*(lenx+lenp+len_bkwargs),chunk_size=chunk_size)
    x_input = [xi.flatten(end_dim=bdims-1) if bdims>0 else xi[None,...] for xi in x]
    p_input = [pi.flatten(end_dim=bdims-1) if bdims>0 else pi[None,...] for pi in p]
    bkwargs_vals_input = [bkwargs[key].flatten(end_dim=bdims-1) if bdims>0 else bkwargs[key][None,...] for key in bkwargs_keys]
    y,vjpys = vjpfwrapvec(*x_input,*p_input,*bkwargs_vals_input)
    if isinstance(y,torch.Tensor):
        if bdims==0:
            return (vjpy[0] for vjpy in vjpys),y[0]
        else:
            return (vjpy.reshape(batch_shape+list(vjpy.shape[1:])) for vjpy in vjpys),y.reshape(batch_shape+list(y.shape[1:]))
    else:
        if bdims==0:
            return (vjpy[0] for vjpy in vjpys),(yk[0] for yk in y)
        else:
            return (vjpy.reshape(batch_shape+list(vjpy.shape[1:])) for vjpy in vjpys),(yk.reshape(batch_shape+list(yk.shape[1:])) for yk in y)

Misc utils

Timer 

Timer(device)

Timer compatible with CPU and GPU operations.

Parameters:

Name Type Description Default
device Union[device, str]

device to perform timing for.

  • For CPU devices, time.perf_counter() is used.
  • For CUDA and MPS GPU devices, torch.{cuda,mps}.Event(enable_timing=True) is used.
 required
Source code in agsutil/utils.py 
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
def __init__(self, device):
    r"""
    Args:
        device (Union[torch.device,str]): device to perform timing for. 

            - For CPU devices, `time.perf_counter()` is used. 
            - For CUDA and MPS GPU devices, `torch.{cuda,mps}.Event(enable_timing=True)` is used. 
    """

    device_str = str(device) 
    if "cpu" in device_str:
        self.torch_backend = torch.cpu 
    elif "cuda" in device_str:
        self.torch_backend = torch.cuda 
    elif "mps" in device_str:
        self.torch_backend = torch.mps 
    else:
        raise Exception("undetected device = %s, should have 'cpu', 'cuda', or 'mps' in it."%device_str)

tic 

tic()

Start the stopwatch.

Source code in agsutil/utils.py 
86
87
88
89
90
91
92
93
94
95
96
def tic(self):
    r"""
    Start the stopwatch.
    """
    if self.torch_backend==torch.cpu:
        self.t0 = time.perf_counter()
    else:
        self.torch_backend.empty_cache()
        self.t0 = self.torch_backend.Event(enable_timing=True)
        self.tend = self.torch_backend.Event(enable_timing=True)
        self.t0.record()

toc 

toc()

Lap the stopwatch.

Returns:

Name Type Description
tdelta float

time elapsed between the start of the stopwatch and the current lap.
Source code in agsutil/utils.py 
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
def toc(self):
    r"""
    Lap the stopwatch. 

    Returns: 
        tdelta (float): time elapsed between the start of the stopwatch and the current lap.
    """
    if self.torch_backend==torch.cpu:
        tdelta = time.perf_counter()-self.t0
    else:
        self.tend.record()
        self.torch_backend.synchronize()
        tdelta = self.t0.elapsed_time(self.tend)/1000
    return float(tdelta)

print_data_signatures 

print_data_signatures(data, name='data', print_devices=False, print_dtypes=False, verbose_indent=4)

Print data shapes and (optionally) devices.

Parameters:

Name Type Description Default
data dict

Dictiony with items that are tensors or dictionaries of tensors.

required
print_devices bool

If True, also print the device.

False
print_dtypes bool

If True, also print the dtypes.

False
verbose_indent int

Non-negative number of indentation spaces for logging.

 4

Examples:

>>> data = {
...     "a": torch.rand(2,3,4),
...     "b": torch.rand(3,4,5),
...     "subdata": {
...         "aa": torch.rand(2,3),
...         "bb": torch.rand(2,3),
...         "subnontensor": ["ags",7,7,7],
...         },
...     "nontensor": [7,7,7,"ags"]
...     }
>>> print_data_signatures(data,print_devices=True,print_dtypes=True,verbose_indent=0)
data['a'].shape = (2, 3, 4) on device = cpu with dtype = torch.float64
data['b'].shape = (3, 4, 5) on device = cpu with dtype = torch.float64
data['subdata']
    data['subdata']['aa'].shape = (2, 3) on device = cpu with dtype = torch.float64
    data['subdata']['bb'].shape = (2, 3) on device = cpu with dtype = torch.float64
    data['subdata']['subnontensor'] a list of length 4
data['nontensor'] a list of length 4
Source code in agsutil/utils.py 
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
def print_data_signatures(data, name="data", print_devices=False, print_dtypes=False, verbose_indent=4):
    r""" 
    Print data shapes and (optionally) devices. 

    Args: 
        data (dict): Dictiony with items that are tensors or dictionaries of tensors. 
        print_devices (bool): If `True`, also print the device. 
        print_dtypes (bool): If `True`, also print the dtypes. 
        verbose_indent (int): Non-negative number of indentation spaces for logging.

    Examples:
        >>> data = {
        ...     "a": torch.rand(2,3,4),
        ...     "b": torch.rand(3,4,5),
        ...     "subdata": {
        ...         "aa": torch.rand(2,3),
        ...         "bb": torch.rand(2,3),
        ...         "subnontensor": ["ags",7,7,7],
        ...         },
        ...     "nontensor": [7,7,7,"ags"]
        ...     }
        >>> print_data_signatures(data,print_devices=True,print_dtypes=True,verbose_indent=0)
        data['a'].shape = (2, 3, 4) on device = cpu with dtype = torch.float64
        data['b'].shape = (3, 4, 5) on device = cpu with dtype = torch.float64
        data['subdata']
            data['subdata']['aa'].shape = (2, 3) on device = cpu with dtype = torch.float64
            data['subdata']['bb'].shape = (2, 3) on device = cpu with dtype = torch.float64
            data['subdata']['subnontensor'] a list of length 4
        data['nontensor'] a list of length 4
    """ 
    for key,val in data.items():
        if isinstance(val,torch.Tensor):
            _s = "%s['%s'].shape = %s"%(name,key,str(tuple(data[key].shape)))
            if print_devices:
                _s += " on device = %s"%str(data[key].device)
            if print_dtypes:
                _s += " with dtype = %s"%str(data[key].dtype)
            print(" "*verbose_indent+_s)
        elif isinstance(val,dict):
            print(" "*verbose_indent+"data['%s']"%key)
            for kkey,vval in val.items():
                if isinstance(vval,torch.Tensor):
                    _s = "%s['%s']['%s'].shape = %s"%(name,key,kkey,str(tuple(data[key][kkey].shape)))
                    if print_devices:
                        _s += " on device = %s"%str(data[key][kkey].device)
                    if print_dtypes:
                        _s += " with dtype = %s"%str(data[key][kkey].dtype)
                    print(" "*(verbose_indent+4)+_s)
                elif isinstance(vval,list):
                    print(" "*(verbose_indent+4)+"%s['%s']['%s'] a list of length %d"%(name,key,kkey,len(data[key][kkey])))
                else:
                    print(" "*(verbose_indent+4)+"%s['%s']['%s'] = %s"%(name,key,kkey,str(data[key][kkey])))
        elif isinstance(val,list):
            print(" "*verbose_indent+"%s['%s'] a list of length %d"%(name,key,len(data[key])))
        else:
            print(" "*verbose_indent+"%s['%s'] = %s"%(name,key,str(data[key])))

to_unitary_expskewh 

to_unitary_expskewh(theta, n, complex_case=False)

Transform to a unitary matrix using the exponential of a skew Hermitian matrix.

