-
Notifications
You must be signed in to change notification settings - Fork 346
/
_trainingplans.py
1108 lines (987 loc) · 38 KB
/
_trainingplans.py
1
2
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
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
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
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
896
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
from functools import partial
from inspect import getfullargspec, signature
from typing import Callable, Dict, Optional, Union
import jax
import jax.numpy as jnp
import numpy as np
import optax
import pyro
import pytorch_lightning as pl
import torch
from pyro.nn import PyroModule
from torch.optim.lr_scheduler import ReduceLROnPlateau
from torchmetrics import MetricCollection
from scvi import REGISTRY_KEYS
from scvi._compat import Literal
from scvi.module import Classifier
from scvi.module.base import (
BaseModuleClass,
JaxModuleWrapper,
LossRecorder,
PyroBaseModuleClass,
TrainStateWithBatchNorm,
)
from scvi.nn import one_hot
from ._metrics import ElboMetric
def _compute_kl_weight(
epoch: int,
step: int,
n_epochs_kl_warmup: Optional[int],
n_steps_kl_warmup: Optional[int],
max_kl_weight: float = 1.0,
min_kl_weight: float = 0.0,
) -> float:
"""
Computes the kl weight for the current step or epoch depending on
`n_epochs_kl_warmup` and `n_steps_kl_warmup`. If both `n_epochs_kl_warmup` and
`n_steps_kl_warmup` are None `max_kl_weight` is returned.
Parameters
----------
epoch
Current epoch.
step
Current step.
n_epochs_kl_warmup
Number of training epochs to scale weight on KL divergences from
`min_kl_weight` to `max_kl_weight`
n_steps_kl_warmup
Number of training steps (minibatches) to scale weight on KL divergences from
`min_kl_weight` to `max_kl_weight`
max_kl_weight
Maximum scaling factor on KL divergence during training.
min_kl_weight
Minimum scaling factor on KL divergence during training.
"""
if min_kl_weight > max_kl_weight:
raise ValueError(
f"min_kl_weight={min_kl_weight} is larger than max_kl_weight={max_kl_weight}."
)
slope = max_kl_weight - min_kl_weight
if n_epochs_kl_warmup:
if epoch < n_epochs_kl_warmup:
return slope * (epoch / n_epochs_kl_warmup) + min_kl_weight
elif n_steps_kl_warmup:
if step < n_steps_kl_warmup:
return slope * (step / n_steps_kl_warmup) + min_kl_weight
return max_kl_weight
class TrainingPlan(pl.LightningModule):
"""
Lightning module task to train scvi-tools modules.
The training plan is a PyTorch Lightning Module that is initialized
with a scvi-tools module object. It configures the optimizers, defines
the training step and validation step, and computes metrics to be recorded
during training. The training step and validation step are functions that
take data, run it through the model and return the loss, which will then
be used to optimize the model parameters in the Trainer. Overall, custom
training plans can be used to develop complex inference schemes on top of
modules.
The following developer tutorial will familiarize you more with training plans
and how to use them: :doc:`/tutorials/notebooks/model_user_guide`.
Parameters
----------
module
A module instance from class ``BaseModuleClass``.
lr
Learning rate used for optimization.
weight_decay
Weight decay used in optimizatoin.
eps
eps used for optimization.
optimizer
One of "Adam" (:class:`~torch.optim.Adam`), "AdamW" (:class:`~torch.optim.AdamW`).
n_steps_kl_warmup
Number of training steps (minibatches) to scale weight on KL divergences from
`min_kl_weight` to `max_kl_weight`. Only activated when `n_epochs_kl_warmup` is
set to None.
n_epochs_kl_warmup
Number of epochs to scale weight on KL divergences from `min_kl_weight` to
`max_kl_weight`. Overrides `n_steps_kl_warmup` when both are not `None`.
reduce_lr_on_plateau
Whether to monitor validation loss and reduce learning rate when validation set
`lr_scheduler_metric` plateaus.
lr_factor
Factor to reduce learning rate.
lr_patience
Number of epochs with no improvement after which learning rate will be reduced.
lr_threshold
Threshold for measuring the new optimum.
lr_scheduler_metric
Which metric to track for learning rate reduction.
lr_min
Minimum learning rate allowed.
max_kl_weight
Maximum scaling factor on KL divergence during training.
min_kl_weight
Minimum scaling factor on KL divergence during training.
