-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy path_patched_UNet_trainer.py
634 lines (493 loc) · 30.1 KB
/
_patched_UNet_trainer.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
from ._base_trainer import _BaseTrainer, MeasureMemory
import pathlib
import torch.multiprocessing as mp
import torch
from torch import nn
import horovod.torch as hvd
import numpy as np
import xarray as xr
import itertools
from .flow_dataset import FlowDataset
from .unet import UNet
from .visualization import save_flows
from .converter import save_as_netcdf
class PatchedUNetTrainer(_BaseTrainer):
def __init__(self, **kwargs):
super().__init__(**kwargs)
self.model_name = 'patched_UNet'
def _initialize(self, **kwargs):
# Horovod: Initialize library
hvd.init()
torch.manual_seed(self.seed)
if self.device == 'cuda':
# Horovod: Pin GPU to be used to process local rank (one GPU per process)
torch.cuda.set_device(hvd.local_rank())
torch.cuda.manual_seed(self.seed)
# Horovod: limit # of CPU threads to be used per worker.
torch.set_num_threads(1)
self.rank, self.size = hvd.rank(), hvd.size()
self.master = self.rank == 0
super()._prepare_dirs()
self.train_loader, self.val_loader, self.test_loader = super()._dataloaders()
self.model0, self.model1, self.model2 = self._get_model(self.run_number)
self.model0 = self.model0.to(self.device)
self.model1 = self.model1.to(self.device)
self.model2 = self.model2.to(self.device)
## Optimizers
# By default, Adasum doesn't need scaling up leraning rate
lr_scaler = hvd.size() if not self.use_adasum else 1
if self.device == 'cuda' and self.use_adasum and hvd.nccl_built():
lr_scaler = hvd.local_size()
lr = self.lr * lr_scaler
self.opt0 = torch.optim.Adam(self.model0.parameters(), lr=lr, betas=(self.beta_1, self.beta_2))
self.opt1 = torch.optim.Adam(self.model1.parameters(), lr=lr, betas=(self.beta_1, self.beta_2))
self.opt2 = torch.optim.Adam(self.model2.parameters(), lr=lr, betas=(self.beta_1, self.beta_2))
# Horovod: broadcast parameters & optimizer state.
hvd.broadcast_parameters(self.model0.state_dict(), root_rank=0)
hvd.broadcast_parameters(self.model1.state_dict(), root_rank=0)
hvd.broadcast_parameters(self.model2.state_dict(), root_rank=0)
hvd.broadcast_optimizer_state(self.opt0, root_rank=0)
hvd.broadcast_optimizer_state(self.opt1, root_rank=0)
hvd.broadcast_optimizer_state(self.opt2, root_rank=0)
# Horovod: (optional) compression algorithm.
compression = hvd.Compression.fp16 if self.fp16_allreduce else hvd.Compression.none
# Horovod: wrap optimizer with DistributedOptimizer.
self.opt0 = hvd.DistributedOptimizer(self.opt0,
named_parameters=self.model0.named_parameters(),
compression=compression,
op=hvd.Adasum if self.use_adasum else hvd.Average,
gradient_predivide_factor=self.gradient_predivide_factor)
self.opt1 = hvd.DistributedOptimizer(self.opt1,
named_parameters=self.model1.named_parameters(),
compression=compression,
op=hvd.Adasum if self.use_adasum else hvd.Average,
gradient_predivide_factor=self.gradient_predivide_factor)
self.opt2 = hvd.DistributedOptimizer(self.opt2,