Parameters:

Name Type Description Default
theta Tensor

With theta.size(-1) == n*(n-1)//2 in the real case and theta.size(-1) == n**2 in the complex case.

 required
complex_case bool

If True, parameterize a complex matrix, otherwise a real one.

False

Returns:

Name Type Description
Q Tensor

With shape (*theta.shape[:-1],n,n) unitary matrices

Examples:

>>> torch.set_default_dtype(torch.float64)
>>> rng = torch.Generator().manual_seed(7)

Single matrix

>>> n = 5
>>> theta = torch.rand(n*(n-1)//2,generator=rng)
>>> Q = to_unitary_expskewh(theta,n)
>>> Q 
tensor([[ 0.5027, -0.1038, -0.2508,  0.3769,  0.7291],
        [-0.6870,  0.3679,  0.2777,  0.3454,  0.4430],
        [-0.2330, -0.7975,  0.3425, -0.2999,  0.3201],
        [-0.4638, -0.3141, -0.8037,  0.1779, -0.0934],
        [-0.0769,  0.3453, -0.3111, -0.7855,  0.4013]])
>>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q),torch.eye(n))
True

Single complex matrix

>>> n = 4
>>> theta = torch.rand(n**2,generator=rng)
>>> Q = to_unitary_expskewh(theta,n,complex_case=True)
>>> Q 
tensor([[ 0.1358-0.3038j, -0.4798+0.2431j,  0.1105+0.2308j, -0.2229+0.6963j],
        [-0.4395+0.6202j,  0.1556-0.1974j, -0.1838-0.0838j, -0.1181+0.5516j],
        [-0.4750+0.0762j, -0.7622+0.0020j, -0.2319-0.0263j,  0.2346-0.2795j],
        [-0.0745-0.2729j,  0.2531+0.0496j, -0.7624+0.5079j,  0.1119+0.0410j]])
>>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q.conj()),torch.eye(n,dtype=torch.complex128))
True

Two matrices

>>> n = 3
>>> theta = torch.rand(2,n*(n-1)//2,generator=rng)
>>> Q = to_unitary_expskewh(theta,n)
>>> Q 
tensor([[[ 0.3004,  0.6612,  0.6874],
         [-0.7630,  0.5990, -0.2428],
         [-0.5723, -0.4516,  0.6845]],
<BLANKLINE>
        [[ 0.4583,  0.7552,  0.4687],
         [-0.7977,  0.5820, -0.1579],
         [-0.3921, -0.3015,  0.8691]]])
>>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q),torch.eye(n))
True

Two complex matrices

>>> n = 3
>>> theta = torch.rand(2,n**2,generator=rng)
>>> Q = to_unitary_expskewh(theta,n,complex_case=True)
>>> Q 
tensor([[[ 0.5591+0.3745j, -0.4342+0.5076j, -0.1283+0.2907j],
         [-0.4802+0.4013j,  0.5099+0.4179j, -0.2657+0.3211j],
         [-0.3843+0.0887j, -0.2972+0.1756j,  0.8185+0.2353j]],
<BLANKLINE>
        [[ 0.4860+0.5953j, -0.1642+0.0458j, -0.3162+0.5295j],
         [-0.5638+0.0805j,  0.3753+0.0607j,  0.2470+0.6856j],
         [-0.2898+0.0325j, -0.8621+0.2884j,  0.2790+0.1037j]]])
>>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q.conj()),torch.eye(n,dtype=torch.complex128))
True

Batch support

>>> n = 10
>>> theta = torch.rand(2,3,4,n*(n-1)//2,generator=rng)
>>> Q = to_unitary_expskewh(theta,n)
>>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q),torch.eye(n))
True
>>> theta = torch.rand(2,3,4,n**2,generator=rng)
>>> Q = to_unitary_expskewh(theta,n,complex_case=True)
>>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q.conj()),torch.eye(n,dtype=torch.complex128))
True
Source code in agsutil/utils.py 
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
def to_unitary_expskewh(theta, n, complex_case=False):
    r"""
    Transform to a unitary matrix using the exponential of a skew Hermitian matrix.

    Args:
        theta (torch.Tensor): With `theta.size(-1) == n*(n-1)//2` in the real case and `theta.size(-1) == n**2` in the complex case.
        complex_case (bool): If `True`, parameterize a complex matrix, otherwise a real one. 

    Returns:
        Q (torch.Tensor): With shape `(*theta.shape[:-1],n,n)` unitary matrices

    Examples:
        >>> torch.set_default_dtype(torch.float64)
        >>> rng = torch.Generator().manual_seed(7)

    Single matrix

        >>> n = 5
        >>> theta = torch.rand(n*(n-1)//2,generator=rng)
        >>> Q = to_unitary_expskewh(theta,n)
        >>> Q 
        tensor([[ 0.5027, -0.1038, -0.2508,  0.3769,  0.7291],
                [-0.6870,  0.3679,  0.2777,  0.3454,  0.4430],
                [-0.2330, -0.7975,  0.3425, -0.2999,  0.3201],
                [-0.4638, -0.3141, -0.8037,  0.1779, -0.0934],
                [-0.0769,  0.3453, -0.3111, -0.7855,  0.4013]])
        >>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q),torch.eye(n))
        True

    Single complex matrix

        >>> n = 4
        >>> theta = torch.rand(n**2,generator=rng)
        >>> Q = to_unitary_expskewh(theta,n,complex_case=True)
        >>> Q 
        tensor([[ 0.1358-0.3038j, -0.4798+0.2431j,  0.1105+0.2308j, -0.2229+0.6963j],
                [-0.4395+0.6202j,  0.1556-0.1974j, -0.1838-0.0838j, -0.1181+0.5516j],
                [-0.4750+0.0762j, -0.7622+0.0020j, -0.2319-0.0263j,  0.2346-0.2795j],
                [-0.0745-0.2729j,  0.2531+0.0496j, -0.7624+0.5079j,  0.1119+0.0410j]])
        >>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q.conj()),torch.eye(n,dtype=torch.complex128))
        True