**loss_kwargs
Keyword args to pass to the loss method of the `module`.
`kl_weight` should not be passed here and is handled automatically.
"""
def __init__(
self,
module: BaseModuleClass,
lr: float = 1e-3,
weight_decay: float = 1e-6,
eps: float = 0.01,
optimizer: Literal["Adam", "AdamW"] = "Adam",
n_steps_kl_warmup: Union[int, None] = None,
n_epochs_kl_warmup: Union[int, None] = 400,
reduce_lr_on_plateau: bool = False,
lr_factor: float = 0.6,
lr_patience: int = 30,
lr_threshold: float = 0.0,
lr_scheduler_metric: Literal[
"elbo_validation", "reconstruction_loss_validation", "kl_local_validation"
] = "elbo_validation",
lr_min: float = 0,
max_kl_weight: float = 1.0,
min_kl_weight: float = 0.0,
**loss_kwargs,
):
super(TrainingPlan, self).__init__()
self.module = module
self.lr = lr
self.weight_decay = weight_decay
self.eps = eps
self.optimizer_name = optimizer
self.n_steps_kl_warmup = n_steps_kl_warmup
self.n_epochs_kl_warmup = n_epochs_kl_warmup
self.reduce_lr_on_plateau = reduce_lr_on_plateau
self.lr_factor = lr_factor
self.lr_patience = lr_patience
self.lr_scheduler_metric = lr_scheduler_metric
self.lr_threshold = lr_threshold
self.lr_min = lr_min
self.loss_kwargs = loss_kwargs
self.min_kl_weight = min_kl_weight
self.max_kl_weight = max_kl_weight
self._n_obs_training = None
self._n_obs_validation = None
# automatic handling of kl weight
self._loss_args = getfullargspec(self.module.loss)[0]
if "kl_weight" in self._loss_args:
self.loss_kwargs.update({"kl_weight": self.kl_weight})
self.initialize_train_metrics()
self.initialize_val_metrics()
@staticmethod
def _create_elbo_metric_components(mode: str, n_total: Optional[int] = None):
"""Initialize ELBO metric and the metric collection."""
rec_loss = ElboMetric("reconstruction_loss", mode, "obs")
kl_local = ElboMetric("kl_local", mode, "obs")
kl_global = ElboMetric("kl_global", mode, "obs")
# n_total can be 0 if there is no validation set, this won't ever be used
# in that case anyway
n = 1 if n_total is None or n_total < 1 else n_total
elbo = rec_loss + kl_local + (1 / n) * kl_global
elbo.name = f"elbo_{mode}"
collection = MetricCollection(
{metric.name: metric for metric in [elbo, rec_loss, kl_local, kl_global]}
)
return elbo, rec_loss, kl_local, kl_global, collection
def initialize_train_metrics(self):
"""Initialize train related metrics."""
(
self.elbo_train,
self.rec_loss_train,
self.kl_local_train,
self.kl_global_train,
self.train_metrics,
) = self._create_elbo_metric_components(
mode="train", n_total=self.n_obs_training
)
self.elbo_train.reset()
def initialize_val_metrics(self):
"""Initialize val related metrics."""
(
self.elbo_val,
self.rec_loss_val,
self.kl_local_val,
self.kl_global_val,
self.val_metrics,
) = self._create_elbo_metric_components(
mode="validation", n_total=self.n_obs_validation
)
self.elbo_val.reset()
@property
def n_obs_training(self):
"""
Number of observations in the training set.
This will update the loss kwargs for loss rescaling.
Notes
-----
This can get set after initialization
"""
return self._n_obs_training
@n_obs_training.setter
def n_obs_training(self, n_obs: int):
if "n_obs" in self._loss_args:
self.loss_kwargs.update({"n_obs": n_obs})
self._n_obs_training = n_obs
self.initialize_train_metrics()
@property
def n_obs_validation(self):
"""
Number of observations in the validation set.
