-
Notifications
You must be signed in to change notification settings - Fork 4
/
main_jhmdb.py
447 lines (338 loc) · 17 KB
/
main_jhmdb.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
import sys
import os
import torch
import time
import random
import imageio
import argparse
import datetime
import numpy as np
import warnings
warnings.filterwarnings("ignore")
import torch.nn as nn
from torch import optim
from torchvision import datasets, transforms
from torch.utils.data import DataLoader
from tqdm import tqdm
from tensorboardX import SummaryWriter
from utils.losses import SpreadLoss, DiceLoss, weighted_mse_loss
from utils.metrics import get_accuracy, IOU2
from utils.helpers import measure_pixelwise_gradient, measure_pixelwise_var_v2
from utils import ramp_ups
import wandb
# WANDB INIT
def val_model_interface(minibatch):
data = minibatch['data'].type(torch.cuda.FloatTensor)
action = minibatch['action'].cuda()
segmentation = minibatch['loc_msk']
empty_vector = torch.zeros(action.shape[0]).cuda()
output, predicted_action, _ = model(data, action, empty_vector, 0, 0)
class_loss, abs_class_loss = criterion_cls(predicted_action, action)
loss1 = criterion_seg_1(output, segmentation.float().cuda())
loss2 = criterion_seg_2(output, segmentation.float().cuda())
seg_loss = loss1 + loss2
total_loss = seg_loss + class_loss
return (output, predicted_action, segmentation, action, total_loss, seg_loss, class_loss)
def train_model_interface(args, label_minibatch, unlabel_minibatch, epoch, wt_ramp):
label_data = label_minibatch['data'].type(torch.cuda.FloatTensor)
fl_label_data = label_minibatch['aug_data'].type(torch.cuda.FloatTensor)
unlabel_data = unlabel_minibatch['data'].type(torch.cuda.FloatTensor)
fl_unlabel_data = unlabel_minibatch['aug_data'].type(torch.cuda.FloatTensor)
label_action = label_minibatch['action'].cuda()
unlabel_action = unlabel_minibatch['action'].cuda()
label_segmentation = label_minibatch['loc_msk']
unlabel_segmentation = unlabel_minibatch['loc_msk']
concat_data = torch.cat([label_data, unlabel_data], dim=0)
concat_fl_data = torch.cat([fl_label_data, fl_unlabel_data], dim=0)
concat_action = torch.cat([label_action, unlabel_action], dim=0)
concat_seg = torch.cat([label_segmentation, unlabel_segmentation], dim=0)
ones_tensor = torch.ones(len(label_action))
zeros_tensor = torch.zeros(len(unlabel_action))
concat_labels = torch.cat([ones_tensor, zeros_tensor], dim=0).cuda()
random_indices = torch.randperm(len(concat_labels))
concat_data = concat_data[random_indices, :, :, :, :]
concat_fl_data = concat_fl_data[random_indices, :, :, :,:]
concat_action = concat_action[random_indices]
concat_labels = concat_labels[random_indices]
concat_seg = concat_seg[random_indices, :, :, :, :]
labeled_vid_index = torch.where(concat_labels==1)[0]
output, predicted_action, feat = model(concat_data, concat_action, concat_labels, epoch, args.thresh_epoch)
flip_op, _, _ = model(concat_fl_data, concat_action, concat_labels, epoch, args.thresh_epoch)
# SEG LOSS SUPERVISED
labeled_op = output[labeled_vid_index]
labeled_seg_data = concat_seg[labeled_vid_index]
seg_loss_1 = criterion_seg_1(labeled_op, labeled_seg_data.float().cuda())
seg_loss_2 = criterion_seg_2(labeled_op, labeled_seg_data.float().cuda())
# Classification loss SUPERVISED
labeled_cls = concat_action[labeled_vid_index]
labeled_pred_action = predicted_action[labeled_vid_index]
class_loss, abs_class_loss = criterion_cls(labeled_pred_action, labeled_cls)
# CONST LOSS
flipped_pred_seg_map = torch.flip(flip_op, [4])
####################################
# Equal weighted MSE Loss #
####################################
# CHECKED - THIS OUTPUTS SAME AS - nn.MSELoss()
equal_wt = torch.ones_like(output, dtype=torch.double)
equal_wt = equal_wt.type(torch.cuda.FloatTensor)
loss_wt_simple_l2 = weighted_mse_loss(flipped_pred_seg_map, output, equal_wt)
####################################