    Two matrices

        >>> n = 3
        >>> theta = torch.rand(2,n*(n-1)//2,generator=rng)
        >>> Q = to_unitary_expskewh(theta,n)
        >>> Q 
        tensor([[[ 0.3004,  0.6612,  0.6874],
                 [-0.7630,  0.5990, -0.2428],
                 [-0.5723, -0.4516,  0.6845]],
        <BLANKLINE>
                [[ 0.4583,  0.7552,  0.4687],
                 [-0.7977,  0.5820, -0.1579],
                 [-0.3921, -0.3015,  0.8691]]])
        >>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q),torch.eye(n))
        True

    Two complex matrices

        >>> n = 3
        >>> theta = torch.rand(2,n**2,generator=rng)
        >>> Q = to_unitary_expskewh(theta,n,complex_case=True)
        >>> Q 
        tensor([[[ 0.5591+0.3745j, -0.4342+0.5076j, -0.1283+0.2907j],
                 [-0.4802+0.4013j,  0.5099+0.4179j, -0.2657+0.3211j],
                 [-0.3843+0.0887j, -0.2972+0.1756j,  0.8185+0.2353j]],
        <BLANKLINE>
                [[ 0.4860+0.5953j, -0.1642+0.0458j, -0.3162+0.5295j],
                 [-0.5638+0.0805j,  0.3753+0.0607j,  0.2470+0.6856j],
                 [-0.2898+0.0325j, -0.8621+0.2884j,  0.2790+0.1037j]]])
        >>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q.conj()),torch.eye(n,dtype=torch.complex128))
        True

    Batch support

        >>> n = 10
        >>> theta = torch.rand(2,3,4,n*(n-1)//2,generator=rng)
        >>> Q = to_unitary_expskewh(theta,n)
        >>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q),torch.eye(n))
        True
        >>> theta = torch.rand(2,3,4,n**2,generator=rng)
        >>> Q = to_unitary_expskewh(theta,n,complex_case=True)
        >>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q.conj()),torch.eye(n,dtype=torch.complex128))
        True
    """
    batch_shape = tuple(theta.shape[:-1])
    batch_ones = torch.ones(batch_shape,dtype=int,device=theta.device)
    iut = torch.triu_indices(n,n,offset=1,device=theta.device)
    iutf = torch.einsum("...,i->...i",batch_ones,iut[0]*n+iut[1])
    iltf = torch.einsum("...,i->...i",batch_ones,iut[1]*n+iut[0])
    if complex_case:
        assert theta.size(-1)==(n**2)
        complex_dtype = (theta+0j).dtype
        H = torch.zeros((*batch_shape,n*n),dtype=complex_dtype,device=theta.device)
        idiag = torch.arange(n,device=theta.device)
        idiagf = torch.einsum("...,i->...i",batch_ones,idiag*n+idiag)
        diag = 1j*theta[...,:n].to(complex_dtype)
        off_real,off_imag = theta[...,n:n+(n**2-n)//2],theta[...,n+(n**2-n)//2:]
        off_vals = off_real.to(complex_dtype)+1j*off_imag.to(complex_dtype)
        H = H.scatter_add(-1,idiagf,diag)
        H = H.scatter_add(-1,iutf,off_vals)
        H = H.scatter_add(-1,iltf,-off_vals.conj())
    else:
        assert theta.size(-1)==(n*(n-1)//2)
        H = torch.zeros((*batch_shape,n*n),dtype=theta.dtype,device=theta.device)
        H = H.scatter_add(-1,iutf,theta)
        H = H.scatter_add(-1,iltf,-theta)
    H = H.reshape((*batch_shape,n,n))
    return torch.matrix_exp(H)

from_unitary_expskewh 

from_unitary_expskewh(Q, complex_case=False)

Transform from a unitary matrix using the exponential of a skew Hermitian matrix.

Parameters:

Name Type Description Default
Q Tensor

With shape (*theta.shape[:-1],n,n) unitary matrices

 required
complex_case bool

If True, parameterize a complex matrix, otherwise a real one.

False

Returns:

Name Type Description
theta Tensor

With theta.size(-1) == n*(n-1)//2 in the real case and theta.size(-1) == n**2 in the complex case.
Note that this theta is not unique, as shown in the following doctests.

Examples:

>>> torch.set_default_dtype(torch.float64)
>>> rng = torch.Generator().manual_seed(7)

Single matrix

>>> n = 10
>>> theta = torch.rand(n*(n-1)//2,generator=rng)
>>> Q = to_unitary_expskewh(theta,n)
>>> theta2 = from_unitary_expskewh(Q)
>>> torch.allclose(theta,theta2)
False
>>> Q2 = to_unitary_expskewh(theta2,n)
>>> torch.allclose(Q,Q2)
True

Single complex matrix

>>> n = 10
>>> theta = torch.rand(n**2,generator=rng)
>>> Q = to_unitary_expskewh(theta,n,complex_case=True)
>>> theta2 = from_unitary_expskewh(Q,complex_case=True)
>>> torch.allclose(theta,theta2)
False
>>> Q2 = to_unitary_expskewh(theta2,n,complex_case=True)
>>> torch.allclose(Q,Q2)
True

Two matrices

>>> n = 10
>>> theta = torch.rand(2,n*(n-1)//2,generator=rng)
>>> Q = to_unitary_expskewh(theta,n)
>>> theta2 = from_unitary_expskewh(Q)
>>> torch.allclose(theta,theta2)
False
>>> Q2 = to_unitary_expskewh(theta2,n)
>>> torch.allclose(Q,Q2)
True

Two complex matrices

>>> n = 3
>>> theta = torch.rand(2,n**2,generator=rng)
>>> Q = to_unitary_expskewh(theta,n,complex_case=True)
>>> theta2 = from_unitary_expskewh(Q,complex_case=True)
>>> Q2 = to_unitary_expskewh(theta2,n,complex_case=True)
>>> torch.allclose(Q,Q2)
True

Batch support

>>> n = 10
>>> theta = torch.rand(2,3,4,n*(n-1)//2,generator=rng)
>>> Q = to_unitary_expskewh(theta,n)
>>> theta2 = from_unitary_expskewh(Q)
>>> Q2 = to_unitary_expskewh(theta2,n)
>>> torch.allclose(Q,Q2)
True
>>> theta = torch.rand(2,3,4,n**2,generator=rng)
>>> Q = to_unitary_expskewh(theta,n,complex_case=True)
>>> theta2 = from_unitary_expskewh(Q,complex_case=True)
>>> Q2 = to_unitary_expskewh(theta2,n,complex_case=True)
>>> torch.allclose(Q,Q2)
True
Source code in agsutil/utils.py 
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
def from_unitary_expskewh(Q, complex_case=False):
    r"""
    Transform from a unitary matrix using the exponential of a skew Hermitian matrix.