This will update the loss kwargs for loss rescaling.
Notes
-----
This can get set after initialization
"""
return self._n_obs_validation
@n_obs_validation.setter
def n_obs_validation(self, n_obs: int):
self._n_obs_validation = n_obs
self.initialize_val_metrics()
def forward(self, *args, **kwargs):
"""Passthrough to `model.forward()`."""
return self.module(*args, **kwargs)
@torch.no_grad()
def compute_and_log_metrics(
self,
loss_recorder: LossRecorder,
metrics: MetricCollection,
mode: str,
):
"""
Computes and logs metrics.
Parameters
----------
loss_recorder
LossRecorder object from scvi-tools module
metric_attr_name
The name of the torch metric object to use
mode
Postfix string to add to the metric name of
extra metrics
"""
rec_loss = loss_recorder.reconstruction_loss
n_obs_minibatch = rec_loss.shape[0]
rec_loss = rec_loss.sum()
kl_local = loss_recorder.kl_local.sum()
kl_global = loss_recorder.kl_global
# use the torchmetric object for the ELBO
metrics.update(
reconstruction_loss=rec_loss,
kl_local=kl_local,
kl_global=kl_global,
n_obs_minibatch=n_obs_minibatch,
)
# pytorch lightning handles everything with the torchmetric object
self.log_dict(
metrics,
on_step=False,
on_epoch=True,
batch_size=n_obs_minibatch,
)
# accumlate extra metrics passed to loss recorder
for extra_metric in loss_recorder.extra_metric_attrs:
met = getattr(loss_recorder, extra_metric)
if isinstance(met, torch.Tensor):
if met.shape != torch.Size([]):
raise ValueError("Extra tracked metrics should be 0-d tensors.")
met = met.detach()
self.log(
f"{extra_metric}_{mode}",
met,
on_step=False,
on_epoch=True,
batch_size=n_obs_minibatch,
)
def training_step(self, batch, batch_idx, optimizer_idx=0):
if "kl_weight" in self.loss_kwargs:
self.loss_kwargs.update({"kl_weight": self.kl_weight})
_, _, scvi_loss = self.forward(batch, loss_kwargs=self.loss_kwargs)
self.log("train_loss", scvi_loss.loss, on_epoch=True)
self.compute_and_log_metrics(scvi_loss, self.train_metrics, "train")
return scvi_loss.loss
def validation_step(self, batch, batch_idx):
# loss kwargs here contains `n_obs` equal to n_training_obs
# so when relevant, the actual loss value is rescaled to number
# of training examples
_, _, scvi_loss = self.forward(batch, loss_kwargs=self.loss_kwargs)
self.log("validation_loss", scvi_loss.loss, on_epoch=True)
self.compute_and_log_metrics(scvi_loss, self.val_metrics, "validation")
def configure_optimizers(self):
params = filter(lambda p: p.requires_grad, self.module.parameters())
if self.optimizer_name == "Adam":
optim_cls = torch.optim.Adam
elif self.optimizer_name == "AdamW":
optim_cls = torch.optim.AdamW
else:
raise ValueError("Optimizer not understood.")
optimizer = optim_cls(
params, lr=self.lr, eps=self.eps, weight_decay=self.weight_decay
)
config = {"optimizer": optimizer}
if self.reduce_lr_on_plateau:
scheduler = ReduceLROnPlateau(
optimizer,
patience=self.lr_patience,
factor=self.lr_factor,
threshold=self.lr_threshold,
min_lr=self.lr_min,
threshold_mode="abs",
verbose=True,
)
config.update(
{
"lr_scheduler": scheduler,
"monitor": self.lr_scheduler_metric,
},
)
return config
@property
def kl_weight(self):
"""Scaling factor on KL divergence during training."""
return _compute_kl_weight(
self.current_epoch,
self.global_step,
self.n_epochs_kl_warmup,
self.n_steps_kl_warmup,
self.max_kl_weight,
self.min_kl_weight,
)
class AdversarialTrainingPlan(TrainingPlan):
"""
Train vaes with adversarial loss option to encourage latent space mixing.