named_parameters=self.model2.named_parameters(),
compression=compression,
op=hvd.Adasum if self.use_adasum else hvd.Average,
gradient_predivide_factor=self.gradient_predivide_factor)
self.criterion = nn.L1Loss() if self.loss_type == 'mae_loss' else nn.MSELoss(reduction='mean')
# Set normalization coefficients
super()._set_normalization_coefs(shape=[1,1,1,-1,1,1])
# Memory measurement
self.memory = MeasureMemory(device=self.device)
# Synchronize
if self.device == 'cuda':
torch.cuda.synchronize() # Waits for everything to finish running
def _initialize_for_inference(self, **kwargs):
# Set output directory
super()._prepare_dirs()
self.train_loader, self.val_loader, self.test_loader = super()._dataloaders()
self.model0, self.model1, self.model2 = self._get_model(self.run_number)
self.model0 = self.model0.to(self.device)
self.model1 = self.model1.to(self.device)
self.model2 = self.model2.to(self.device)
# Set normalization coefficients
super()._set_normalization_coefs(shape=[1,1,1,-1,1,1])
# Memory measurement
self.memory = MeasureMemory(device=self.device)
# Synchronize
if self.device == 'cuda':
torch.cuda.synchronize() # Waits for everything to finish running
def _get_model(self, run_number):
model0 = UNet(dim=self.dim, padding_mode=self.padding_mode)
model1 = UNet(dim=self.dim, padding_mode=self.padding_mode)
model2 = UNet(dim=self.dim, padding_mode=self.padding_mode)
if self.inference_mode:
self.epoch_start = self.load_nth_state_file
# To load the state file for inference
rank = 0
model0.load_state_dict( torch.load(f'{self.state_file_dir}/model0_{rank}_{self.epoch_start:03}.pt') )
model1.load_state_dict( torch.load(f'{self.state_file_dir}/model1_{rank}_{self.epoch_start:03}.pt') )
model2.load_state_dict( torch.load(f'{self.state_file_dir}/model2_{rank}_{self.epoch_start:03}.pt') )
else:
self.epoch_start = 0
if run_number > 0:
if self.master:
print(f'restart, {run_number}')
# Load model states from previous run
prev_run_number = run_number - 1
prev_result_filename = self.out_dir / f'flow_cnn_result_rank{self.rank}_rst{prev_run_number:03}.h5'
if not prev_result_filename.is_file():
raise IOError(f'prev_result_filename')
ds_prev = xr.open_dataset(prev_result_filename, engine='netcdf4')
# To load the previous files
epoch_end = ds_prev.attrs['epoch_end']
model0.load_state_dict( torch.load(f'{self.model_dir}/model0_{self.rank}_{epoch_end:03}.pt') )
model1.load_state_dict( torch.load(f'{self.model_dir}/model1_{self.rank}_{epoch_end:03}.pt') )
model2.load_state_dict( torch.load(f'{self.model_dir}/model2_{self.rank}_{epoch_end:03}.pt') )
# Next epoch should start from epoch_end + 1
self.epoch_start = int(epoch_end) + 1
return model0, model1, model2
def _save_models(self, total_epoch):
torch.save(self.model0.state_dict(), f'{self.model_dir}/model0_{self.rank}_{total_epoch:03}.pt')
torch.save(self.model1.state_dict(), f'{self.model_dir}/model1_{self.rank}_{total_epoch:03}.pt')
torch.save(self.model2.state_dict(), f'{self.model_dir}/model2_{self.rank}_{total_epoch:03}.pt')
########### Main scripts
def _train(self, data_loader, epoch):