    Args:
        Q (torch.Tensor): With shape `(*theta.shape[:-1],n,n)` unitary matrices
        complex_case (bool): If `True`, parameterize a complex matrix, otherwise a real one. 

    Returns:
        theta (torch.Tensor): With `theta.size(-1) == n*(n-1)//2` in the real case and `theta.size(-1) == n**2` in the complex case.  
            Note that this `theta` is not unique, as shown in the following doctests.

    Examples:
        >>> torch.set_default_dtype(torch.float64)
        >>> rng = torch.Generator().manual_seed(7)

    Single matrix

        >>> n = 10
        >>> theta = torch.rand(n*(n-1)//2,generator=rng)
        >>> Q = to_unitary_expskewh(theta,n)
        >>> theta2 = from_unitary_expskewh(Q)
        >>> torch.allclose(theta,theta2)
        False
        >>> Q2 = to_unitary_expskewh(theta2,n)
        >>> torch.allclose(Q,Q2)
        True

    Single complex matrix

        >>> n = 10
        >>> theta = torch.rand(n**2,generator=rng)
        >>> Q = to_unitary_expskewh(theta,n,complex_case=True)
        >>> theta2 = from_unitary_expskewh(Q,complex_case=True)
        >>> torch.allclose(theta,theta2)
        False
        >>> Q2 = to_unitary_expskewh(theta2,n,complex_case=True)
        >>> torch.allclose(Q,Q2)
        True

    Two matrices

        >>> n = 10
        >>> theta = torch.rand(2,n*(n-1)//2,generator=rng)
        >>> Q = to_unitary_expskewh(theta,n)
        >>> theta2 = from_unitary_expskewh(Q)
        >>> torch.allclose(theta,theta2)
        False
        >>> Q2 = to_unitary_expskewh(theta2,n)
        >>> torch.allclose(Q,Q2)
        True

    Two complex matrices

        >>> n = 3
        >>> theta = torch.rand(2,n**2,generator=rng)
        >>> Q = to_unitary_expskewh(theta,n,complex_case=True)
        >>> theta2 = from_unitary_expskewh(Q,complex_case=True)
        >>> Q2 = to_unitary_expskewh(theta2,n,complex_case=True)
        >>> torch.allclose(Q,Q2)
        True

    Batch support

        >>> n = 10
        >>> theta = torch.rand(2,3,4,n*(n-1)//2,generator=rng)
        >>> Q = to_unitary_expskewh(theta,n)
        >>> theta2 = from_unitary_expskewh(Q)
        >>> Q2 = to_unitary_expskewh(theta2,n)
        >>> torch.allclose(Q,Q2)
        True
        >>> theta = torch.rand(2,3,4,n**2,generator=rng)
        >>> Q = to_unitary_expskewh(theta,n,complex_case=True)
        >>> theta2 = from_unitary_expskewh(Q,complex_case=True)
        >>> Q2 = to_unitary_expskewh(theta2,n,complex_case=True)
        >>> torch.allclose(Q,Q2)
        True
    """
    n = Q.size(-1)
    L = logm_unitary(Q)
    iut = torch.triu_indices(n,n,offset=1,device=Q.device)
    if complex_case:
        idiag = torch.arange(n,device=Q.device)
        diag = L[...,idiag,idiag].imag
        off_diag = L[...,iut[0],iut[1]]
        theta = torch.cat([diag,off_diag.real,off_diag.imag],dim=-1)
    else:
        theta = L[...,iut[0],iut[1]].real
    return theta

to_unitary_qr 

to_unitary_qr(A)

Transform to a unitary matrix using the QR decomposition.

Parameters:

Name Type Description Default
A Tensor

With A.shape == (...,n,n).

 required

Returns:

Name Type Description
Q Tensor

With shape (*A.shape[:-1],n,n) unitary matrices

Examples:

>>> torch.set_default_dtype(torch.float64)
>>> rng = torch.Generator().manual_seed(7)

Single matrix

>>> n = 5
>>> A = torch.rand(n,n,generator=rng)
>>> Q = to_unitary_qr(A)
>>> Q 
tensor([[ 0.1819,  0.2300,  0.8956,  0.3042,  0.1391],
        [ 0.5466, -0.1828, -0.3033,  0.7517, -0.1037],
        [ 0.0410,  0.0450,  0.1626, -0.0886, -0.9808],
        [ 0.5518, -0.6553,  0.2046, -0.4684,  0.0693],
        [ 0.6017,  0.6944, -0.1939, -0.3392,  0.0555]])
>>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q),torch.eye(n))
True

Single complex matrix

>>> n = 4
>>> A = torch.rand(n,n,generator=rng,dtype=torch.complex128)
>>> Q = to_unitary_qr(A)
>>> Q 
tensor([[ 0.3820+4.9592e-01j,  0.6227-3.3313e-01j,  0.2301+4.5375e-04j,
         -0.1762-1.5935e-01j],
        [ 0.3750+1.7952e-01j, -0.4719+1.6687e-01j,  0.1957+5.4894e-02j,
          0.2213-6.9735e-01j],
        [ 0.4702+1.8505e-01j,  0.0477+2.3842e-01j, -0.8138+2.9071e-02j,
          0.0823+1.2523e-01j],
        [ 0.4241+1.1515e-02j,  0.0064+4.3764e-01j,  0.4874-7.1170e-02j,
          0.2837+5.5258e-01j]])
>>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q.conj()),torch.eye(n,dtype=torch.complex128))
True

Two matrices

>>> n = 3
>>> A = torch.rand(2,n,n,generator=rng)
>>> Q = to_unitary_qr(A)
>>> Q 
tensor([[[ 0.6587,  0.2967, -0.6914],
         [ 0.6269, -0.7246,  0.2864],
         [ 0.4160,  0.6221,  0.6633]],
<BLANKLINE>
        [[ 0.5190,  0.8009, -0.2985],
         [ 0.6364, -0.1289,  0.7605],
         [ 0.5706, -0.5847, -0.5766]]])
>>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q),torch.eye(n))
True

Two complex matrices

>>> n = 3
>>> A = torch.rand(2,n,n,generator=rng,dtype=torch.complex128)
>>> Q = to_unitary_qr(A)
>>> Q 
tensor([[[ 0.5834+0.1938j, -0.0560+0.3895j, -0.3185+0.6048j],
         [ 0.4323+0.4517j,  0.6040-0.4547j,  0.0490-0.1873j],
         [ 0.4700+0.1009j, -0.3015+0.4275j,  0.4046-0.5759j]],
<BLANKLINE>
        [[ 0.2769+0.2957j, -0.1093+0.8955j, -0.0100+0.1480j],
         [ 0.0858+0.6104j,  0.0647-0.1696j,  0.7218-0.2572j],
         [ 0.2022+0.6443j,  0.2591-0.2933j, -0.5677+0.2621j]]])
>>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q.conj()),torch.eye(n,dtype=torch.complex128))
True

Batch support

>>> n = 10
>>> A = torch.rand(2,3,4,n,n,generator=rng)
>>> Q = to_unitary_qr(A)
>>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q),torch.eye(n))
True
>>> A = torch.rand(2,3,4,n,n,generator=rng,dtype=torch.complex128)
>>> Q = to_unitary_qr(A)
>>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q.conj()),torch.eye(n,dtype=torch.complex128))
True
Source code in agsutil/utils.py 
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
def to_unitary_qr(A):
    r"""
    Transform to a unitary matrix using the QR decomposition.