# Weighted MSE Loss - simple var #
####################################
if args.bv:
# CLCK+ANTICLCK
batch_variance_clck = measure_pixelwise_var_v2(output, torch.flip(flipped_pred_seg_map, [2]), frames_cnt=args.n_frames, use_sig_output=args.predict_maps)
batch_variance_anticlck = measure_pixelwise_var_v2(torch.flip(output, [2]), flipped_pred_seg_map, frames_cnt=args.n_frames, use_sig_output=args.predict_maps)
batch_variance_clck = batch_variance_clck.type(torch.cuda.FloatTensor)
batch_variance_anticlck = batch_variance_anticlck.type(torch.cuda.FloatTensor)
loss_wt_var_1 = weighted_mse_loss(flipped_pred_seg_map, output, batch_variance_clck)
loss_wt_var_2 = weighted_mse_loss(flipped_pred_seg_map, output, torch.flip(batch_variance_anticlck, [2]))
total_seg_cons_loss = (wt_ramp * (loss_wt_var_1 + loss_wt_var_2)) + ((1 - wt_ramp) * loss_wt_simple_l2)
####################################
# Weighted MSE Loss - gradients #
####################################
if args.gv:
batch_grad = measure_pixelwise_gradient(output, conf_thresh_lower=args.lower_thresh, conf_thresh_upper=args.upper_thresh)
batch_grad = batch_grad.type(torch.cuda.FloatTensor)
loss_wt_grad = weighted_mse_loss(flipped_pred_seg_map, output, batch_grad)
total_seg_cons_loss = loss_wt_grad
# total_seg_cons_loss = (wt_ramp * loss_wt_grad) + ((1 - wt_ramp) * loss_wt_simple_l2)
total_cons_loss = total_seg_cons_loss
seg_loss = seg_loss_1 + seg_loss_2
total_loss = args.wt_seg * seg_loss + args.wt_cls * class_loss + args.wt_cons * total_cons_loss
return (output, predicted_action, concat_seg, concat_action, total_loss, seg_loss, class_loss, total_cons_loss)
def train(args, model, labeled_train_loader, unlabeled_train_loader, optimizer, epoch, save_path, writer, ramp_wt):
model.train(mode=True)
model.training = True
total_loss = []
accuracy = []
seg_loss = []
class_loss = []
class_consistency_loss = []
steps = len(unlabeled_train_loader)
start_time = time.time()
labeled_iterloader = iter(labeled_train_loader)
for batch_id, unlabel_minibatch in enumerate(unlabeled_train_loader):
optimizer.zero_grad()
try:
label_minibatch = next(labeled_iterloader)
except StopIteration:
labeled_iterloader = iter(labeled_train_loader)
label_minibatch = next(labeled_iterloader)
output, predicted_action, segmentation, action, loss, s_loss, c_loss, cc_loss =\
train_model_interface(args, label_minibatch, unlabel_minibatch, epoch, ramp_wt(epoch))
loss.backward()
optimizer.step()
total_loss.append(loss.item())
seg_loss.append(s_loss.item())
class_loss.append(c_loss.item())
class_consistency_loss.append(cc_loss.item())
accuracy.append(get_accuracy(predicted_action, action))
if (batch_id + 1) % args.pf == 0:
r_total = np.array(total_loss).mean()
r_seg = np.array(seg_loss).mean()
r_class = np.array(class_loss).mean()
r_const = np.array(class_consistency_loss).mean()
r_acc = np.array(accuracy).mean()
print(f'[TRAIN] epoch-{epoch:0{len(str(args.epochs))}}/{args.epochs}, batch-{batch_id+1:0{len(str(steps))}}/{steps},' \
f'loss-{r_total:.3f}, acc-{r_acc:.3f}' \
f'\t [LOSS ] cls-{r_class:.3f}, seg-{r_seg:.3f}, const-{r_const:.3f}')
# summary writing
total_step = (epoch-1)*len(unlabeled_train_loader) + batch_id + 1
info_loss = {
'loss': r_total,
'loss_seg': r_seg,
'loss_cls': r_class,
'loss_consistency':r_const
}
info_acc = {
'acc': r_acc
}
wandb.log({
"loss_total": r_total,
"loss_cls": r_class,
"loss_seg": r_seg,
"loss_loc_const": r_const,
"acc": r_acc
})
writer.add_scalars('train/loss', info_loss, total_step)
writer.add_scalars('train/acc', info_acc, total_step)
sys.stdout.flush()
end_time = time.time()
train_epoch_time = end_time - start_time
print("Training time: ", train_epoch_time)
train_total_loss = np.array(total_loss).mean()
return train_total_loss
def validate(model, val_data_loader, epoch):
steps = len(val_data_loader)
model.eval()
model.training = False
total_loss = []
accuracy = []
seg_loss = []
class_loss = []
total_IOU = 0
validiou = 0
print('validating...')