Parameters
----------
module
A module instance from class ``BaseModuleClass``.
lr
Learning rate used for optimization :class:`~torch.optim.Adam`.
weight_decay
Weight decay used in :class:`~torch.optim.Adam`.
n_steps_kl_warmup
Number of training steps (minibatches) to scale weight on KL divergences from 0 to 1.
Only activated when `n_epochs_kl_warmup` is set to None.
n_epochs_kl_warmup
Number of epochs to scale weight on KL divergences from 0 to 1.
Overrides `n_steps_kl_warmup` when both are not `None`.
reduce_lr_on_plateau
Whether to monitor validation loss and reduce learning rate when validation set
`lr_scheduler_metric` plateaus.
lr_factor
Factor to reduce learning rate.
lr_patience
Number of epochs with no improvement after which learning rate will be reduced.
lr_threshold
Threshold for measuring the new optimum.
lr_scheduler_metric
Which metric to track for learning rate reduction.
lr_min
Minimum learning rate allowed
adversarial_classifier
Whether to use adversarial classifier in the latent space
scale_adversarial_loss
Scaling factor on the adversarial components of the loss.
By default, adversarial loss is scaled from 1 to 0 following opposite of
kl warmup.
**loss_kwargs
Keyword args to pass to the loss method of the `module`.
`kl_weight` should not be passed here and is handled automatically.
"""
def __init__(
self,
module: BaseModuleClass,
lr=1e-3,
weight_decay=1e-6,
n_steps_kl_warmup: Union[int, None] = None,
n_epochs_kl_warmup: Union[int, None] = 400,
reduce_lr_on_plateau: bool = False,
lr_factor: float = 0.6,
lr_patience: int = 30,
lr_threshold: float = 0.0,
lr_scheduler_metric: Literal[
"elbo_validation", "reconstruction_loss_validation", "kl_local_validation"
] = "elbo_validation",
lr_min: float = 0,
adversarial_classifier: Union[bool, Classifier] = False,
scale_adversarial_loss: Union[float, Literal["auto"]] = "auto",
**loss_kwargs,
):
super().__init__(
module=module,
lr=lr,
weight_decay=weight_decay,
n_steps_kl_warmup=n_steps_kl_warmup,
n_epochs_kl_warmup=n_epochs_kl_warmup,
reduce_lr_on_plateau=reduce_lr_on_plateau,
lr_factor=lr_factor,
lr_patience=lr_patience,
lr_threshold=lr_threshold,
lr_scheduler_metric=lr_scheduler_metric,
lr_min=lr_min,
**loss_kwargs,
)
if adversarial_classifier is True:
self.n_output_classifier = self.module.n_batch
self.adversarial_classifier = Classifier(
n_input=self.module.n_latent,
n_hidden=32,
n_labels=self.n_output_classifier,
n_layers=2,
logits=True,
)
else:
self.adversarial_classifier = adversarial_classifier
self.scale_adversarial_loss = scale_adversarial_loss
def loss_adversarial_classifier(self, z, batch_index, predict_true_class=True):
n_classes = self.n_output_classifier
cls_logits = torch.nn.LogSoftmax(dim=1)(self.adversarial_classifier(z))
if predict_true_class:
cls_target = one_hot(batch_index, n_classes)
else:
one_hot_batch = one_hot(batch_index, n_classes)
cls_target = torch.zeros_like(one_hot_batch)
# place zeroes where true label is
cls_target.masked_scatter_(
~one_hot_batch.bool(), torch.ones_like(one_hot_batch) / (n_classes - 1)
)
l_soft = cls_logits * cls_target
loss = -l_soft.sum(dim=1).mean()
return loss
def training_step(self, batch, batch_idx, optimizer_idx=0):
if "kl_weight" in self.loss_kwargs:
self.loss_kwargs.update({"kl_weight": self.kl_weight})
