-
Notifications
You must be signed in to change notification settings - Fork 3
/
train.py
155 lines (130 loc) · 5.98 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
import os
from tqdm import tqdm
import torch
import argparse
import shutil
import numpy as np
import torch.nn.functional as F
from torchvision import transforms
from torch.utils.data import DataLoader
from libs.data import ICUDataset, ToTensor, GenerateHeatmap
from libs.loss import JointsMSELoss, Debug_JointsMSELoss, Hard_mining_JointsMSELoss, KpLoss
import extramodel.densenet as densenet
import extramodel.resnet as resnet
import copy
import matplotlib
matplotlib.use("Agg")
from matplotlib import pyplot as plt
import itertools
from libs.metrics import compute_mse
import cv2
def train_epoch(net, epoch, dataLoader, optimizer, loss_list):
net.train()
dataprocess = tqdm(dataLoader)
for batch_item in dataprocess:
images, label = batch_item['image'], batch_item["label"][:, -3:]
if torch.cuda.is_available():
images, label = images.cuda(), label.cuda()
optimizer.zero_grad()
output = net(images)
loss = torch.mean(torch.abs((label - output)))
loss.backward()
optimizer.step()
dataprocess.set_description_str("epoch:{}".format(epoch))
dataprocess.set_postfix_str("loss:{:.4f}".format(loss.item()))
loss_list.append(loss.float())
def test(net, epoch, dataLoader, parser, test_results_list):
net.eval()
cor_abs_mean_error_list, angle_pred_list, angle_gt_list = [], [], []
for batch_num, batch_item in enumerate(dataLoader):
with torch.no_grad():
images, label = batch_item['image'].cuda(), batch_item['label'][:,-3:].cuda()
out = net(images)
out = out.cpu().detach().numpy()
label = label.cpu().detach().numpy()
for i in range(out.shape[0]):
angle_pred_list.append(out[i])
angle_gt_list.append(label[i])
angle_abs_mean_error = np.mean(np.abs(np.array(angle_pred_list) - np.array(angle_gt_list)), 0)
print('Abs Error: %.4f, %.4f, %.4f' % (angle_abs_mean_error[0] * 90, angle_abs_mean_error[1] * 90, angle_abs_mean_error[2] * 90))
def adjust_lr(optimizer, epoch):
if epoch == 0:
return
if (epoch % 30 == 0) & (epoch != 0):
lr = optimizer.param_groups[0]['lr'] * 0.1
for param_group in optimizer.param_groups:
param_group['lr'] = lr
def main(parser):
if os.path.exists(parser["save_path"]):
shutil.rmtree(parser["save_path"])
os.makedirs(parser["save_path"], exist_ok=True)
if os.path.exists(parser["visual_dir"]):
shutil.rmtree(parser["visual_dir"])
os.makedirs(parser["visual_dir"], exist_ok=True)
kwargs = {'num_workers': 8, 'pin_memory': False} if torch.cuda.is_available() else {}
training_dataset = ICUDataset(data_dir=parser["data_dir"], size=parser["img_size"],
transform=transforms.Compose([GenerateHeatmap(parser["img_size"], parser["stride"], parser["sigma"]),
ToTensor()]))
training_data_batch = DataLoader(training_dataset, batch_size=parser["batch_size"],
shuffle=True, drop_last=True, ** kwargs)
val_dataset = ICUDataset(data_dir=parser["data_dir"], phase='test', size=parser["img_size"],
transform=transforms.Compose([GenerateHeatmap(parser["img_size"], parser["stride"], parser["sigma"]),
ToTensor()]))
val_data_batch = DataLoader(val_dataset, batch_size=parser["test_batch_size"], ** kwargs)
if parser["net"] == "dense121":
net = densenet.densenet121(pretrained=True, num_classes=3)
if parser["net"] == "dense161":
net = densenet.densenet161(pretrained=True, num_classes=3)
if parser["net"] == "dense169":
net = densenet.densenet169(pretrained=True, num_classes=3)
if parser["net"] == "dense201":
net = densenet.densenet201(pretrained=True, num_classes=3)
if parser["net"] == "res18":
net = resnet.resnet18(pretrained=True, num_classes=3)
if parser["net"] == "res50":
net = resnet.resnet50(pretrained=True, num_classes=3)
if parser["net"] == "res101":
net = resnet.resnet101(pretrained=True, num_classes=3)
if parser["net"] == "res152":
net = resnet.resnet152(pretrained=True, num_classes=3)
if torch.cuda.is_available():
net = net.cuda()
net = torch.nn.DataParallel(net)
optimizer = torch.optim.SGD(net.parameters(), lr=parser["lr"], momentum=0.9, weight_decay=parser["weight_decay"])
# optimizer = torch.optim.Adam(net.parameters(), lr=3e-3, weight_decay=parser["weight_decay"])
loss_list, test_results_list = [], []
for epoch in range(parser["epochs"]):
adjust_lr(optimizer, epoch)
train_epoch(net, epoch, training_data_batch, optimizer, loss_list)
test(net, epoch, val_data_batch, parser, test_results_list)
plt.plot(np.linspace(0, parser["epochs"], len(loss_list)), loss_list)
plt.savefig(os.path.join('./', 'angle_training_loss.png'))
torch.save(net, os.path.join(os.getcwd(), parser["save_path"], "angle_baseline.pth"))
test_results = np.array(test_results_list)
for k, v in parser.items():
print('{}: {}'.format(k, v))
print()
test_results = np.mean(test_results[-10:, :], 0)
print('min_abs_angle: %.4f, %.4f, %.4f' % (test_results[0]*90, test_results[1]*90, test_results[2]*90))
if __name__ == "__main__":
os.environ["CUDA_VISIBLE_DEVICES"] = "0,1,2,3"
platform = 'sever'
parser = {}
parser["epochs"] = 90
parser["weight_decay"] = 1e-4
parser["lr"] = 1e-2
parser["save_path"] = "logs"
parser["visual_dir"] = ""
parser["img_size"] = [512, 256]
parser["sigma"] = 0.01
parser["stride"] = 4
parser["data_dir"] = ""
parser["batch_size"] = 32
parser["test_batch_size"] = 32
parser["net"] = "dense169"
if platform == 'win':
parser["data_dir"] = ""
parser["net_growth_rate"] = 1
parser["batch_size"] = 2
parser["test_batch_size"] = 2
main(parser)