name = 'train'
self.model0.train()
self.model1.train()
self.model2.train()
log_loss = [0] * 3
nb_samples = len(data_loader.sampler)
for i, (sdf, flows) in enumerate(data_loader):
# Load data and meta-data
sdf_Lv0, sdf_Lv1, sdf_Lv2 = sdf
flows_Lv0, flows_Lv1, flows_Lv2 = flows
_, patch_y_Lv1, patch_x_Lv1, *_ = sdf_Lv1.shape
_, patch_y_Lv2, patch_x_Lv2, *_ = sdf_Lv2.shape
# Number of sub patches in each level
nb_patches_Lv2 = patch_y_Lv2 * patch_x_Lv2
nb_patches_Lv1 = patch_y_Lv1 * patch_x_Lv1
# Sub patch inside the Lv1 patch
patch_y_Lv2 = patch_y_Lv2 // patch_y_Lv1
patch_x_Lv2 = patch_x_Lv2 // patch_x_Lv1
batch_len = len(sdf_Lv0)
## To device
self.timer.start()
sdf_Lv0, sdf_Lv1, sdf_Lv2 = sdf_Lv0.to(self.device), sdf_Lv1.to(self.device), sdf_Lv2.to(self.device)
flows_Lv0, flows_Lv1, flows_Lv2 = flows_Lv0.to(self.device), flows_Lv1.to(self.device), flows_Lv2.to(self.device)
self.timer.stop()
self.elapsed_times[f'MemcpyH2D_{name}'].append(self.timer.elapsed_seconds())
# Keep sdfs on CPUs
sdf_Lv0_cpu = sdf_Lv0.to('cpu')
sdf_Lv1_cpu = sdf_Lv1.to('cpu')
sdf_Lv2_cpu = sdf_Lv2.to('cpu')
## Normalization or standardization
sdf_Lv0 = super()._preprocess(sdf_Lv0, self.sdf_Lv0_var0, self.sdf_Lv0_var1)
sdf_Lv1 = super()._preprocess(sdf_Lv1, self.sdf_Lv1_var0, self.sdf_Lv1_var1)
sdf_Lv2 = super()._preprocess(sdf_Lv2, self.sdf_Lv2_var0, self.sdf_Lv2_var1)
flows_Lv0 = super()._preprocess(flows_Lv0, self.flows_Lv0_var0, self.flows_Lv0_var1)
flows_Lv1 = super()._preprocess(flows_Lv1, self.flows_Lv1_var0, self.flows_Lv1_var1)
flows_Lv2 = super()._preprocess(flows_Lv2, self.flows_Lv2_var0, self.flows_Lv2_var1)
# Objectives: construct pred_flows_Lv0-Lv2
pred_flows_Lv0_ = torch.zeros_like(flows_Lv0, device='cpu')
pred_flows_Lv1_ = torch.zeros_like(flows_Lv1, device='cpu')
pred_flows_Lv2_ = torch.zeros_like(flows_Lv2, device='cpu')
#### Train Lv0
self.timer.start()
sdf_Lv0_ = sdf_Lv0[:, 0, 0]
flows_Lv0_ = flows_Lv0[:, 0, 0]
### Update weights
pred_flows_Lv0 = self.model0(sdf_Lv0_)
loss_mae = self.criterion(pred_flows_Lv0, flows_Lv0_)
self.opt0.zero_grad()
loss_mae.backward()
self.opt0.step()
### Log losses
level = 0
log_loss[level] += loss_mae.item() / nb_samples
### Destandardization and save
pred_flows_Lv0 = super()._postprocess(pred_flows_Lv0, self.flows_Lv0_var0, self.flows_Lv0_var1)
pred_flows_Lv0_[:, 0, 0, :, :, :] = pred_flows_Lv0.detach().cpu()
self.timer.stop()
self.elapsed_times[f'{name}_Lv{level}'].append(self.timer.elapsed_seconds())
### Train Lv1
for iy_Lv1, ix_Lv1 in itertools.product(range(patch_y_Lv1), range(patch_x_Lv1)):
self.timer.start()
sdf_Lv1_ = sdf_Lv1[:, iy_Lv1, ix_Lv1]
flows_Lv1_ = flows_Lv1[:, iy_Lv1, ix_Lv1]
### Update weights
pred_flows_Lv1 = self.model1(sdf_Lv1_)
loss_mae = self.criterion(pred_flows_Lv1, flows_Lv1_)
self.opt1.zero_grad()
loss_mae.backward()
self.opt1.step()
### Log losses
level = 1
log_loss[level] += loss_mae.item() / (nb_samples * nb_patches_Lv1)
pred_flows_Lv1 = super()._postprocess(pred_flows_Lv1, self.flows_Lv1_var0, self.flows_Lv1_var1)
pred_flows_Lv1_[:, iy_Lv1, ix_Lv1, :, :, :] = pred_flows_Lv1.detach().cpu()