kappa = (
1 - self.kl_weight
if self.scale_adversarial_loss == "auto"
else self.scale_adversarial_loss
)
batch_tensor = batch[REGISTRY_KEYS.BATCH_KEY]
if optimizer_idx == 0:
inference_outputs, _, scvi_loss = self.forward(
batch, loss_kwargs=self.loss_kwargs
)
loss = scvi_loss.loss
# fool classifier if doing adversarial training
if kappa > 0 and self.adversarial_classifier is not False:
z = inference_outputs["z"]
fool_loss = self.loss_adversarial_classifier(z, batch_tensor, False)
loss += fool_loss * kappa
self.log("train_loss", loss, on_epoch=True)
self.compute_and_log_metrics(scvi_loss, self.train_metrics, "train")
return loss
# train adversarial classifier
# this condition will not be met unless self.adversarial_classifier is not False
if optimizer_idx == 1:
inference_inputs = self.module._get_inference_input(batch)
outputs = self.module.inference(**inference_inputs)
z = outputs["z"]
loss = self.loss_adversarial_classifier(z.detach(), batch_tensor, True)
loss *= kappa
return loss
def configure_optimizers(self):
params1 = filter(lambda p: p.requires_grad, self.module.parameters())
optimizer1 = torch.optim.Adam(
params1, lr=self.lr, eps=0.01, weight_decay=self.weight_decay
)
config1 = {"optimizer": optimizer1}
if self.reduce_lr_on_plateau:
scheduler1 = ReduceLROnPlateau(
optimizer1,
patience=self.lr_patience,
factor=self.lr_factor,
threshold=self.lr_threshold,
min_lr=self.lr_min,
threshold_mode="abs",
verbose=True,
)
config1.update(
{
"lr_scheduler": scheduler1,
"monitor": self.lr_scheduler_metric,
},
)
if self.adversarial_classifier is not False:
params2 = filter(
lambda p: p.requires_grad, self.adversarial_classifier.parameters()
)
optimizer2 = torch.optim.Adam(
params2, lr=1e-3, eps=0.01, weight_decay=self.weight_decay
)
config2 = {"optimizer": optimizer2}
# bug in pytorch lightning requires this way to return
opts = [config1.pop("optimizer"), config2["optimizer"]]
if "lr_scheduler" in config1:
config1["scheduler"] = config1.pop("lr_scheduler")
scheds = [config1]
return opts, scheds
else:
return opts
return config1
class SemiSupervisedTrainingPlan(TrainingPlan):
"""
Lightning module task for SemiSupervised Training.
Parameters
----------
module
A module instance from class ``BaseModuleClass``.
classification_ratio
Weight of the classification_loss in loss function
lr
Learning rate used for optimization :class:`~torch.optim.Adam`.
weight_decay
Weight decay used in :class:`~torch.optim.Adam`.
n_steps_kl_warmup
Number of training steps (minibatches) to scale weight on KL divergences from 0 to 1.
Only activated when `n_epochs_kl_warmup` is set to None.
n_epochs_kl_warmup
Number of epochs to scale weight on KL divergences from 0 to 1.
Overrides `n_steps_kl_warmup` when both are not `None`.
reduce_lr_on_plateau
Whether to monitor validation loss and reduce learning rate when validation set
`lr_scheduler_metric` plateaus.
lr_factor
Factor to reduce learning rate.
lr_patience
Number of epochs with no improvement after which learning rate will be reduced.
lr_threshold
Threshold for measuring the new optimum.
lr_scheduler_metric
Which metric to track for learning rate reduction.
**loss_kwargs
Keyword args to pass to the loss method of the `module`.
`kl_weight` should not be passed here and is handled automatically.
