-
Notifications
You must be signed in to change notification settings - Fork 16
/
train.py
249 lines (216 loc) · 10.2 KB
/
train.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
"""
Author: Amr Elsersy
email: [email protected]
-----------------------------------------------------------------------------------
Description: Training & Validation
"""
import numpy as np
import argparse
import logging
import time
import os
from tqdm import tqdm
import torch
import torch.nn as nn
import torch.optim
import torch.utils.tensorboard as tensorboard
from dataset import WFLW_Dataset
from dataset import create_test_loader, create_train_loader
from visualization import WFLW_Visualizer
from model.Loss import PFLD_L2Loss
from model.model import PFLD, AuxiliaryNet
from model.DepthSepConv import DepthSepConvBlock
from model.BottleneckResidual import BottleneckResidualBlock
from utils import to_numpy_image
import torch.backends.cudnn as cudnn
cudnn.benchmark = True
cudnn.determinstic = True
cudnn.enabled = True
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('--epochs', type=int, default=800, help='num of training epochs')
parser.add_argument('--batch_size', type=int, default=24, help="training batch size")
parser.add_argument('--tensorboard', type=str, default='checkpoint/tensorboard', help='path log dir of tensorboard')
parser.add_argument('--logging', type=str, default='checkpoint/logging', help='path of logging')
parser.add_argument('--lr', type=float, default=0.00007, help='learning rate')
parser.add_argument('--weight_decay', type=float, default=1e-6, help='optimizer weight decay')
parser.add_argument('--datapath', type=str, default='data', help='root path of augumented WFLW dataset')
parser.add_argument('--pretrained', type=str,default='checkpoint/model_weights/weights.pth1.tar',help='load checkpoint')
parser.add_argument('--resume', action='store_true', help='resume from pretrained path specified in prev arg')
parser.add_argument('--savepath', type=str, default='checkpoint/model_weights', help='save checkpoint path')
parser.add_argument('--savefreq', type=int, default=1, help="save weights each freq num of epochs")
parser.add_argument('--logdir', type=str, default='checkpoint/logging', help='logging')
parser.add_argument("--lr_patience", default=40, type=int)
args = parser.parse_args()
return args
# ======================================================================
# device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# args
args = parse_args()
# logging
logging.basicConfig(
format='[%(asctime)s] [p%(process)s] [%(pathname)s:%(lineno)d] [%(levelname)s] %(message)s',
level=logging.INFO,
handlers=[logging.FileHandler(args.logdir, mode='w'), logging.StreamHandler()])
# tensorboard
writer = tensorboard.SummaryWriter(args.tensorboard)
def main():
# ========= dataloaders ===========
train_dataloader = create_train_loader(root=args.datapath,batch_size=args.batch_size)
test_dataloader = create_test_loader(root=args.datapath, batch_size=args.batch_size)
start_epoch = 0
# ======== models & loss ==========
pfld = PFLD().to(device)
auxiliarynet = AuxiliaryNet().to(device)
loss = PFLD_L2Loss().to(device)
# ========= load weights ===========
if args.resume:
checkpoint = torch.load(args.pretrained)
pfld.load_state_dict(checkpoint["pfld"], strict=False)
auxiliarynet.load_state_dict(checkpoint["auxiliary"])
start_epoch = checkpoint['epoch'] + 1
print(f'\tLoaded checkpoint from {args.pretrained}\n')
# logging.info(f'\tLoaded checkpoint from {args.pretrained}\n')
time.sleep(1)
else:
print("******************* Start training from scratch *******************\n")
time.sleep(5)
# =========== optimizer ===========
# parameters = list(pfld.parameters()) + list(auxiliarynet.parameters())
parameters = [
{ 'params': pfld.parameters() },
{ 'params': auxiliarynet.parameters() }
]
optimizer = torch.optim.Adam(parameters, lr=args.lr, weight_decay=args.weight_decay)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', patience=args.lr_patience, verbose=True)
# ========================================================================
for epoch in range(start_epoch, args.epochs):
# =========== train / validate ===========
w_train_loss, train_loss = train_one_epoch(pfld, auxiliarynet, loss, optimizer, train_dataloader, epoch)
val_loss = validate(pfld, auxiliarynet, loss, test_dataloader, epoch)
scheduler.step(val_loss)
logging.info(f"\ttraining epoch={epoch} .. weighted_loss= {w_train_loss} ... loss={train_loss}")
# ============= tensorboard =============
# writer.add_scalar('train_weighted_loss',w_train_loss, epoch)
writer.add_scalar('train_loss',train_loss, epoch)