self.timer.stop()
self.elapsed_times[f'{name}_Lv{level}'].append(self.timer.elapsed_seconds())
### Train Lv2
for iy_Lv2, ix_Lv2 in itertools.product(range(patch_y_Lv2), range(patch_x_Lv2)):
self.timer.start()
global_ix_Lv2 = ix_Lv2 + (ix_Lv1 * patch_x_Lv2)
global_iy_Lv2 = iy_Lv2 + (iy_Lv1 * patch_y_Lv2)
sdf_Lv2_ = sdf_Lv2[:, global_iy_Lv2, global_ix_Lv2]
flows_Lv2_ = flows_Lv2[:, global_iy_Lv2, global_ix_Lv2]
### Update generator weights
pred_flows_Lv2 = self.model2(sdf_Lv2_)
loss_mae = self.criterion(pred_flows_Lv2, flows_Lv2_)
self.opt2.zero_grad()
### Measure memory usage before backward
self.memory.measure()
if 'reserved' not in self.memory_consumption:
self.memory_consumption['reserved'] = self.memory.reserved()
self.memory_consumption['alloc'] = self.memory.alloc()
loss_mae.backward()
self.opt2.step()
### Log losses
level = 2
log_loss[level] += loss_mae.item() / (nb_samples * nb_patches_Lv2)
### Destandardization and save
pred_flows_Lv2 = super()._postprocess(pred_flows_Lv2, self.flows_Lv2_var0, self.flows_Lv2_var1)
pred_flows_Lv2_[:, global_iy_Lv2, global_ix_Lv2, :, :, :] = pred_flows_Lv2.detach().cpu()
self.timer.stop()
self.elapsed_times[f'{name}_Lv{level}'].append(self.timer.elapsed_seconds())
# Saving figures
if i == 0:
self.timer.start()
flows_Lv0 = super()._postprocess(flows_Lv0, self.flows_Lv0_var0, self.flows_Lv0_var1)
flows_Lv1 = super()._postprocess(flows_Lv1, self.flows_Lv1_var0, self.flows_Lv1_var1)
flows_Lv2 = super()._postprocess(flows_Lv2, self.flows_Lv2_var0, self.flows_Lv2_var1)
### Zeros inside objects
pred_flows_Lv0_ = super()._zeros_inside_objects(pred_flows_Lv0_, sdf_Lv0_cpu)
pred_flows_Lv1_ = super()._zeros_inside_objects(pred_flows_Lv1_, sdf_Lv1_cpu)
pred_flows_Lv2_ = super()._zeros_inside_objects(pred_flows_Lv2_, sdf_Lv2_cpu)
### Lv0 figures
level = 0
save_flows(flows_Lv0, name=name, img_dir = self.sub_img_dir, type_name = 'ref', level = level, epoch=epoch)
save_flows(pred_flows_Lv0_, name=name, img_dir = self.sub_img_dir, type_name = 'pred', level = level, epoch=epoch)
save_flows(pred_flows_Lv0_-flows_Lv0.cpu(), name=name, img_dir = self.sub_img_dir, type_name = 'error', level = level, epoch=epoch)
### Lv1 figures
level = 1
save_flows(flows_Lv1, name=name, img_dir = self.sub_img_dir, type_name = 'ref', level = level, epoch=epoch)
save_flows(pred_flows_Lv1_, name=name, img_dir = self.sub_img_dir, type_name = 'pred', level = level, epoch=epoch)
save_flows(pred_flows_Lv1_-flows_Lv1.cpu(), name=name, img_dir = self.sub_img_dir, type_name = 'error', level = level, epoch=epoch)
### Lv2 figures
level = 2
save_flows(flows_Lv2, name=name, img_dir = self.sub_img_dir, type_name = 'ref', level = level, epoch=epoch)
save_flows(pred_flows_Lv2_, name=name, img_dir = self.sub_img_dir, type_name = 'pred', level = level, epoch=epoch)
save_flows(pred_flows_Lv2_-flows_Lv2.cpu(), name=name, img_dir = self.sub_img_dir, type_name = 'error', level = level, epoch=epoch)
self.timer.stop()
self.elapsed_times[f'save_figs_{name}'].append(self.timer.elapsed_seconds())