"""
def __init__(
self,
module: BaseModuleClass,
classification_ratio: int = 50,
lr=1e-3,
weight_decay=1e-6,
n_steps_kl_warmup: Union[int, None] = None,
n_epochs_kl_warmup: Union[int, None] = 400,
reduce_lr_on_plateau: bool = False,
lr_factor: float = 0.6,
lr_patience: int = 30,
lr_threshold: float = 0.0,
lr_scheduler_metric: Literal[
"elbo_validation", "reconstruction_loss_validation", "kl_local_validation"
] = "elbo_validation",
**loss_kwargs,
):
super(SemiSupervisedTrainingPlan, self).__init__(
module=module,
lr=lr,
weight_decay=weight_decay,
n_steps_kl_warmup=n_steps_kl_warmup,
n_epochs_kl_warmup=n_epochs_kl_warmup,
reduce_lr_on_plateau=reduce_lr_on_plateau,
lr_factor=lr_factor,
lr_patience=lr_patience,
lr_threshold=lr_threshold,
lr_scheduler_metric=lr_scheduler_metric,
**loss_kwargs,
)
self.loss_kwargs.update({"classification_ratio": classification_ratio})
def training_step(self, batch, batch_idx, optimizer_idx=0):
# Potentially dangerous if batch is from a single dataloader with two keys
if len(batch) == 2:
full_dataset = batch[0]
labelled_dataset = batch[1]
else:
full_dataset = batch
labelled_dataset = None
if "kl_weight" in self.loss_kwargs:
self.loss_kwargs.update({"kl_weight": self.kl_weight})
input_kwargs = dict(
feed_labels=False,
labelled_tensors=labelled_dataset,
)
input_kwargs.update(self.loss_kwargs)
_, _, scvi_losses = self.forward(full_dataset, loss_kwargs=input_kwargs)
loss = scvi_losses.loss
self.log(
"train_loss",
loss,
on_epoch=True,
batch_size=len(scvi_losses.reconstruction_loss),
)
self.compute_and_log_metrics(scvi_losses, self.train_metrics, "train")
return loss
def validation_step(self, batch, batch_idx, optimizer_idx=0):
# Potentially dangerous if batch is from a single dataloader with two keys
if len(batch) == 2:
full_dataset = batch[0]
labelled_dataset = batch[1]
else:
full_dataset = batch
labelled_dataset = None
input_kwargs = dict(
feed_labels=False,
labelled_tensors=labelled_dataset,
)
input_kwargs.update(self.loss_kwargs)
_, _, scvi_losses = self.forward(full_dataset, loss_kwargs=input_kwargs)
loss = scvi_losses.loss
self.log(
"validation_loss",
loss,
on_epoch=True,
batch_size=len(scvi_losses.reconstruction_loss),
)
self.compute_and_log_metrics(scvi_losses, self.val_metrics, "validation")
class PyroTrainingPlan(pl.LightningModule):
"""
Lightning module task to train Pyro scvi-tools modules.
Parameters
----------
pyro_module
An instance of :class:`~scvi.module.base.PyroBaseModuleClass`. This object
should have callable `model` and `guide` attributes or methods.
loss_fn
A Pyro loss. Should be a subclass of :class:`~pyro.infer.ELBO`.
If `None`, defaults to :class:`~pyro.infer.Trace_ELBO`.
optim
A Pyro optimizer instance, e.g., :class:`~pyro.optim.Adam`. If `None`,
defaults to :class:`pyro.optim.Adam` optimizer with a learning rate of `1e-3`.
optim_kwargs
Keyword arguments for **default** optimiser :class:`pyro.optim.Adam`.
n_steps_kl_warmup
Number of training steps (minibatches) to scale weight on KL divergences from 0 to 1.
Only activated when `n_epochs_kl_warmup` is set to None.
n_epochs_kl_warmup
Number of epochs to scale weight on KL divergences from 0 to 1.
Overrides `n_steps_kl_warmup` when both are not `None`.
scale_elbo
Scale ELBO using :class:`~pyro.poutine.scale`. Potentially useful for avoiding
numerical inaccuracy when working with very large ELBO.
