-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain.py
executable file
·221 lines (211 loc) · 9.68 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
import os
import cv2
import torch
import torch.utils.data
import torch.utils.tensorboard
import PIL.Image
import numpy as np
from datasets import get_train_val_loaders
from deep_3d_face import (
get_encoder,
Decoder,
Renderer,
PerceptualLoss,
LandmarkLoss,
PhotometricLoss,
CoefficientRegularizationLoss
)
# data
bfm_params_dir = "params/bfm"
data_dir = "data/data"
batch_size = 32
num_workers = 2
# optimizer
learning_rate = 1e-4
weight_decay = 0
# training & info
num_epochs = 500
print_iter = 20
val_iter = 2000
summary_dir = "ckpts/summary"
# resume training related
decoder_trainable = False
encoder_fixed = False
decoder_fixed = False
resume_encoder_params = ""
resume_decoder_params = ""
vis_dir = "ckpts/images"
save_encoder_path = "ckpts/encoder_%03d.pth"
save_decoder_path = "ckpts/decoder_%03d.pth"
def save_images(original, rendered, gt_landmarks, pred_landmarks, epoch, iteration, save_dir):
"""
:param original: (bs, 3, h, w) (-1, 1) RGB GPU
:param rendered: (bs, 3, h, w) (-1, 1) RGB GPU
:param gt_landmarks: (bs, 68, 2) GPU
:param pred_landmarks: (bs, 68, 2) GPU
:param epoch:
:param iteration:
:param save_dir: directory to save images
:return: image: (numpy.ndarray, numpy.uint8, 0-255) an integrated image
"""
bs, c, h, w = original.shape
ori = original * 127.5 + 127.5
ren = rendered * 127.5 + 127.5
ori = ori.cpu().detach().numpy().transpose(0, 2, 3, 1)
ren = ren.cpu().detach().numpy().transpose(0, 2, 3, 1)
gt_ldmk = gt_landmarks.cpu().detach().numpy().reshape((-1, 68, 2))
pd_ldmk = pred_landmarks.cpu().detach().numpy().reshape((-1, 68, 2))
length = bs if bs < 8 else 8
image = np.zeros([h * 4, w * length, c], dtype=np.uint8)
for i in range(length):
ori_cp = np.ascontiguousarray(ori[i].copy(), dtype=np.uint8)
ren_cp = np.ascontiguousarray(ren[i].copy(), dtype=np.uint8)
image[0:h, i*w:(i+1)*w, :] = ori_cp
image[h:2*h, i*w:(i+1)*w, :] = ren_cp
for j in range(68):
ori_cp = cv2.circle(ori_cp, (gt_ldmk[i][j][0], gt_ldmk[i][j][1]),
radius=1, color=(255, 255, 255), thickness=1)
ren_cp = cv2.circle(ren_cp, (pd_ldmk[i][j][0], pd_ldmk[i][j][1]),
radius=1, color=(255, 255, 255), thickness=1)
image[2 * h:3 * h, i * w:(i + 1) * w, :] = ori_cp
image[3 * h:4 * h, i * w:(i + 1) * w, :] = ren_cp
image_pil = PIL.Image.fromarray(image)
if type(epoch) is int:
image_pil.save(os.path.join(save_dir, "%d_%d.png" % (epoch, iteration)))
else:
image_pil.save(os.path.join(save_dir, "val_%d.png" % iteration))
return image
def main():
if not os.path.exists(vis_dir):
os.makedirs(vis_dir)
# dataset
train_loader, val_loader = get_train_val_loaders(data_dir, batch_size, num_workers)
train_length = len(train_loader)
global_val_iteration = 0
# tensorboard
writer = torch.utils.tensorboard.SummaryWriter(summary_dir, flush_secs=30)
# networks
encoder = get_encoder()
decoder = Decoder(bfm_params_dir=bfm_params_dir, trainable=decoder_trainable)
if resume_encoder_params:
encoder.load_state_dict(torch.load(resume_encoder_params))
if decoder_trainable and resume_decoder_params:
decoder.load_state_dict(torch.load(resume_decoder_params))
renderer = Renderer()
# move to cuda
encoder = encoder.cuda()
decoder = decoder.cuda()
renderer = renderer.cuda()
# losses
photometric_loss = PhotometricLoss()
landmark_loss = LandmarkLoss()
perceptual_loss = PerceptualLoss()
coefficient_regularization_loss = CoefficientRegularizationLoss()
photometric_loss = photometric_loss.cuda()
landmark_loss = landmark_loss.cuda()
perceptual_loss = perceptual_loss.cuda()
coefficient_regularization_loss = coefficient_regularization_loss.cuda()
# optimizer
parameters = list()
if not encoder_fixed:
parameters.extend(encoder.parameters())
if decoder_trainable and (not decoder_fixed):
parameters.extend(decoder.parameters())
optimizer = torch.optim.Adam(params=parameters,
lr=learning_rate,