# Horovod: average metric values across workers.
losses = {}
for level in range(3):
losses[f'log_loss_{name}_{self.loss_type}_Lv{level}'] = log_loss[level]
for key, value in losses.items():
loss = super()._metric_average(value, key)
self.loss_dict[key].append(loss)
def _validation(self, data_loader, epoch, name):
self.model0.eval()
self.model1.eval()
self.model2.eval()
log_loss = [0] * 3
nb_samples = len(data_loader.sampler)
for i, (sdf, flows) in enumerate(data_loader):
# Load data and meta-data
sdf_Lv0, sdf_Lv1, sdf_Lv2 = sdf
flows_Lv0, flows_Lv1, flows_Lv2 = flows
_, patch_y_Lv1, patch_x_Lv1, *_ = sdf_Lv1.shape
_, patch_y_Lv2, patch_x_Lv2, *_ = sdf_Lv2.shape
# Number of sub patches in each level
nb_patches_Lv2 = patch_y_Lv2 * patch_x_Lv2
nb_patches_Lv1 = patch_y_Lv1 * patch_x_Lv1
# Sub patch inside the Lv1 patch
patch_y_Lv2 = patch_y_Lv2 // patch_y_Lv1
patch_x_Lv2 = patch_x_Lv2 // patch_x_Lv1
batch_len = len(sdf_Lv0)
## To device
self.timer.start()
sdf_Lv0, sdf_Lv1, sdf_Lv2 = sdf_Lv0.to(self.device), sdf_Lv1.to(self.device), sdf_Lv2.to(self.device)
flows_Lv0, flows_Lv1, flows_Lv2 = flows_Lv0.to(self.device), flows_Lv1.to(self.device), flows_Lv2.to(self.device)
self.timer.stop()
self.elapsed_times[f'MemcpyH2D_{name}'].append(self.timer.elapsed_seconds())
# Keep sdfs on CPUs
sdf_Lv0_cpu = sdf_Lv0.to('cpu')
sdf_Lv1_cpu = sdf_Lv1.to('cpu')
sdf_Lv2_cpu = sdf_Lv2.to('cpu')
## Normalization or standardization
sdf_Lv0 = super()._preprocess(sdf_Lv0, self.sdf_Lv0_var0, self.sdf_Lv0_var1)
sdf_Lv1 = super()._preprocess(sdf_Lv1, self.sdf_Lv1_var0, self.sdf_Lv1_var1)
sdf_Lv2 = super()._preprocess(sdf_Lv2, self.sdf_Lv2_var0, self.sdf_Lv2_var1)
flows_Lv0 = super()._preprocess(flows_Lv0, self.flows_Lv0_var0, self.flows_Lv0_var1)
flows_Lv1 = super()._preprocess(flows_Lv1, self.flows_Lv1_var0, self.flows_Lv1_var1)
flows_Lv2 = super()._preprocess(flows_Lv2, self.flows_Lv2_var0, self.flows_Lv2_var1)
# Objectives: construct pred_flows_Lv0-Lv2
pred_flows_Lv0_ = torch.zeros_like(flows_Lv0, device='cpu')
pred_flows_Lv1_ = torch.zeros_like(flows_Lv1, device='cpu')
pred_flows_Lv2_ = torch.zeros_like(flows_Lv2, device='cpu')
#### Train Lv0
self.timer.start()
sdf_Lv0_ = sdf_Lv0[:, 0, 0]
flows_Lv0_ = flows_Lv0[:, 0, 0]
### Update weights
pred_flows_Lv0 = self.model0(sdf_Lv0_)
loss_mae = self.criterion(pred_flows_Lv0, flows_Lv0_)
### Log losses
level = 0
log_loss[level] += loss_mae.item() / nb_samples
### Destandardization and save
pred_flows_Lv0 = super()._postprocess(pred_flows_Lv0, self.flows_Lv0_var0, self.flows_Lv0_var1)
pred_flows_Lv0_[:, 0, 0, :, :, :] = pred_flows_Lv0.detach().cpu()
self.timer.stop()
self.elapsed_times[f'{name}_Lv{level}'].append(self.timer.elapsed_seconds())
### Train Lv1
for iy_Lv1, ix_Lv1 in itertools.product(range(patch_y_Lv1), range(patch_x_Lv1)):
self.timer.start()
sdf_Lv1_ = sdf_Lv1[:, iy_Lv1, ix_Lv1]
flows_Lv1_ = flows_Lv1[:, iy_Lv1, ix_Lv1]
### Update weights
pred_flows_Lv1 = self.model1(sdf_Lv1_)
loss_mae = self.criterion(pred_flows_Lv1, flows_Lv1_)
