Repository navigation
Expand file tree
/
Copy pathtrain.py
More file actions
184 lines (163 loc) · 7.88 KB
/
Copy pathtrain.py
File metadata and controls
184 lines (163 loc) · 7.88 KB
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
import argparse
import os
import random
from datetime import datetime
import numpy as np
import torch
import torch.nn.functional as F
import torch.utils.data as D
from tqdm import tqdm
from TransInvNet.model.model import TransInvNet, CONFIGS
from TransInvNet.utils.dataloader import PolypDataset
from TransInvNet.utils.utils import clip_gradient, adjust_lr
from eval import Metrics
def set_seeds(seed=42):
random.seed(seed)
os.environ['PYTHONHASHSEED'] = str(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
def loss_fn(pred, mask):
mask = mask.type_as(pred)
weit = 1 + 5 * torch.abs(F.avg_pool2d(mask, kernel_size=31, stride=1, padding=15) - mask)
wbce = F.binary_cross_entropy_with_logits(pred, mask, reduction='none')
wbce = (weit * wbce).sum(dim=(2, 3)) / weit.sum(dim=(2, 3))
pred = torch.sigmoid(pred)
inter = ((pred * mask) * weit).sum(dim=(2, 3))
union = ((pred + mask) * weit).sum(dim=(2, 3))
wiou = 1 - (inter + 1) / (union - inter + 1)
return (wbce + wiou).mean()
def train(opt, train_loader, model, optimizer, epoch, loss_fn):
model.train()
optimizer.zero_grad()
loss_record = []
tbar = tqdm(train_loader, desc='\r')
for i, pack in enumerate(tbar, start=1):
# ---- data preparation ----
images, gts = pack
images = images.cuda()
gts = gts.float().cuda()
# ---- forward ----
pred = model(images)
# ---- compute loss ----
loss = loss_fn(pred, gts)
# ---- record loss ----
loss_record.append(float(loss.data))
# ---- gradient accumulation ----
loss /= opt.accumulation_step
# ---- backward ----
loss.backward()
clip_gradient(optimizer, opt.clip)
if i % opt.accumulation_step == 0:
optimizer.step()
optimizer.zero_grad()
# ---- train visualization ----
if i == 1 or i % 20 == 0 or i == len(train_loader):
tbar.set_description('{} Epoch [{:03d}/{:03d}], Step [{:04d}/{:04d}], Train loss {:.04f}'.
format(datetime.now(), epoch, opt.epoch, i, len(train_loader),
np.mean(loss_record)))
save_path = 'outputs/{}/train/'.format(opt.output_path)
os.makedirs(save_path, exist_ok=True)
if epoch % 10 == 0:
torch.save(model.state_dict(), save_path + 'TransInvNet-%d.pth' % epoch)
print('[Saving model weight:]', save_path + 'TransInvNet-%d.pth' % epoch)
def test(opt, test_loader, model, epoch, loss_fn, best_loss, best_metrics):
model.eval()
loss_record = []
mDice, mMae, mIou = [], [], []
with torch.no_grad():
tbar = tqdm(test_loader, desc='\r')
for i, pack in enumerate(tbar, start=1):
# ---- data preparation ----
images, gts = pack
images = images.cuda()
gts = gts.cuda()
# ---- forward ----
pred = model(images)
# ---- compute loss ----
loss = loss_fn(pred, gts)
# --- compute metric ----
pred = pred.sigmoid().cpu().numpy().squeeze()
pred = (pred - pred.min()) / (pred.max() - pred.min() + 1e-8)
mask = gts.cpu().numpy().squeeze()
metrics = Metrics(pred, mask)
iou, mae, dice = metrics.calculate_iou(), metrics.calculate_mae(), metrics.calculate_dice()
# ---- record loss and metrics ----
loss_record.append(float(loss.data))
mDice.append(dice)
mMae.append(mae)
mIou.append(iou)
tbar.set_description('{} Epoch [{:03d}/{:03d}], Step[{:03d}/{:03d}], '
'Test loss {:.04f}, Test mean dice {:.04f}, '
'Test mean abs error {:.04f}, Test mean iou {:.04f}'.