    Args:
        A (torch.Tensor): With `A.shape == (...,n,n)`.

    Returns:
        Q (torch.Tensor): With shape `(*A.shape[:-1],n,n)` unitary matrices

    Examples:
        >>> torch.set_default_dtype(torch.float64)
        >>> rng = torch.Generator().manual_seed(7)

    Single matrix

        >>> n = 5
        >>> A = torch.rand(n,n,generator=rng)
        >>> Q = to_unitary_qr(A)
        >>> Q 
        tensor([[ 0.1819,  0.2300,  0.8956,  0.3042,  0.1391],
                [ 0.5466, -0.1828, -0.3033,  0.7517, -0.1037],
                [ 0.0410,  0.0450,  0.1626, -0.0886, -0.9808],
                [ 0.5518, -0.6553,  0.2046, -0.4684,  0.0693],
                [ 0.6017,  0.6944, -0.1939, -0.3392,  0.0555]])
        >>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q),torch.eye(n))
        True

    Single complex matrix

        >>> n = 4
        >>> A = torch.rand(n,n,generator=rng,dtype=torch.complex128)
        >>> Q = to_unitary_qr(A)
        >>> Q 
        tensor([[ 0.3820+4.9592e-01j,  0.6227-3.3313e-01j,  0.2301+4.5375e-04j,
                 -0.1762-1.5935e-01j],
                [ 0.3750+1.7952e-01j, -0.4719+1.6687e-01j,  0.1957+5.4894e-02j,
                  0.2213-6.9735e-01j],
                [ 0.4702+1.8505e-01j,  0.0477+2.3842e-01j, -0.8138+2.9071e-02j,
                  0.0823+1.2523e-01j],
                [ 0.4241+1.1515e-02j,  0.0064+4.3764e-01j,  0.4874-7.1170e-02j,
                  0.2837+5.5258e-01j]])
        >>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q.conj()),torch.eye(n,dtype=torch.complex128))
        True

    Two matrices

        >>> n = 3
        >>> A = torch.rand(2,n,n,generator=rng)
        >>> Q = to_unitary_qr(A)
        >>> Q 
        tensor([[[ 0.6587,  0.2967, -0.6914],
                 [ 0.6269, -0.7246,  0.2864],
                 [ 0.4160,  0.6221,  0.6633]],
        <BLANKLINE>
                [[ 0.5190,  0.8009, -0.2985],
                 [ 0.6364, -0.1289,  0.7605],
                 [ 0.5706, -0.5847, -0.5766]]])
        >>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q),torch.eye(n))
        True

    Two complex matrices

        >>> n = 3
        >>> A = torch.rand(2,n,n,generator=rng,dtype=torch.complex128)
        >>> Q = to_unitary_qr(A)
        >>> Q 
        tensor([[[ 0.5834+0.1938j, -0.0560+0.3895j, -0.3185+0.6048j],
                 [ 0.4323+0.4517j,  0.6040-0.4547j,  0.0490-0.1873j],
                 [ 0.4700+0.1009j, -0.3015+0.4275j,  0.4046-0.5759j]],
        <BLANKLINE>
                [[ 0.2769+0.2957j, -0.1093+0.8955j, -0.0100+0.1480j],
                 [ 0.0858+0.6104j,  0.0647-0.1696j,  0.7218-0.2572j],
                 [ 0.2022+0.6443j,  0.2591-0.2933j, -0.5677+0.2621j]]])
        >>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q.conj()),torch.eye(n,dtype=torch.complex128))
        True

    Batch support

        >>> n = 10
        >>> A = torch.rand(2,3,4,n,n,generator=rng)
        >>> Q = to_unitary_qr(A)
        >>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q),torch.eye(n))
        True
        >>> A = torch.rand(2,3,4,n,n,generator=rng,dtype=torch.complex128)
        >>> Q = to_unitary_qr(A)
        >>> torch.allclose(torch.einsum("...ji,...jk->...ik",Q,Q.conj()),torch.eye(n,dtype=torch.complex128))
        True
    """
    n = A.size(-1)
    assert A.shape[-2:]==(n,n)
    Q,R = torch.linalg.qr(A)
    phases = torch.sgn(torch.diagonal(R,dim1=-2,dim2=-1))
    return Q*phases.unsqueeze(-2)

get_torch_rng 

get_torch_rng(seed=None, device=None)

Get a torch.Generator() with a random seed when seed=None or a fixed seed when seed is an int.
This is necessary because torch.Generator() uses a fixed default seed.

Parameters:

Name Type Description Default
seed Union[None, int, Generator()]

Random seed.

None
device device

Device on which to place the generator.

 None

Returns:

Name Type Description
rng Generator

The random number generator
Source code in agsutil/utils.py 
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
def get_torch_rng(seed=None, device=None):
    r"""
    Get a `torch.Generator()` with a random seed when `seed=None` or a fixed seed when `seed` is an int.  
    This is necessary because torch.Generator() uses a fixed default seed.  


    Args:
        seed (Union[None,int,torch.Generator()]): Random seed. 
        device (torch.device): Device on which to place the generator.

    Returns: 
        rng (torch.Generator): The random number generator
    """
    if device is None: 
        device = torch.get_default_device()
    if isinstance(seed,torch.Generator):
        rng = seed 
    elif seed is None: 
        rng = torch.Generator(device=device)
        rng.seed()
    else:
        rng = torch.Generator(device=device).manual_seed(seed)
    return rng

logmultinomialcoeff 

logmultinomialcoeff(n, *ks)

\(\log \binom{n}{k_1,\dots,k_m} = \log \left(\frac{n!}{k_1! \cdots k_m!}\right)\).
Note that we do not enforce \(k_1+\cdots+k_m=n\).

Parameters:

Name Type Description Default
n Tensor

\(n\).

 required
*ks Tuple

\(k_1,\dots,k_m\) where \(k_j\) is a torch.Tensor.

 ()

Returns:

Name Type Description
y Tensor

\(\log \binom{n}{k_1,\dots,k_m}\).