### Log losses
level = 1
log_loss[level] += loss_mae.item() / (nb_samples * nb_patches_Lv1)
pred_flows_Lv1 = super()._postprocess(pred_flows_Lv1, self.flows_Lv1_var0, self.flows_Lv1_var1)
pred_flows_Lv1_[:, iy_Lv1, ix_Lv1, :, :, :] = pred_flows_Lv1.detach().cpu()
self.timer.stop()
self.elapsed_times[f'{name}_Lv{level}'].append(self.timer.elapsed_seconds())
### Train Lv2
for iy_Lv2, ix_Lv2 in itertools.product(range(patch_y_Lv2), range(patch_x_Lv2)):
self.timer.start()
global_ix_Lv2 = ix_Lv2 + (ix_Lv1 * patch_x_Lv2)
global_iy_Lv2 = iy_Lv2 + (iy_Lv1 * patch_y_Lv2)
sdf_Lv2_ = sdf_Lv2[:, global_iy_Lv2, global_ix_Lv2]
flows_Lv2_ = flows_Lv2[:, global_iy_Lv2, global_ix_Lv2]
### Update generator weights
pred_flows_Lv2 = self.model2(sdf_Lv2_)
loss_mae = self.criterion(pred_flows_Lv2, flows_Lv2_)
### Log losses
level = 2
log_loss[level] += loss_mae.item() / (nb_samples * nb_patches_Lv2)
### Destandardization and save
pred_flows_Lv2 = super()._postprocess(pred_flows_Lv2, self.flows_Lv2_var0, self.flows_Lv2_var1)
pred_flows_Lv2_[:, global_iy_Lv2, global_ix_Lv2, :, :, :] = pred_flows_Lv2.detach().cpu()
self.timer.stop()
self.elapsed_times[f'{name}_Lv{level}'].append(self.timer.elapsed_seconds())
# Saving figures
if i == 0:
self.timer.start()
flows_Lv0 = super()._postprocess(flows_Lv0, self.flows_Lv0_var0, self.flows_Lv0_var1)
flows_Lv1 = super()._postprocess(flows_Lv1, self.flows_Lv1_var0, self.flows_Lv1_var1)
flows_Lv2 = super()._postprocess(flows_Lv2, self.flows_Lv2_var0, self.flows_Lv2_var1)
### Zeros inside objects
pred_flows_Lv0_ = super()._zeros_inside_objects(pred_flows_Lv0_, sdf_Lv0_cpu)
pred_flows_Lv1_ = super()._zeros_inside_objects(pred_flows_Lv1_, sdf_Lv1_cpu)
pred_flows_Lv2_ = super()._zeros_inside_objects(pred_flows_Lv2_, sdf_Lv2_cpu)
### Lv0 figures
level = 0
save_flows(flows_Lv0, name=name, img_dir = self.sub_img_dir, type_name = 'ref', level = level, epoch=epoch)
save_flows(pred_flows_Lv0_, name=name, img_dir = self.sub_img_dir, type_name = 'pred', level = level, epoch=epoch)
save_flows(pred_flows_Lv0_-flows_Lv0.cpu(), name=name, img_dir = self.sub_img_dir, type_name = 'error', level = level, epoch=epoch)
### Lv1 figures
level = 1
save_flows(flows_Lv1, name=name, img_dir = self.sub_img_dir, type_name = 'ref', level = level, epoch=epoch)
save_flows(pred_flows_Lv1_, name=name, img_dir = self.sub_img_dir, type_name = 'pred', level = level, epoch=epoch)
save_flows(pred_flows_Lv1_-flows_Lv1.cpu(), name=name, img_dir = self.sub_img_dir, type_name = 'error', level = level, epoch=epoch)
### Lv2 figures
level = 2
save_flows(flows_Lv2, name=name, img_dir = self.sub_img_dir, type_name = 'ref', level = level, epoch=epoch)
save_flows(pred_flows_Lv2_, name=name, img_dir = self.sub_img_dir, type_name = 'pred', level = level, epoch=epoch)
save_flows(pred_flows_Lv2_-flows_Lv2.cpu(), name=name, img_dir = self.sub_img_dir, type_name = 'error', level = level, epoch=epoch)
self.timer.stop()
self.elapsed_times[f'save_figs_{name}'].append(self.timer.elapsed_seconds())