"""
def __init__(
self,
pyro_module: PyroBaseModuleClass,
loss_fn: Optional[pyro.infer.ELBO] = None,
optim: Optional[pyro.optim.PyroOptim] = None,
optim_kwargs: Optional[dict] = None,
n_steps_kl_warmup: Union[int, None] = None,
n_epochs_kl_warmup: Union[int, None] = 400,
scale_elbo: float = 1.0,
):
super().__init__()
self.module = pyro_module
self._n_obs_training = None
optim_kwargs = optim_kwargs if isinstance(optim_kwargs, dict) else dict()
if "lr" not in optim_kwargs.keys():
optim_kwargs.update({"lr": 1e-3})
self.loss_fn = pyro.infer.Trace_ELBO() if loss_fn is None else loss_fn
self.optim = (
pyro.optim.Adam(optim_args=optim_kwargs) if optim is None else optim
)
self.n_steps_kl_warmup = n_steps_kl_warmup
self.n_epochs_kl_warmup = n_epochs_kl_warmup
self.automatic_optimization = False
self.use_kl_weight = False
if isinstance(self.module.model, PyroModule):
self.use_kl_weight = (
"kl_weight" in signature(self.module.model.forward).parameters
)
elif callable(self.module.model):
self.use_kl_weight = "kl_weight" in signature(self.module.model).parameters
def scale(pyro_obj):
if scale_elbo == 1:
return pyro_obj
else:
return pyro.poutine.scale(pyro_obj, scale_elbo)
self.svi = pyro.infer.SVI(
model=scale(self.module.model),
guide=scale(self.module.guide),
optim=self.optim,
loss=self.loss_fn,
)
self._dummy_param = torch.nn.Parameter(torch.Tensor([0.0]))
@property
def n_obs_training(self):
"""
Number of training examples.
If not `None`, updates the `n_obs` attr
of the Pyro module's `model` and `guide`, if they exist.
"""
return self._n_obs_training
@n_obs_training.setter
def n_obs_training(self, n_obs: int):
# important for scaling log prob in Pyro plates
if n_obs is not None:
if hasattr(self.module.model, "n_obs"):
setattr(self.module.model, "n_obs", n_obs)
if hasattr(self.module.guide, "n_obs"):
setattr(self.module.guide, "n_obs", n_obs)
self._n_obs_training = n_obs
def forward(self, *args, **kwargs):
"""Passthrough to `model.forward()`."""
return self.module(*args, **kwargs)
def training_step(self, batch, batch_idx):
args, kwargs = self.module._get_fn_args_from_batch(batch)
# Set KL weight if necessary.
# Note: if applied, ELBO loss in progress bar is the effective KL annealed loss, not the true ELBO.
if self.use_kl_weight:
kwargs.update({"kl_weight": self.kl_weight})
# pytorch lightning requires a Tensor object for loss
loss = torch.Tensor([self.svi.step(*args, **kwargs)])
_opt = self.optimizers()
_opt.step()
return {"loss": loss}
def training_epoch_end(self, outputs):
elbo = 0
n = 0
for out in outputs:
elbo += out["loss"]
n += 1
elbo /= n
self.log("elbo_train", elbo, prog_bar=True)
def configure_optimizers(self):
"""
PyTorch Lightning shim optimizer.
PyTorch Lightning wants to take steps on an optimizer
returned by this function in order to increment the global
step count. See PyTorch Lighinting optimizer manual loop.
Here we provide a shim optimizer that we can take steps on
at minimal computational cost in order to keep Lightning happy :).
"""
return torch.optim.Adam([self._dummy_param])
def optimizer_step(self, *args, **kwargs):
pass
def backward(self, *args, **kwargs):
pass
@property
def kl_weight(self):
"""Scaling factor on KL divergence during training."""
return _compute_kl_weight(
self.current_epoch,
self.global_step,
self.n_epochs_kl_warmup,
self.n_steps_kl_warmup,
min_kl_weight=1e-3,
)
class ClassifierTrainingPlan(pl.LightningModule):
"""
Lightning module task to train a simple MLP classifier.