Examples:

>>> n = torch.arange(6,12)
>>> k1 = torch.arange(0,2)
>>> k2 = torch.arange(2,4)
>>> k3 = torch.arange(4,6)
>>> logmultinomialcoeff(n[None,:],k1[:,None],k2[:,None],k3[:,None])
tensor([[2.7081e+00, 4.6540e+00, 6.7334e+00, 8.9306e+00, 1.1233e+01, 1.3631e+01],
        [8.8818e-16, 1.9459e+00, 4.0254e+00, 6.2226e+00, 8.5252e+00, 1.0923e+01]])
>>> logmultinomialcoeff(n[:,None,None,None],k1[None,:,None,None],k2[None,None,:,None],k3[None,None,None,:]).shape
torch.Size([6, 2, 2, 2])
Source code in agsutil/utils.py 
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
def logmultinomialcoeff(n, *ks):
    r"""
    $\log \binom{n}{k_1,\dots,k_m} = \log \left(\frac{n!}{k_1! \cdots k_m!}\right)$.  
    Note that we do not enforce $k_1+\cdots+k_m=n$.

    Args:
        n (torch.Tensor): $n$.
        *ks (Tuple): $k_1,\dots,k_m$ where $k_j$ is a `torch.Tensor`.

    Returns:
        y (torch.Tensor): $\log \binom{n}{k_1,\dots,k_m}$.

    Examples:
        >>> n = torch.arange(6,12)
        >>> k1 = torch.arange(0,2)
        >>> k2 = torch.arange(2,4)
        >>> k3 = torch.arange(4,6)
        >>> logmultinomialcoeff(n[None,:],k1[:,None],k2[:,None],k3[:,None])
        tensor([[2.7081e+00, 4.6540e+00, 6.7334e+00, 8.9306e+00, 1.1233e+01, 1.3631e+01],
                [8.8818e-16, 1.9459e+00, 4.0254e+00, 6.2226e+00, 8.5252e+00, 1.0923e+01]])
        >>> logmultinomialcoeff(n[:,None,None,None],k1[None,:,None,None],k2[None,None,:,None],k3[None,None,None,:]).shape
        torch.Size([6, 2, 2, 2])
    """ 
    assert (n>=0).all()
    m = len(ks) 
    assert ((k>=0).all() for k in ks)
    y = torch.lgamma(n+1)
    for k in ks:
        y = y-torch.lgamma(k+1)
    return y

multinomialcoeff 

multinomialcoeff(n, *ks)

\(\binom{n}{k_1,\dots,k_m} = \frac{n!}{k_1! \cdots k_m!}\).
Note that we do not enforce \(k_1+\cdots+k_m=n\), so we do not round the result to the nearest integer.

Parameters:

Name Type Description Default
n Tensor

\(n\).

 required
*ks Tuple

\(k_1,\dots,k_m\) where \(k_j\) is a torch.Tensor.

 ()

Returns:

Name Type Description
y Tensor

\(\binom{n}{k_1,\dots,k_m}\).

Examples:

>>> n = torch.arange(6,12)
>>> k1 = torch.arange(0,2)
>>> k2 = torch.arange(2,4)
>>> k3 = torch.arange(4,6)
>>> multinomialcoeff(n[None,:],k1[:,None],k2[:,None],k3[:,None])
tensor([[1.5000e+01, 1.0500e+02, 8.4000e+02, 7.5600e+03, 7.5600e+04, 8.3160e+05],
        [1.0000e+00, 7.0000e+00, 5.6000e+01, 5.0400e+02, 5.0400e+03, 5.5440e+04]])
>>> multinomialcoeff(n[:,None,None,None],k1[None,:,None,None],k2[None,None,:,None],k3[None,None,None,:]).shape
torch.Size([6, 2, 2, 2])
Source code in agsutil/utils.py 
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
def multinomialcoeff(n, *ks):
    r"""
    $\binom{n}{k_1,\dots,k_m} = \frac{n!}{k_1! \cdots k_m!}$.  
    Note that we do not enforce $k_1+\cdots+k_m=n$, so we do not round the result to the nearest integer.

    Args:
        n (torch.Tensor): $n$.
        *ks (Tuple): $k_1,\dots,k_m$ where $k_j$ is a `torch.Tensor`.

    Returns:
        y (torch.Tensor): $\binom{n}{k_1,\dots,k_m}$.

    Examples:
        >>> n = torch.arange(6,12)
        >>> k1 = torch.arange(0,2)
        >>> k2 = torch.arange(2,4)
        >>> k3 = torch.arange(4,6)
        >>> multinomialcoeff(n[None,:],k1[:,None],k2[:,None],k3[:,None])
        tensor([[1.5000e+01, 1.0500e+02, 8.4000e+02, 7.5600e+03, 7.5600e+04, 8.3160e+05],
                [1.0000e+00, 7.0000e+00, 5.6000e+01, 5.0400e+02, 5.0400e+03, 5.5440e+04]])
        >>> multinomialcoeff(n[:,None,None,None],k1[None,:,None,None],k2[None,None,:,None],k3[None,None,None,:]).shape
        torch.Size([6, 2, 2, 2])
    """ 
    return torch.exp(logmultinomialcoeff(n,*ks))

logfactorial 

logfactorial(n)

\(\log(n!)\)

Parameters:

Name Type Description Default
n Tensor

\(n\).

 required

Returns:

Name Type Description
y Tensor

\(\log(n!)\).

Examples:

>>> logfactorial(torch.arange(1,8))
tensor([0.0000, 0.6931, 1.7918, 3.1781, 4.7875, 6.5793, 8.5252])
Source code in agsutil/utils.py 
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
def logfactorial(n):
    r"""
    $\log(n!)$

    Args:
        n (torch.Tensor): $n$.

    Returns:
        y (torch.Tensor): $\log(n!)$.

    Examples:
        >>> logfactorial(torch.arange(1,8))
        tensor([0.0000, 0.6931, 1.7918, 3.1781, 4.7875, 6.5793, 8.5252])
    """ 
    return logmultinomialcoeff(n)

factorial 

factorial(n)

\(n!\)

Parameters:

Name Type Description Default
n Tensor

\(n\).

 required

Returns:

Name Type Description
y Tensor

\(n!\).

Examples:

>>> factorial(torch.arange(1,8))
tensor([   1,    2,    6,   24,  120,  720, 5040])
Source code in agsutil/utils.py 
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
def factorial(n):
    r"""
    $n!$

    Args:
        n (torch.Tensor): $n$.

    Returns:
        y (torch.Tensor): $n!$.

    Examples:
        >>> factorial(torch.arange(1,8))
        tensor([   1,    2,    6,   24,  120,  720, 5040])
    """ 
    return multinomialcoeff(n).round().to(int)

logcomb 

logcomb(n, k)

\(\log \binom{n}{k} = \log \left(\frac{n!}{k!(n-k)!}\right)\)

Parameters:

Name Type Description Default
n Tensor

\(n\).

 required
k Tensor

\(k\).

 required

Returns:

Name Type Description
y Tensor

\(\log \binom{n}{k}\).