# Horovod: average metric values across workers.
losses = {}
for level in range(3):
losses[f'log_loss_{name}_{self.loss_type}_Lv{level}'] = log_loss[level]
for key, value in losses.items():
loss = super()._metric_average(value, key)
self.loss_dict[key].append(loss)
### For inference
def _infer(self):
with torch.no_grad():
self._convert(data_loader=self.val_loader, name='validation')
self._convert(data_loader=self.test_loader, name='test')
def _convert(self, data_loader, name):
self.model0.eval()
self.model1.eval()
self.model2.eval()
nb_samples = len(data_loader.sampler)
for indices, sdf, flows in data_loader:
# Load data and meta-data
sdf_Lv0, sdf_Lv1, sdf_Lv2 = sdf
flows_Lv0, flows_Lv1, flows_Lv2 = flows
_, patch_y_Lv1, patch_x_Lv1, *_ = sdf_Lv1.shape
_, patch_y_Lv2, patch_x_Lv2, *_ = sdf_Lv2.shape
# Number of sub patches in each level
nb_patches_Lv2 = patch_y_Lv2 * patch_x_Lv2
nb_patches_Lv1 = patch_y_Lv1 * patch_x_Lv1
# Sub patch inside the Lv1 patch
patch_y_Lv2 = patch_y_Lv2 // patch_y_Lv1
patch_x_Lv2 = patch_x_Lv2 // patch_x_Lv1
batch_len = len(sdf_Lv0)
## To device
self.timer.start()
sdf_Lv0, sdf_Lv1, sdf_Lv2 = sdf_Lv0.to(self.device), sdf_Lv1.to(self.device), sdf_Lv2.to(self.device)
flows_Lv0, flows_Lv1, flows_Lv2 = flows_Lv0.to(self.device), flows_Lv1.to(self.device), flows_Lv2.to(self.device)
self.timer.stop()
self.elapsed_times[f'MemcpyH2D_{name}'].append(self.timer.elapsed_seconds())
# Keep sdfs on CPUs
sdf_Lv0_cpu = sdf_Lv0.to('cpu')
sdf_Lv1_cpu = sdf_Lv1.to('cpu')
sdf_Lv2_cpu = sdf_Lv2.to('cpu')
## Normalization or standardization
sdf_Lv0 = super()._preprocess(sdf_Lv0, self.sdf_Lv0_var0, self.sdf_Lv0_var1)
sdf_Lv1 = super()._preprocess(sdf_Lv1, self.sdf_Lv1_var0, self.sdf_Lv1_var1)
sdf_Lv2 = super()._preprocess(sdf_Lv2, self.sdf_Lv2_var0, self.sdf_Lv2_var1)
flows_Lv0 = super()._preprocess(flows_Lv0, self.flows_Lv0_var0, self.flows_Lv0_var1)
flows_Lv1 = super()._preprocess(flows_Lv1, self.flows_Lv1_var0, self.flows_Lv1_var1)
flows_Lv2 = super()._preprocess(flows_Lv2, self.flows_Lv2_var0, self.flows_Lv2_var1)
# Objectives: construct pred_flows_Lv0-Lv2
pred_flows_Lv0_ = torch.zeros_like(flows_Lv0, device='cpu')
pred_flows_Lv1_ = torch.zeros_like(flows_Lv1, device='cpu')
pred_flows_Lv2_ = torch.zeros_like(flows_Lv2, device='cpu')
#### Train Lv0
self.timer.start()
sdf_Lv0_ = sdf_Lv0[:, 0, 0]
flows_Lv0_ = flows_Lv0[:, 0, 0]
### Update weights
pred_flows_Lv0 = self.model0(sdf_Lv0_)
level = 0
### Destandardization and save
pred_flows_Lv0 = super()._postprocess(pred_flows_Lv0, self.flows_Lv0_var0, self.flows_Lv0_var1)
pred_flows_Lv0_[:, 0, 0, :, :, :] = pred_flows_Lv0.detach().cpu()