writer.add_scalar('val_loss',val_loss, epoch)
# ============== save model =============
if epoch % args.savefreq == 0:
checkpoint_state = {
"pfld": pfld.state_dict(),
"auxiliary": auxiliarynet.state_dict(),
"epoch": epoch
}
savepath = os.path.join(args.savepath, f'weights.pth_epoch_{epoch}.tar')
torch.save(checkpoint_state, savepath)
print(f'\n\t*** Saved checkpoint in {savepath} ***\n')
time.sleep(2)
writer.close()
def train_one_epoch(pfld_model, auxiliary_model, criterion, optimizer, dataloader, epoch):
weighted_loss = 0
loss = 0
pfld_model.train()
auxiliary_model.train()
for image, labels in tqdm(dataloader):
euler_angles = labels['euler_angles'].squeeze() # shape (batch, 3)
attributes = labels['attributes'].squeeze() # shape (batch, 6)
landmarks = labels['landmarks'].squeeze() # shape (batch, 98, 2)
landmarks = landmarks.reshape((landmarks.shape[0], 196)) # reshape landmarks to match loss function
image = image.to(device)
landmarks = landmarks.to(device)
euler_angles = euler_angles.to(device)
attributes = attributes.to(device)
pfld_model = pfld_model.to(device)
auxiliary_model = auxiliary_model.to(device)
featrues, pred_landmarks = pfld_model(image)
pred_angles = auxiliary_model(featrues)
weighted_loss, loss = criterion(pred_landmarks, landmarks, pred_angles, euler_angles, attributes)
train_w_loss = round(weighted_loss.item(),3)
train_loss = round(loss.item(),3)
print(f"training epoch={epoch} .. weighted_loss= {train_w_loss} ... loss={train_loss}\n")
optimizer.zero_grad()
weighted_loss.backward()
optimizer.step()
return weighted_loss.item(), loss.item()
def validate(pfld_model, auxiliary_model, criterion, dataloader, epoch):
validation_losses = []
pfld_model.eval()
auxiliary_model.eval()
with torch.no_grad():
for image, labels in tqdm(dataloader):
euler_angles = labels['euler_angles'].squeeze() # shape (batch, 3)
attributes = labels['attributes'].squeeze() # shape (batch, 6)
landmarks = labels['landmarks'].squeeze() # shape (batch, 98, 2)
landmarks = landmarks.reshape((landmarks.shape[0], 196)) # reshape landmarks to match loss function
image = image.to(device)
landmarks = landmarks.to(device)
euler_angles = euler_angles.to(device)
attributes = attributes.to(device)
pfld_model = pfld_model.to(device)
auxiliary_model = auxiliary_model.to(device)
featrues, pred_landmarks = pfld_model(image)
pred_angles = auxiliary_model(featrues)
weighted_loss, loss = criterion(pred_landmarks, landmarks, pred_angles, euler_angles, attributes)
weighted_loss = round(weighted_loss.item(),3)
loss = round(loss.item(),3)
print(f"\tval epoch={epoch} .. val_weighted_loss= {weighted_loss} ... val_loss={loss}\n")
# logging.info(f"\tval epoch={epoch} .. val_weighted_loss= {weighted_loss} ... val_loss={loss}\n")
validation_losses.append(loss)
avg_val_loss = round(np.mean(validation_losses).item(),3)
print('*'*70,f'\n\tEvaluation average loss= {avg_val_loss}\n')
logging.info('*'*70 + f'\n\tEvaluation average loss= {avg_val_loss}\n')
time.sleep(1)
return avg_val_loss
if __name__ == "__main__":
main()
# import torchvision.transforms.transforms as transforms
# transform = transforms.Compose([transforms.ToTensor()])
# dataset = WFLW_Dataset(mode='train', transform=True)
# # ============ From tensor to image ... crahses in any function in cv2 ==================
# dataloader = create_train_loader(transform=True)
# for images, labels in dataloader:
# print(images.shape)
# image = images[0]
# print(image.shape)
# image = to_numpy_image(images[0])
# print(image.shape)
# import cv2
# landmarks = labels['landmarks'].squeeze()[0]
# euler_angles = labels['euler_angles'].squeeze()[0]
# attributes = labels['attributes'].squeeze()[0]
# l = {}
# l['landmarks'] = landmarks.numpy()
# l['euler_angles'] = euler_angles.numpy()
# l['attributes'] = attributes.numpy()
# visualizer = WFLW_Visualizer()
# visualizer.visualize(image, l)
# print(image.shape, image)
# ====================================================
# ======= Habd
# cv2.circle(image, (40,50), 30, (245,0,0), -1)
# cv2.imshow("I", image)
# cv2.waitKey(0)
# datase2 = WFLW_Dataset(transform=False, mode='val')
# image2, labels2 = datase2[0]
# print(image.shape, image2.shape, type(image), type(image2))
# ============= Test reshape and back reshape (works well) ======
# x = np.array([[
# [1,2],
# [3,4],
# [5,6]
# ]])
# print(x.shape)
# xx = transform(x)
# print("transform",xx,'\n')
# xx = xx.reshape((1,1,6))
# print("flatten",xx,'\n')
# xx = xx.reshape((1,3,2))
# print("reshape",xx,'\n')
# ======== Test Landmarks reshape =========
# x = torch.tensor([
# 1,2,3,4,5,6
# ])
# print(x.shape)
# x = x.reshape((1,3,2))
# print("reshape",x,'\n')