forked from Eskender-B/roi-tanh
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
292 lines (207 loc) · 9.37 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
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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import numpy as np
from preprocess import ToTensor, ImageDataset
from torch.utils.data import DataLoader
from torchvision import transforms, utils
import argparse
from utils import LOG_INFO
import pickle
from model import Model
parser = argparse.ArgumentParser()
parser.add_argument("--batch_size", default=10, type=int, help="Batch size to use during training.")
parser.add_argument("--display_freq", default=10, type=int, help="Display frequency")
parser.add_argument("--lr", default=0.001, type=float, help="Learning rate for optimizer")
parser.add_argument("--epochs", default=10, type=int, help="Number of epochs to train")
parser.add_argument("--load_model", default=False, type=bool, help="Wether to continue training from last checkpoint")
args = parser.parse_args()
print(args)
if torch.cuda.is_available():
device = torch.device("cuda")
else:
device = torch.device("cpu")
train_dataset = ImageDataset(txt_file='exemplars.txt',
root_dir='data/SmithCVPR2013_dataset_warped',
bg_indexs=set([0]),
transform=transforms.Compose([ToTensor()]))
train_loader = DataLoader(train_dataset, batch_size=args.batch_size,
shuffle=True, num_workers=4)
valid_dataset = ImageDataset(txt_file='tuning.txt',
root_dir='data/SmithCVPR2013_dataset_warped',
bg_indexs=set([0]),
transform=transforms.Compose([ToTensor()]))
valid_loader = DataLoader(valid_dataset, batch_size=args.batch_size,
shuffle=True, num_workers=4)
test_dataset = ImageDataset(txt_file='testing.txt',
root_dir='data/SmithCVPR2013_dataset_warped',
bg_indexs=set([0]),
transform=transforms.Compose([ToTensor()]))
test_loader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=True, num_workers=4)
model = Model()
optimizer = optim.Adam(model.parameters(), lr=args.lr)
model = model.to(device)
criterion1 = nn.L1Loss().to(device)
criterion2 = []
for i in range(6):
criterion2.append(nn.CrossEntropyLoss().to(device))
criterion3 = nn.CrossEntropyLoss().to(device)
def train1(epoch, model, train_loader, optimizer):
loss_list = []
model.train()
for i, batch in enumerate(train_loader):
optimizer.zero_grad()
image, rects = batch['image'].to(device), batch['rects'].to(device)
pred_rects = model(image,rects_only=True)
loss = criterion1(pred_rects, rects)
loss.backward()
optimizer.step()
loss_list.append(loss.item())
if i % args.display_freq == 0:
msg = "Epoch %02d, Iter [%03d/%03d], train loss = %.4f" % (
epoch, i, len(train_loader), np.mean(loss_list))
LOG_INFO(msg)
loss_list.clear()
def train2(epoch, model, train_loader, optimizer):
loss_list = []
loss_list1 = []
loss_list2 = []
loss_list3 = []
model.train()
for j, batch in enumerate(train_loader):
optimizer.zero_grad()
image, labels, rects = batch['image'].to(device), batch['labels'].to(device), batch['rects'].to(device)
pred_rects, segm, full = model(image)
## Loss1
loss1 = criterion1(pred_rects, rects)
## Loss2
parts = [torch.tensor([]).to(device).long()] * 6
b,c,h,w = image.shape
rects_ = pred_rects.round().long().to('cpu').clamp(0,511)
for bb in range(b):
# Non-mouth
for i in range(5):
x1, y1, x2, y2 = rects_[bb][i*4:(i+1)*4]
if x2-x1 <= 0 or y2-y1<=0:
x1, y1, x2, y2 = torch.tensor([0,0,511,511]).long()
p = torch.gather(labels[bb][2+i], 0, torch.arange(y1, y2+1).unsqueeze(1).repeat_interleave(w,dim=1).to(device))
p = torch.gather(p, 1, torch.arange(x1, x2+1).unsqueeze(0).repeat_interleave(y2-y1+1,dim=0).to(device))
p = torch.stack([1-p, p]) # background
parts[i] = torch.cat([parts[i], F.interpolate(p.float().unsqueeze(0), size=[128,128], mode='bilinear').squeeze(0).argmax(dim=0)], dim=0)
# Mouth
x1, y1, x2, y2 = rects_[bb][20:24]
if x2-x1 <= 0 or y2-y1<=0:
x1, y1, x2, y2 = torch.tensor([0,0,511,511]).long()
p = torch.gather(labels[bb][7:10], 1, torch.arange(y1, y2+1).unsqueeze(1).repeat_interleave(w,dim=1).unsqueeze(0).repeat_interleave(3,dim=0).to(device))
p = torch.gather(p, 2, torch.arange(x1, x2+1).unsqueeze(0).repeat_interleave(y2-y1+1,dim=0).unsqueeze(0).repeat_interleave(3,dim=0).to(device))
p = torch.cat([1-p.sum(dim=0,keepdim=True), p], 0) # background
parts[5] = torch.cat([parts[5], F.interpolate(p.float().unsqueeze(0), size=[128,128], mode='bilinear').squeeze(0).argmax(dim=0)], dim=0)
loss2 = []