Parameters
----------
classifier
A model instance from :class:`~scvi.module.Classifier`.
lr
Learning rate used for optimization.
weight_decay
Weight decay used in optimizatoin.
eps
eps used for optimization.
optimizer
One of "Adam" (:class:`~torch.optim.Adam`), "AdamW" (:class:`~torch.optim.AdamW`).
data_key
Key for classifier input in tensor dict minibatch
labels_key
Key for classifier label in tensor dict minibatch
loss
PyTorch loss to use
"""
def __init__(
self,
classifier: BaseModuleClass,
lr: float = 1e-3,
weight_decay: float = 1e-6,
eps: float = 0.01,
optimizer: Literal["Adam", "AdamW"] = "Adam",
data_key: str = REGISTRY_KEYS.X_KEY,
labels_key: str = REGISTRY_KEYS.LABELS_KEY,
loss: Callable = torch.nn.CrossEntropyLoss,
):
super().__init__()
self.module = classifier
self.lr = lr
self.weight_decay = weight_decay
self.eps = eps
self.optimizer_name = optimizer
self.data_key = data_key
self.labels_key = labels_key
self.loss_fn = loss()
if self.module.logits is False and loss == torch.nn.CrossEntropyLoss:
raise UserWarning(
"classifier should return logits when using CrossEntropyLoss."
)
def forward(self, *args, **kwargs):
"""Passthrough to `model.forward()`."""
return self.module(*args, **kwargs)
def training_step(self, batch, batch_idx, optimizer_idx=0):
soft_prediction = self.forward(batch[self.data_key])
loss = self.loss_fn(soft_prediction, batch[self.labels_key].view(-1).long())
self.log("train_loss", loss, on_epoch=True)
return loss
def validation_step(self, batch, batch_idx):
soft_prediction = self.forward(batch[self.data_key])
loss = self.loss_fn(soft_prediction, batch[self.labels_key].view(-1).long())
self.log("validation_loss", loss)
return loss
def configure_optimizers(self):
params = filter(lambda p: p.requires_grad, self.module.parameters())
if self.optimizer_name == "Adam":
optim_cls = torch.optim.Adam
elif self.optimizer_name == "AdamW":
optim_cls = torch.optim.AdamW
else:
raise ValueError("Optimizer not understood.")
optimizer = optim_cls(
params, lr=self.lr, eps=self.eps, weight_decay=self.weight_decay
)
return optimizer
class JaxTrainingPlan(pl.LightningModule):
"""
Lightning module task to train Pyro scvi-tools modules.
Parameters
----------
module
An instance of :class:`~scvi.module.base.JaxModuleWraper`.
n_steps_kl_warmup
Number of training steps (minibatches) to scale weight on KL divergences from 0 to 1.
Only activated when `n_epochs_kl_warmup` is set to None.
n_epochs_kl_warmup
Number of epochs to scale weight on KL divergences from 0 to 1.
Overrides `n_steps_kl_warmup` when both are not `None`.
"""
def __init__(
self,
module: JaxModuleWrapper,
n_steps_kl_warmup: Union[int, None] = None,
n_epochs_kl_warmup: Union[int, None] = 400,
optim_kwargs: Optional[dict] = None,
**loss_kwargs,
):
super().__init__()
self.module = module
self._n_obs_training = None
self.loss_kwargs = loss_kwargs
self.n_steps_kl_warmup = n_steps_kl_warmup
self.n_epochs_kl_warmup = n_epochs_kl_warmup
self.automatic_optimization = False
# automatic handling of kl weight
self._loss_args = signature(self.module.loss).parameters
if "kl_weight" in self._loss_args:
self.loss_kwargs.update({"kl_weight": self.kl_weight})
# set optim kwargs
self.optim_kwargs = dict(learning_rate=1e-3, eps=0.01, weight_decay=1e-6)
if optim_kwargs is not None:
self.optim_kwargs.update(optim_kwargs)
def set_train_state(self, params, batch_stats=None):
if self.module.train_state is not None:
return
weight_decay = self.optim_kwargs.pop("weight_decay")
# replicates PyTorch Adam
optimizer = optax.chain(
optax.additive_weight_decay(weight_decay=weight_decay),
optax.adam(**self.optim_kwargs),
)
train_state = TrainStateWithBatchNorm.create(
apply_fn=self.module.apply,
params=params,
tx=optimizer,
batch_stats=batch_stats,
)
self.module.train_state = train_state
@staticmethod
@jax.jit
def jit_training_step(
state: TrainStateWithBatchNorm,
batch: Dict[str, np.ndarray],
rngs: Dict[str, jnp.ndarray],
**kwargs,
):
# batch stats can't be passed here
def loss_fn(params):