-
Notifications
You must be signed in to change notification settings - Fork 0
/
percent_lbl_train_aug.py
287 lines (223 loc) · 9.9 KB
/
percent_lbl_train_aug.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
import sys
import os
import torch
import time
import random
import imageio
import argparse
import numpy as np
import warnings
warnings.filterwarnings("ignore")
import torch.nn as nn
from torch import optim
from torch.utils.data import DataLoader
from torchvision.utils import make_grid
from tqdm import tqdm
from tensorboardX import SummaryWriter
from models.capsules_ucf101 import CapsNet
from utils.losses import SpreadLoss, DiceLoss
from utils.metrics import get_accuracy, IOU2
from utils.commons import init_seeds
def val_model_interface(minibatch):
data = minibatch['weak_data'].cuda()
action = minibatch['action'].cuda()
label_mask = minibatch['weak_mask'].cuda()
empty_vector = torch.zeros(action.shape[0]).cuda()
output, predicted_action, _ = model(data, action)
class_loss, abs_class_loss = criterion_cls(predicted_action, action)
loc_loss1 = criterion_loc_1(output, label_mask)
loc_loss2 = criterion_loc_2(output, label_mask)
loc_loss = loc_loss1 + loc_loss2
total_loss = loc_loss + class_loss
return output, predicted_action, label_mask, action, total_loss, loc_loss, class_loss
def train_model_interface(minibatch):
data = minibatch['strong_data'].cuda()
action = minibatch['action'].cuda()
label_mask = minibatch['strong_mask'].cuda()
empty_vector = torch.zeros(action.shape[0]).cuda()
output, predicted_action, _ = model(data, action)
# for viz_num in range(8):
# visualize_clips(data[viz_num], viz_num, 'sup_strong_clip')
# visualize_clips(label_mask[viz_num], viz_num, 'sup_strong_mask')
# visualize_pred_grad(output[viz_num], viz_num, 'sup_pred_strong_mask')
class_loss, abs_class_loss = criterion_cls(predicted_action, action)
loc_loss1 = criterion_loc_1(output, label_mask)
loc_loss2 = criterion_loc_2(output, label_mask)
loc_loss = loc_loss1 + loc_loss2
total_loss = loc_loss + class_loss
return output, predicted_action, label_mask, action, total_loss, loc_loss, class_loss
def train(args, model, train_loader, optimizer, epoch, save_path, writer):
start_time = time.time()
steps = len(train_loader)
model.train(mode=True)
model.training = True
total_loss = []
accuracy = []
loc_loss = []
class_loss = []
start_time = time.time()
for batch_id, minibatch in enumerate(train_loader):
optimizer.zero_grad()
output, predicted_action, _, action, loss, s_loss, c_loss = train_model_interface(minibatch)
loss.backward()
optimizer.step()
total_loss.append(loss.item())
loc_loss.append(s_loss.item())
class_loss.append(c_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(loc_loss).mean()
r_class = np.array(class_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}')
# summary writing
total_step = (epoch - 1) * len(train_loader) + batch_id + 1
info_loss = {
'loss': r_total,
'loss_seg': r_seg,
'loss_cls': r_class,
}
info_acc = {
'acc': r_acc
}
writer.add_scalars('train/loss', info_loss, total_step)
writer.add_scalars('train/acc', info_acc, total_step)
sys.stdout.flush()
del minibatch, output
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 = []
loc_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())
loc_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['weak_data'].shape[0]):
iou = IOU2(truth_np[a], maskout_np[a])
if iou == iou:
total_IOU += iou
validiou += 1
val_epoch_time = time.time() - start_time
print("Validation time: ", val_epoch_time)
r_total = np.array(total_loss).mean()
r_seg = np.array(loc_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='add_losses')
parser.add_argument('--bs', type=int, default=8, help='mini-batch size')
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('--pf', type=int, default=50, help='print frequency every batch')
parser.add_argument('--lr', type=float, default=0.0001, help='learning rate')
parser.add_argument('--loc_loss', type=str, default='dice', help='dice or iou loss')
parser.add_argument('--exp_id', type=str, default='debug_ucf101', help='experiment name')
# parser.add_argument('--pkl_file_label', type=str, default="train_annots_8_labeled_random.pkl",
# help='experiment name')
parser.add_argument('--pkl_file_label', type=str, default="training_annots_with_labels.pkl",
help='experiment name')
parser.add_argument('-at', '--aug_type', type=int, help="0-spatial, 1- temporal, 2 - both, 3-basic(only flip)")
# define seed params
parser.add_argument('--seed', type=int, default=47, help='seed for initializing training.')
parser.add_argument('--seed_data', type=int, default=47, help='seed variation pickle files')
args = parser.parse_args()
return args
if __name__ == '__main__':
args = parse_args()
print(vars(args))
init_seeds(args.seed)
USE_CUDA = True if torch.cuda.is_available() else False
TRAIN_BATCH_SIZE = args.bs
VAL_BATCH_SIZE = args.bs
N_EPOCHS = args.epochs
LR = args.lr
loc_loss_criteria = args.loc_loss
from datasets.sup_ucf_dataloader_st_augs_v1_speedup import UCF101DataLoader, collate_fn_train, collate_fn_test
labeled_trainset = UCF101DataLoader('train', [224, 224], cl=8, file_id=args.pkl_file_label,
aug_mode=args.aug_type, subset_seed=args.seed_data)
validationset = UCF101DataLoader('validation', [224, 224], cl=8, file_id="test_annots.pkl",
aug_mode=0, subset_seed=args.seed_data)
print(len(labeled_trainset), len(validationset))
labeled_train_data_loader = DataLoader(
dataset=labeled_trainset,
batch_size=TRAIN_BATCH_SIZE,
num_workers=8,
shuffle=True,
collate_fn=collate_fn_train
)
val_data_loader = DataLoader(
dataset=validationset,
batch_size=VAL_BATCH_SIZE,
num_workers=8,
shuffle=False,
collate_fn=collate_fn_test
)
print(len(labeled_train_data_loader), len(val_data_loader))
# Load pretrained weights
model = CapsNet(pretrained_load=True)
if USE_CUDA:
model = model.cuda()
# losses
global criterion_cls
global criterion_loc_1
global criterion_loc_2
criterion_cls = SpreadLoss(num_class=24, m_min=0.2, m_max=0.9)
criterion_loc_1 = nn.BCEWithLogitsLoss(size_average=True)
criterion_loc_2 = DiceLoss()
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)
exp_id = args.exp_id
save_path = os.path.join('sup_train_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)
prev_best_val_loss = 10000
prev_best_train_loss = 10000
prev_best_val_loss_model_path = None
prev_best_train_loss_model_path = None
for e in tqdm(range(1, N_EPOCHS + 1), total=N_EPOCHS, desc="Epochs"):
train_loss = train(args, model, labeled_train_data_loader, optimizer, e, save_path, writer)
val_loss = validate(model, val_data_loader, e)
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:
os.remove(prev_best_train_loss_model_path)
prev_best_train_loss_model_path = train_model_path
scheduler.step(train_loss)