Examples:

>>> logcomb(torch.arange(8)[:,None],torch.arange(6)[None,:])
tensor([[0.0000,   -inf,   -inf,   -inf,   -inf,   -inf],
        [0.0000, 0.0000,   -inf,   -inf,   -inf,   -inf],
        [0.0000, 0.6931, 0.0000,   -inf,   -inf,   -inf],
        [0.0000, 1.0986, 1.0986, 0.0000,   -inf,   -inf],
        [0.0000, 1.3863, 1.7918, 1.3863, 0.0000,   -inf],
        [0.0000, 1.6094, 2.3026, 2.3026, 1.6094, 0.0000],
        [0.0000, 1.7918, 2.7081, 2.9957, 2.7081, 1.7918],
        [0.0000, 1.9459, 3.0445, 3.5553, 3.5553, 3.0445]])
Source code in agsutil/utils.py 
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
def logcomb(n,k):
    r"""
    $\log \binom{n}{k} = \log \left(\frac{n!}{k!(n-k)!}\right)$

    Args:
        n (torch.Tensor): $n$.
        k (torch.Tensor): $k$.

    Returns:
        y (torch.Tensor): $\log \binom{n}{k}$.

    Examples:
        >>> logcomb(torch.arange(8)[:,None],torch.arange(6)[None,:])
        tensor([[0.0000,   -inf,   -inf,   -inf,   -inf,   -inf],
                [0.0000, 0.0000,   -inf,   -inf,   -inf,   -inf],
                [0.0000, 0.6931, 0.0000,   -inf,   -inf,   -inf],
                [0.0000, 1.0986, 1.0986, 0.0000,   -inf,   -inf],
                [0.0000, 1.3863, 1.7918, 1.3863, 0.0000,   -inf],
                [0.0000, 1.6094, 2.3026, 2.3026, 1.6094, 0.0000],
                [0.0000, 1.7918, 2.7081, 2.9957, 2.7081, 1.7918],
                [0.0000, 1.9459, 3.0445, 3.5553, 3.5553, 3.0445]])
    """ 
    return logmultinomialcoeff(n,k,n-k)

comb 

comb(n, k)

\(\binom{n}{k} = \frac{n!}{k!(n-k)!}\)

Parameters:

Name Type Description Default
n Tensor

\(n\).

 required
k Tensor

\(k\).

 required

Returns:

Name Type Description
y Tensor

\(\binom{n}{k}\).

Examples:

>>> comb(torch.arange(8)[:,None],torch.arange(6)[None,:])
tensor([[ 1,  0,  0,  0,  0,  0],
        [ 1,  1,  0,  0,  0,  0],
        [ 1,  2,  1,  0,  0,  0],
        [ 1,  3,  3,  1,  0,  0],
        [ 1,  4,  6,  4,  1,  0],
        [ 1,  5, 10, 10,  5,  1],
        [ 1,  6, 15, 20, 15,  6],
        [ 1,  7, 21, 35, 35, 21]])
Source code in agsutil/utils.py 
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
def comb(n,k):
    r"""
    $\binom{n}{k} = \frac{n!}{k!(n-k)!}$

    Args:
        n (torch.Tensor): $n$.
        k (torch.Tensor): $k$.

    Returns:
        y (torch.Tensor): $\binom{n}{k}$.

    Examples:
        >>> comb(torch.arange(8)[:,None],torch.arange(6)[None,:])
        tensor([[ 1,  0,  0,  0,  0,  0],
                [ 1,  1,  0,  0,  0,  0],
                [ 1,  2,  1,  0,  0,  0],
                [ 1,  3,  3,  1,  0,  0],
                [ 1,  4,  6,  4,  1,  0],
                [ 1,  5, 10, 10,  5,  1],
                [ 1,  6, 15, 20, 15,  6],
                [ 1,  7, 21, 35, 35, 21]])
    """ 
    return multinomialcoeff(n,k,n-k).round().to(int)

enumerate_sums 

enumerate_sums(s, t)

Generator of all possible ways to choose \(s \in \mathbb{N}_0\) non-negative integers which sum to \(t \in \mathbb{N}_0\).
There are a total of \(\binom{t+s-1}{s-1} = \frac{(t+s-1)!}{t!(s-1)!}\) choices.

Parameters:

Name Type Description Default
s int

Number of integers to choose.

required
t int

Number of items to allocate.

required

Returns:

Name Type Description
g generator

Genrator of tuples of length \(s\).

Examples:

2 non-negative integers that sum to 3 

>>> for v in enumerate_sums(2,3): 
...     print(v)
(0, 3)
(1, 2)
(2, 1)
(3, 0)
>>> len_enumerate_sums(2,3)
4

3 non-negative integers that sum to 2 

>>> for v in enumerate_sums(3,2): 
...     print(v)
(0, 0, 2)
(0, 1, 1)
(0, 2, 0)
(1, 0, 1)
(1, 1, 0)
(2, 0, 0)
>>> len_enumerate_sums(3,2)
6
Source code in agsutil/utils.py 
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
def enumerate_sums(s, t):
    r"""
    Generator of all possible ways to choose $s \in \mathbb{N}_0$ non-negative integers which sum to $t \in \mathbb{N}_0$.  
    There are a total of $\binom{t+s-1}{s-1} = \frac{(t+s-1)!}{t!(s-1)!}$ choices. 

    Args:
        s (int): Number of integers to choose. 
        t (int): Number of items to allocate. 

    Returns: 
        g (generator): Genrator of tuples of length $s$. 

    Examples: 

    2 non-negative integers that sum to 3 

        >>> for v in enumerate_sums(2,3): 
        ...     print(v)
        (0, 3)
        (1, 2)
        (2, 1)
        (3, 0)
        >>> len_enumerate_sums(2,3)
        4

    3 non-negative integers that sum to 2 

        >>> for v in enumerate_sums(3,2): 
        ...     print(v)
        (0, 0, 2)
        (0, 1, 1)
        (0, 2, 0)
        (1, 0, 1)
        (1, 1, 0)
        (2, 0, 0)
        >>> len_enumerate_sums(3,2)
        6
    """ 
    for combo in itertools.combinations(range(t+s-1),s-1):
        bars = (-1,)+combo+(t+s-1,)
        yield tuple(bars[i+1]-bars[i]-1 for i in range(s))

len_enumerate_sums 

len_enumerate_sums(s, t)

\(\binom{t+s-1}{s-1} = \frac{(t+s-1)!}{t!(s-1)!}\), the number of ways to choose \(s \in \mathbb{N}_0\) non-negative integers which sum to \(t \in \mathbb{N}_0\).

Parameters:

Name Type Description Default
s int

Number of integers to choose.

required
t int

Number of items to allocate.

required

Returns:

Name Type Description
y int

\(\binom{t+s-1}{s-1}\).