for i in range(6):
loss2.append(criterion2[i](segm[i], parts[i].view(b,128,128)))
## Loss3
face_hair = labels.float().index_select(1, torch.tensor([1,10]).long().to(device))
bg = 1. - face_hair.sum(dim=1, keepdim=True)
face_hair_bg = torch.cat([bg,face_hair], dim=1)
full_l = F.interpolate(face_hair_bg, size=[128,128], mode='bilinear').argmax(dim=1)
loss3 = criterion3(full, full_l)
## Total loss
tot_loss = loss1 + sum(loss2) + loss3
tot_loss.backward()
optimizer.step()
loss_list1.append(loss1.item())
loss_list2.append((sum(loss2)/6.).item())
loss_list3.append(loss3.item())
loss_list.append(tot_loss.item())
if j % args.display_freq == 0:
msg = "Epoch %02d, Iter [%03d/%03d], train loss = %.4f, %.4f, %.4f, %.4f" % (
epoch, j, len(train_loader), np.mean(loss_list1),np.mean(loss_list2),np.mean(loss_list3),np.mean(loss_list))
LOG_INFO(msg)
loss_list.clear()
loss_list1.clear()
loss_list2.clear()
loss_list3.clear()
def evaluate1(model, loader):
epoch_loss = 0
model.eval()
with torch.no_grad():
for batch in loader:
image, rects = batch['image'].to(device), batch['rects'].to(device)
pred_rects = model(image,rects_only=True)
loss = criterion1(pred_rects, rects)
epoch_loss += loss.item()
return epoch_loss / len(loader)
def evaluate2(model, loader):
epoch_loss1 = 0
epoch_loss2 = 0
epoch_loss3 = 0
epoch_tot = 0
model.eval()
with torch.no_grad():
for batch in loader:
image, labels, rects = batch['image'].to(device), batch['labels'].to(device), batch['rects'].to(device)
pred_rects, segm, full = model(image)
## Loss1
loss1 = criterion1(pred_rects, rects)
## Loss2
parts = [torch.tensor([]).to(device).long()] * 6
b,c,h,w = image.shape
rects_ = pred_rects.round().long().to('cpu').clamp(0,511)
for bb in range(b):
# Non-mouth
for i in range(5):
x1, y1, x2, y2 = rects_[bb][i*4:(i+1)*4]
if x2-x1 <= 0 or y2-y1<=0:
x1, y1, x2, y2 = torch.tensor([0,0,511,511]).long()
p = torch.gather(labels[bb][2+i], 0, torch.arange(y1, y2+1).unsqueeze(1).repeat_interleave(w,dim=1).to(device))
p = torch.gather(p, 1, torch.arange(x1, x2+1).unsqueeze(0).repeat_interleave(y2-y1+1,dim=0).to(device))
p = torch.stack([1-p, p]) # background
parts[i] = torch.cat([parts[i], F.interpolate(p.float().unsqueeze(0), size=[128,128], mode='bilinear').squeeze(0).argmax(dim=0)], dim=0)
# Mouth
x1, y1, x2, y2 = rects_[bb][20:24]
if x2-x1 <= 0 or y2-y1<=0:
x1, y1, x2, y2 = torch.tensor([0,0,511,511]).long()
p = torch.gather(labels[bb][7:10], 1, torch.arange(y1, y2+1).unsqueeze(1).repeat_interleave(w,dim=1).unsqueeze(0).repeat_interleave(3,dim=0).to(device))
p = torch.gather(p, 2, torch.arange(x1, x2+1).unsqueeze(0).repeat_interleave(y2-y1+1,dim=0).unsqueeze(0).repeat_interleave(3,dim=0).to(device))
p = torch.cat([1-p.sum(dim=0,keepdim=True), p], 0) # background
parts[5] = torch.cat([parts[5], F.interpolate(p.float().unsqueeze(0), size=[128,128], mode='bilinear').squeeze(0).argmax(dim=0)], dim=0)
loss2 = []
for i in range(6):
loss2.append(criterion2[i](segm[i], parts[i].view(b,128,128)))
## Loss3
face_hair = labels.float().index_select(1, torch.tensor([1,10]).long().to(device))
bg = 1. - face_hair.sum(dim=1, keepdim=True)
face_hair_bg = torch.cat([bg,face_hair], dim=1)
full_l = F.interpolate(face_hair_bg, size=[128,128], mode='bilinear').argmax(dim=1)
loss3 = criterion3(full, full_l)
## Total loss
tot_loss = loss1 + sum(loss2) + loss3
epoch_loss1 += loss1.item()
epoch_loss2 += (sum(loss2)/6).item()
epoch_loss3 += loss3.item()
epoch_tot += tot_loss.item()
return [epoch_loss1/len(loader), epoch_loss2/len(loader), epoch_loss3/len(loader), epoch_tot/len(loader)]
LOSS = 100
epoch_min = 1
if args.load_model:
model = pickle.load(open('res/saved-model.pth', 'rb'))
for epoch in range(1, 1 + 1):
train1(epoch, model, train_loader, optimizer)
valid_loss = evaluate1(model, valid_loader)
msg = '...Epoch %02d, val loss = %.4f' % (epoch, valid_loss)
LOG_INFO(msg)
for epoch in range(1, args.epochs + 1):
train2(epoch, model, train_loader, optimizer)
l1,l2,l3,lt = evaluate2(model, valid_loader)
if l3 < LOSS:
LOSS = l3
epoch_min = epoch
pickle.dump(model, open('res/saved-model.pth', 'wb'))
msg = '...Epoch %02d, val loss = %.4f, %.4f, %.4f, %.4f' % (epoch, l1,l2,l3,lt)
LOG_INFO(msg)
model = pickle.load(open('res/saved-model.pth', 'rb'))
msg = 'Min @ Epoch %02d, val loss = %.4f' % (epoch_min, LOSS)
LOG_INFO(msg)
l1,l2,l3,lt = evaluate2(model, test_loader)
LOG_INFO('Finally, test loss = %.4f, %.4f, %.4f, %.4f' % (l1,l2,l3,lt))