start_time = time.time()
with torch.no_grad():
for batch_id, minibatch in enumerate(val_data_loader):
output, predicted_action, segmentation, action, loss, s_loss, c_loss = val_model_interface(minibatch)
total_loss.append(loss.item())
seg_loss.append(s_loss.item())
class_loss.append(c_loss.item())
accuracy.append(get_accuracy(predicted_action, action))
maskout = output.cpu()
maskout_np = maskout.data.numpy()
# utils.show(maskout_np[0])
# use threshold to make mask binary
maskout_np[maskout_np > 0] = 1
maskout_np[maskout_np < 1] = 0
# utils.show(maskout_np[0])
truth_np = segmentation.cpu().data.numpy()
for a in range(minibatch['data'].shape[0]):
iou = IOU2(truth_np[a], maskout_np[a])
if iou == iou:
total_IOU += iou
validiou += 1
else:
print('bad IOU')
val_epoch_time = time.time() - start_time
print("Validation time: ", val_epoch_time)
r_total = np.array(total_loss).mean()
r_seg = np.array(seg_loss).mean()
r_class = np.array(class_loss).mean()
r_acc = np.array(accuracy).mean()
average_IOU = total_IOU / validiou
print(f'[VAL] epoch-{epoch}, loss-{r_total:.3f}, acc-{r_acc:.3f} [IOU ] {average_IOU:.3f}')
sys.stdout.flush()
return r_total
def parse_args():
parser = argparse.ArgumentParser(description='loc var const')
parser.add_argument('--bs', type=int, default=16, help='mini-batch size')
parser.add_argument('--pf', type=int, default=50, help='print frequency every batch')
parser.add_argument('--epochs', type=int, default=1, help='number of total epochs to run')
parser.add_argument('--model_name', type=str, default='i3d', help='model name')
parser.add_argument('--lr', type=float, default=0.001, help='learning rate')
parser.add_argument('--seg_loss', type=str, default='dice', help='dice or iou loss')
parser.add_argument('--exp_id', type=str, default='debug', help='experiment name')
parser.add_argument('--pkl_file_label', type=str, default='jhmdb_classes_list_per_20_labeled.txt', help='label subset')
parser.add_argument('--pkl_file_unlabel', type=str, default='jhmdb_classes_list_per_80_unlabeled.txt', help='unlabele subset')
parser.add_argument('--const_loss', type=str, default="l2", help='consistency loss type')
parser.add_argument('--wt_seg', type=float, default=1, help='segmentation loss weight')
parser.add_argument('--wt_cls', type=float, default=1, help='Classification loss weight')
parser.add_argument('--wt_cons', type=float, default=1, help='class consistency loss weight')
parser.add_argument('--seed', type=int, default=47, help='seed for initializing training.')
parser.add_argument('--thresh_epoch', type=int, default=11, help='thresh epoch to introduce pseudo labels')
parser.add_argument('--n_frames', type=int, default=3, help='batch variance frames number.')