weight_decay=weight_decay)
# train & val
for epoch in range(num_epochs):
for iteration, (images, landmarks, masks) in enumerate(train_loader):
optimizer.zero_grad()
encoder.train()
decoder.train()
renderer.train()
# data
images = images.cuda()
landmarks = landmarks.cuda()
masks = masks.cuda()
# forward
coeff = encoder(images)
face_projection, face_color, landmarks_3d, tri = decoder(coeff)
rendered_images, _ = renderer(face_projection, face_color, tri)
# loss
loss_photometric = photometric_loss(images, rendered_images, masks)
loss_landmark, landmarks_2d = landmark_loss(landmarks_3d, landmarks)
loss_perceptual = perceptual_loss(images, rendered_images, masks)
loss_coeff_reg = coefficient_regularization_loss(coeff)
total_loss = 1.9 * loss_photometric + \
1.6e-3 * loss_landmark + \
0.2 * loss_perceptual + \
3e-4 * loss_coeff_reg
# clean up
total_loss.backward()
optimizer.step()
# info output
if iteration % print_iter == 0:
# print and save images
print(("[Training] Epoch: %d, Iteration: %d: Total_loss=%0.6f " +
"L_photo=%0.6f L_landmark=%0.6f L_perc=%0.6f L_coeff_reg=%0.6f") %
(epoch, iteration,
total_loss.detach().cpu().numpy().item(),
loss_photometric.detach().cpu().numpy().item(),
loss_landmark.detach().cpu().numpy().item(),
loss_perceptual.detach().cpu().numpy().item(),
loss_coeff_reg.detach().cpu().numpy().item()))
image = save_images(images, rendered_images, landmarks, landmarks_2d, epoch, iteration, vis_dir)
# tensorboard
global_step = train_length * epoch + iteration
writer.add_scalar("train/total_loss", total_loss, global_step)
writer.add_scalar("train/loss_photometric", loss_photometric, global_step)
writer.add_scalar("train/loss_landmark", loss_landmark, global_step)
writer.add_scalar("train/loss_perceptual", loss_perceptual, global_step)
writer.add_scalar("train/loss_coeff_reg", loss_coeff_reg, global_step)
writer.add_image("train/visualization", image, global_step, dataformats="HWC")
if iteration % val_iter == 0:
# val
encoder.eval()
decoder.eval()
renderer.eval()
for local_val_iteration, (images, landmarks, masks) in enumerate(val_loader):
images = images.cuda()
landmarks = landmarks.cuda()
masks = masks.cuda()
# forward
coeff = encoder(images)
face_projection, face_color, landmarks_3d, tri = decoder(coeff)
rendered_images, _ = renderer(face_projection, face_color, tri)
# loss
loss_photometric = photometric_loss(images, rendered_images, masks)
loss_landmark, landmarks_2d = landmark_loss(landmarks_3d, landmarks)
loss_perceptual = perceptual_loss(images, rendered_images, masks)
loss_coeff_reg = coefficient_regularization_loss(coeff)
total_loss = 1.9 * loss_photometric + \
1.6e-3 * loss_landmark + \
0.2 * loss_perceptual + \
3e-4 * loss_coeff_reg
# print and save images
print(("[Validation] Iteration: %d: Total_loss=%0.6f " +
"L_photo=%0.6f L_landmark=%0.6f L_perc=%0.6f L_coeff_reg=%0.6f") %
(global_val_iteration,
total_loss.detach().cpu().numpy().item(),
loss_photometric.detach().cpu().numpy().item(),
loss_landmark.detach().cpu().numpy().item(),
loss_perceptual.detach().cpu().numpy().item(),
loss_coeff_reg.detach().cpu().numpy().item()))
image = save_images(images, rendered_images, landmarks, landmarks_2d,
None, global_val_iteration, vis_dir)
# tensorboard
writer.add_scalar("val/total_loss", total_loss, global_val_iteration)
writer.add_scalar("val/loss_photometric", loss_photometric, global_val_iteration)
writer.add_scalar("val/loss_landmark", loss_landmark, global_val_iteration)
writer.add_scalar("val/loss_perceptual", loss_perceptual, global_val_iteration)
writer.add_scalar("val/loss_coeff_reg", loss_coeff_reg, global_val_iteration)
writer.add_image("val/visualization", image, global_val_iteration, dataformats="HWC")
global_val_iteration += 1
if local_val_iteration > 8:
break
torch.save(encoder.state_dict(), save_encoder_path % epoch)
if decoder_trainable:
torch.save(decoder.state_dict(), save_decoder_path % epoch)
if __name__ == '__main__':
main()