Examples:

2 non-negative integers that sum to 3 

>>> len_enumerate_sums(2,3)
4

3 non-negative integers that sum to 2 

>>> len_enumerate_sums(3,2)
6
Source code in agsutil/utils.py 
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
def len_enumerate_sums(s, t):
    r"""
    $\binom{t+s-1}{s-1} = \frac{(t+s-1)!}{t!(s-1)!}$, the number of ways to choose $s \in \mathbb{N}_0$ non-negative integers which sum to $t \in \mathbb{N}_0$.  

    Args:
        s (int): Number of integers to choose. 
        t (int): Number of items to allocate. 

    Returns: 
        y (int): $\binom{t+s-1}{s-1}$. 

    Examples: 

    2 non-negative integers that sum to 3 

        >>> len_enumerate_sums(2,3)
        4

    3 non-negative integers that sum to 2 

        >>> len_enumerate_sums(3,2)
        6
    """ 
    return comb(torch.tensor(t+s-1),torch.tensor(s-1)).item()

enumerate_partitions 

enumerate_partitions(n, max_val=None)

Enumerate all partitions sizes of \(n\) objects.

Parameters:

Name Type Description Default
n int

Number of objects

 required

Returns:

Name Type Description
g generator

Generator of partitions.

Examples:

>>> for p in enumerate_partitions(3):
...     print(p)
[3]
[2, 1]
[1, 1, 1]
>>> for p in enumerate_partitions(5):
...     print(p)
[5]
[4, 1]
[3, 2]
[3, 1, 1]
[2, 2, 1]
[2, 1, 1, 1]
[1, 1, 1, 1, 1]
Source code in agsutil/utils.py 
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
def enumerate_partitions(n, max_val=None):
    r"""
    Enumerate all partitions sizes of $n$ objects.

    Args:
        n (int): Number of objects

    Returns: 
        g (generator): Generator of partitions. 

    Examples: 
        >>> for p in enumerate_partitions(3):
        ...     print(p)
        [3]
        [2, 1]
        [1, 1, 1]
        >>> for p in enumerate_partitions(5):
        ...     print(p)
        [5]
        [4, 1]
        [3, 2]
        [3, 1, 1]
        [2, 2, 1]
        [2, 1, 1, 1]
        [1, 1, 1, 1, 1]
    """
    if max_val is None:
        max_val = n
    if n == 0:
        yield []
        return
    for i in range(min(n,max_val),0,-1):
        for p in enumerate_partitions(n-i,i):
            yield [i]+p

icdf_std_normal 

icdf_std_normal(x)

Inverse CDF of the standard normal distribution \(\mathcal{N}(0,1)\).

Examples:

>>> torch.set_default_dtype(torch.float64)
>>> rng = torch.Generator().manual_seed(7)

>>> x = torch.rand(3,4,generator=rng)
>>> g = icdf_std_normal(x)
>>> g 
tensor([[-0.5847, -0.6017,  1.0898,  0.4034],
        [ 1.4223,  0.9926, -0.5397,  0.1528],
        [ 0.7161,  0.7042, -1.5299, -1.5756]])
>>> g2 = icdf_normal = torch.distributions.Normal(loc=torch.zeros(1),scale=torch.ones(1)).icdf(x)
>>> torch.allclose(g,g2)
True
Source code in agsutil/utils.py 
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
def icdf_std_normal(x):
    r"""
    Inverse CDF of the standard normal distribution $\mathcal{N}(0,1)$. 

    Examples:
        >>> torch.set_default_dtype(torch.float64)
        >>> rng = torch.Generator().manual_seed(7)

        >>> x = torch.rand(3,4,generator=rng)
        >>> g = icdf_std_normal(x)
        >>> g 
        tensor([[-0.5847, -0.6017,  1.0898,  0.4034],
                [ 1.4223,  0.9926, -0.5397,  0.1528],
                [ 0.7161,  0.7042, -1.5299, -1.5756]])
        >>> g2 = icdf_normal = torch.distributions.Normal(loc=torch.zeros(1),scale=torch.ones(1)).icdf(x)
        >>> torch.allclose(g,g2)
        True
    """
    return np.sqrt(2)*torch.erfinv(2*x-1)

matplotlib plotting utils

mpl_setup 

mpl_setup()

Setup matplotlib default parameters

Returns:

Name Type Description
mplparams dict

matplotlib helpful parameters
Source code in agsutil/plots.py 
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
def mpl_setup():
    r""" 
    Setup matplotlib default parameters

    Returns:
        mplparams (dict): matplotlib helpful parameters"""
    from matplotlib import pyplot 
    import seaborn as sns
    pyplot.style.use("seaborn-v0_8-whitegrid")
    sns.set_palette("colorblind") # Options: "deep", "muted", "pastel", "bright", "dark", "colorblind
    COLORS = pyplot.rcParams['axes.prop_cycle'].by_key()['color']
    PW = 30 # page width in inches
    FS = 30 # font size
    LINESTYLES = [
        'solid',
        'dotted',
        'dashdot',
        'dashed',
        (0, (3, 5, 1, 5, 1, 5)),
        (5, (10, 3)),
        (0, (1, 1)),
        (0, (1, 10)),
        (0, (1, 5)),
        (0, (5, 5)),
        ]
    MARKERS = [
        "o",
        "s",
        "D",
        "P",
        "^",
        "v",
        "<",
        ">",
        "$a$",
        "$b$",
        "$c$",
        "$d$",
        ]
    pyplot.rcParams['xtick.labelsize'] = FS
    pyplot.rcParams['ytick.labelsize'] = FS
    pyplot.rcParams['ytick.labelsize'] = FS
    pyplot.rcParams['axes.titlesize'] = FS
    pyplot.rcParams['figure.titlesize'] = FS
    pyplot.rcParams["axes.labelsize"] = FS
    pyplot.rcParams['legend.fontsize'] = FS
    pyplot.rcParams['font.size'] = FS
    pyplot.rcParams['lines.linewidth'] = 5
    pyplot.rcParams['lines.markersize'] = 15
    mplparams = {
        "PW": PW,
        "FS": FS,
        "COLORS": COLORS,
        "LINESTYLES": LINESTYLES,
        "MARKERS": MARKERS,
        }
    return mplparams

set_aspect 

set_aspect(ax, ratio=1)

Set aspect ratio of the ax.

Parameters:

Name Type Description Default
ax Axes

axes to set the aspect ratio of.

required
ratio float

positive aspect ratio for the axis

 1
Source code in agsutil/plots.py 
61
62
63
64
65
66
67
68
69
70
71
72
73
def set_aspect(ax, ratio=1):
    r"""
    Set aspect ratio of the ax. 

    Args:
        ax (Axes): axes to set the aspect ratio of. 
        ratio (float): positive aspect ratio for the axis
    """
    assert ratio>0
    xmin,xmax = ax.get_xlim()
    ymin,ymax = ax.get_ylim()
    aspect = ratio*(xmax-xmin)/(ymax-ymin)
    ax.set_aspect(aspect)