format(datetime.now(), epoch, opt.epoch, i, len(test_loader),
np.mean(loss_record), np.mean(mDice),
np.mean(mMae), np.mean(mIou)))
current_loss = np.mean(loss_record)
metrics = (np.mean(mDice) + np.mean(mIou) + (1 - np.mean(mMae))) / 3
if current_loss < best_loss[-1] or metrics > best_metrics[-1]:
best_loss.append(current_loss)
best_metrics.append(metrics)
save_path = 'outputs/{}/train/'.format(opt.output_path)
torch.save(model.state_dict(), save_path + 'TransInvNet-best.pth')
print('[Saving model weight], current best test loss is {:.04f}, current best mean dice is {:.04f}, '
'current best mean abs error is {:.04f}, current best mean iou is {:.04f}'
.format(current_loss, np.mean(mDice), np.mean(mMae), np.mean(mIou)))
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--epoch', type=int,
default=100, help='epoch number')
parser.add_argument('--lr', type=float,
default=1e-4, help='learning rate')
parser.add_argument('--batch_size', type=int,
default=4, help='training batch size')
parser.add_argument('--accumulation_step', type=int,
default=4, help='accumulation step to implement gradient accumulation')
parser.add_argument('--img_size', type=int,
default=352, help='training dataset size')
parser.add_argument('--clip', type=float,
default=0.5, help='gradient clipping margin')
parser.add_argument('--cfg', type=str,
default='ViT-B_16', help='configs for our model, choose from ViT-B_8, '
'ViT-B_16, ViT-B_32, ViT-L_16, ViT-L_32')
parser.add_argument('--train_path', type=str,
default='datasets/polyp-dataset/TrainDataset', help='path to train dataset')
parser.add_argument('--test_path', type=str,
default='datasets/polyp-dataset/TestDataset/Kvasir', help='path to test dataset')
parser.add_argument('--output_path', type=str,
default='exp{}'.format(datetime.now().strftime('%m%d%H%M')), help='path to output')
parser.add_argument('--seed', type=int,
default=1234, help='random seed')
opt = parser.parse_args()
set_seeds(seed=opt.seed) # set seed
# ---- prepare dataset ----
train_dataset = PolypDataset(opt.train_path, image_dir='images', mask_dir='masks',
new_size=(opt.img_size, opt.img_size))
test_dataset = PolypDataset(opt.test_path, image_dir='images', mask_dir='masks',
new_size=(opt.img_size, opt.img_size))
train_loader = D.DataLoader(train_dataset, batch_size=opt.batch_size, shuffle=True)
test_loader = D.DataLoader(test_dataset, batch_size=1, shuffle=True)
# ---- build models ----
torch.cuda.set_device(0) # set your gpu device
cfg = CONFIGS[opt.cfg]
model = TransInvNet(cfg, opt.img_size, vis=True, pretrained=True).cuda()
# ---- flops and params ----
# from TransInvNet.utils.utils import CalParams
# x = torch.randn(1, 3, 352, 352).cuda()
# CalParams(model, x)
rednet_params = model.rednet.parameters()
vit_params = model.transformer.parameters()
decoder_params = model.decoder.parameters()
optimizer = torch.optim.AdamW([
{"params": rednet_params},
{"params": vit_params},
{"params": decoder_params, "lr": opt.lr * 10},
], opt.lr, weight_decay=5e-4)
print("#" * 20, "Start Training", "#" * 20)
best_loss, best_metrics = [np.inf], [-np.inf]
for epoch in range(1, opt.epoch + 1):
adjust_lr(optimizer, epoch, opt.epoch, 0.9)
train(opt, train_loader, model, optimizer, epoch, loss_fn)
test(opt, test_loader, model, epoch, loss_fn, best_loss, best_metrics)