parser.add_argument('--bv', action='store_true', help='use batch variance')
parser.add_argument('--predict_maps', action='store_true', help='use sigmoid outputs')
parser.add_argument('--cyclic', action='store_true', help='use batch variance')
parser.add_argument('--gv', action='store_true', help='use grad variance')
parser.add_argument('--lower_thresh', type=float, default=None, help='lower conf thresh')
parser.add_argument('--upper_thresh', type=float, default=None, help='upper conf thresh')
parser.add_argument('--viz', action='store_true', help='map visuzlization debug')
parser.add_argument('--seed_num', type=int, default=47, help='seed variation pickle files')
args = parser.parse_args()
return args
if __name__ == '__main__':
args = parse_args()
print(vars(args))
torch.backends.cudnn.benchmark = True
torch.backends.cudnn.enabled = True
if args.seed:
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
USE_CUDA = True if torch.cuda.is_available() else False
if torch.cuda.is_available() and not USE_CUDA:
print("WARNING: You have a CUDA device, so you should probably run with --cuda")
TRAIN_BATCH_SIZE = args.bs
VAL_BATCH_SIZE = args.bs
N_EPOCHS = args.epochs
LR = args.lr
seg_loss_criteria = args.seg_loss
from datasets.load_jhmdb_pytorch_multi import JHMDB
labeled_trainset = JHMDB('train', [224, 224], file_id=args.pkl_file_label, use_random_start_frame=False)
unlabeled_trainset = JHMDB('train', [224, 224], file_id=args.pkl_file_unlabel, use_random_start_frame=False)
validationset = JHMDB('test',[224, 224], file_id='testlist.txt', use_random_start_frame=False)
print(len(labeled_trainset), len(unlabeled_trainset), len(validationset))
labeled_train_data_loader = DataLoader(
dataset=labeled_trainset,
batch_size=(TRAIN_BATCH_SIZE)//2,
num_workers=8,
shuffle=True
)
unlabeled_train_data_loader = DataLoader(
dataset=unlabeled_trainset,
batch_size=(TRAIN_BATCH_SIZE)//2,
num_workers=8,
shuffle=True
)
val_data_loader = DataLoader(
dataset=validationset,
batch_size=VAL_BATCH_SIZE,
num_workers=8,
shuffle=False
)
print(len(labeled_train_data_loader), len(unlabeled_train_data_loader), len(val_data_loader))
from models.capsules_jhmdb_semi_sup_pa import CapsNet
# Load pretrained weights
model = CapsNet(pretrained_load=True)
if USE_CUDA:
model = model.cuda()
# losses
global criterion_cls
global criterion_seg_1
global criterion_seg_2
global consistency_criterion
criterion_cls = SpreadLoss(num_class=21, m_min=0.2, m_max=0.9)
criterion_seg_1 = nn.BCEWithLogitsLoss(size_average=True)
criterion_seg_2 = DiceLoss()
if args.const_loss == 'jsd':
consistency_criterion = torch.nn.KLDivLoss(size_average=False, reduce=False).cuda()
elif args.const_loss == 'l2':
consistency_criterion = nn.MSELoss()
elif args.const_loss == 'l1':
consistency_criterion = nn.L1Loss()
else:
print("no consistency criterion found. Exiting the code!!!")
exit()
print("Consistency criterion: ", consistency_criterion)
optimizer = optim.Adam(model.parameters(), lr=LR, weight_decay=0, eps=1e-6)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', min_lr=1e-7, patience=5, factor=0.1, verbose=True)
ramp_wt = ramp_ups.exp_rampup(N_EPOCHS)
exp_id = args.exp_id
save_path = os.path.join('./train_log_wts', exp_id)
model_save_dir = os.path.join(save_path,time.strftime('%m-%d-%H-%M'))
writer = SummaryWriter(model_save_dir)
if not os.path.exists(model_save_dir):
os.makedirs(model_save_dir)
# WANDB RUN NAME DECLARATION
wandb.run.name = args.exp_id
prev_best_val_loss = 10000
prev_best_train_loss = 10000
prev_best_val_loss_model_path = None
prev_best_train_loss_model_path = None
wandb.watch(model)
for e in tqdm(range(1, N_EPOCHS + 1)):
train_loss = train(args, model, labeled_train_data_loader, unlabeled_train_data_loader, optimizer, e, save_path, writer, ramp_wt)
val_loss = validate(model, val_data_loader, e)
if val_loss < prev_best_val_loss:
print("Yay!!! Got the val loss down...")
val_model_path = os.path.join(model_save_dir, f'best_model_val_loss_{e}.pth')
torch.save(model.state_dict(), val_model_path)
prev_best_val_loss = val_loss;
if prev_best_val_loss_model_path and e< 4:
os.remove(prev_best_val_loss_model_path)
prev_best_val_loss_model_path = val_model_path
if train_loss < prev_best_train_loss:
print("Yay!!! Got the train loss down...")
train_model_path = os.path.join(model_save_dir, f'best_model_train_loss_{e}.pth')
torch.save(model.state_dict(), train_model_path)
prev_best_train_loss = train_loss
if prev_best_train_loss_model_path and e<4:
os.remove(prev_best_train_loss_model_path)
prev_best_train_loss_model_path = train_model_path
scheduler.step(train_loss)