self.timer.stop()
self.elapsed_times[f'{name}_Lv{level}'].append(self.timer.elapsed_seconds())
### Train Lv1
for iy_Lv1, ix_Lv1 in itertools.product(range(patch_y_Lv1), range(patch_x_Lv1)):
self.timer.start()
sdf_Lv1_ = sdf_Lv1[:, iy_Lv1, ix_Lv1]
flows_Lv1_ = flows_Lv1[:, iy_Lv1, ix_Lv1]
### Update weights
pred_flows_Lv1 = self.model1(sdf_Lv1_)
### Log losses
level = 1
pred_flows_Lv1 = super()._postprocess(pred_flows_Lv1, self.flows_Lv1_var0, self.flows_Lv1_var1)
pred_flows_Lv1_[:, iy_Lv1, ix_Lv1, :, :, :] = pred_flows_Lv1.detach().cpu()
self.timer.stop()
self.elapsed_times[f'{name}_Lv{level}'].append(self.timer.elapsed_seconds())
### Train Lv2
for iy_Lv2, ix_Lv2 in itertools.product(range(patch_y_Lv2), range(patch_x_Lv2)):
self.timer.start()
global_ix_Lv2 = ix_Lv2 + (ix_Lv1 * patch_x_Lv2)
global_iy_Lv2 = iy_Lv2 + (iy_Lv1 * patch_y_Lv2)
sdf_Lv2_ = sdf_Lv2[:, global_iy_Lv2, global_ix_Lv2]
flows_Lv2_ = flows_Lv2[:, global_iy_Lv2, global_ix_Lv2]
### Update generator weights
pred_flows_Lv2 = self.model2(sdf_Lv2_)
level = 2
### Destandardization and save
pred_flows_Lv2 = super()._postprocess(pred_flows_Lv2, self.flows_Lv2_var0, self.flows_Lv2_var1)
pred_flows_Lv2_[:, global_iy_Lv2, global_ix_Lv2, :, :, :] = pred_flows_Lv2.detach().cpu()
self.timer.stop()
self.elapsed_times[f'{name}_Lv{level}'].append(self.timer.elapsed_seconds())
# Save the data in netcdf format
self.timer.start()
flows_Lv0 = super()._postprocess(flows_Lv0, self.flows_Lv0_var0, self.flows_Lv0_var1)
flows_Lv1 = super()._postprocess(flows_Lv1, self.flows_Lv1_var0, self.flows_Lv1_var1)
flows_Lv2 = super()._postprocess(flows_Lv2, self.flows_Lv2_var0, self.flows_Lv2_var1)
### Zeros inside objects
pred_flows_Lv0_ = super()._zeros_inside_objects(pred_flows_Lv0_, sdf_Lv0_cpu)
pred_flows_Lv1_ = super()._zeros_inside_objects(pred_flows_Lv1_, sdf_Lv1_cpu)
pred_flows_Lv2_ = super()._zeros_inside_objects(pred_flows_Lv2_, sdf_Lv2_cpu)
### Lv0 figures
level = 0
save_as_netcdf(sdf=sdf_Lv0_cpu, real_flows=flows_Lv0.cpu(), pred_flows=pred_flows_Lv0_,
indices=indices, epoch=self.epoch_start, level=level, name=name, data_dir=self.inference_dir)
### Lv1 figures
level = 1
save_as_netcdf(sdf=sdf_Lv1_cpu, real_flows=flows_Lv1.cpu(), pred_flows=pred_flows_Lv1_,
indices=indices, epoch=self.epoch_start, level=level, name=name, data_dir=self.inference_dir)
### Lv2 figures
level = 2
save_as_netcdf(sdf=sdf_Lv2_cpu, real_flows=flows_Lv2.cpu(), pred_flows=pred_flows_Lv2_,
indices=indices, epoch=self.epoch_start, level=level, name=name, data_dir=self.inference_dir)
self.timer.stop()
self.elapsed_times[f'save_figs_{name}'].append(self.timer.elapsed_seconds())