-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
189 lines (160 loc) · 7.62 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
# python imports
import os
import glob
import warnings
import time
# external imports
import torch
import numpy as np
import SimpleITK as sitk
import math
from torch.optim import Adam
import torch.utils.data as Data
import torch.nn.functional as F
import itertools
# internal imports
from Model import losses
from Model.config import args
from Model.datagenerators_affine import Dataset
from Model.model import U_Network, SpatialTransformer, Extractor
from settings import setting
os.environ["CUDA_VISIBLE_DEVICES"] = "2"
dim = 2
simfunctions = {
"euclidean" : lambda x, y: -torch.norm(x - y, p=2, dim=1).mean(),
"L1" : lambda x, y: -torch.norm(x - y, p=1, dim=1).mean(),
"MSE" : lambda x, y: -(x - y).pow(2).mean(),
"L3" : lambda x, y: -torch.norm(x - y, p=3, dim=1).mean(),
"Linf" : lambda x, y: -torch.norm(x - y, p=float("inf"), dim=1).mean(),
"soft_corr" : lambda x, y: F.softplus(x*y).sum(axis=1),
"corr" : lambda x, y: (x*y).sum(axis=1),
"cosine" : lambda x, y: F.cosine_similarity(x, y, dim=1, eps=1e-8).mean(),
"angular" : lambda x, y: F.cosine_similarity(x, y, dim=1, eps=1e-8).acos().mean() / math.pi,
}
sim_loss_fn = simfunctions["cosine"]
def count_parameters(model):
model_parameters = filter(lambda p: p.requires_grad, model.parameters())
params = sum([np.prod(p.size()) for p in model_parameters])
return params
def make_dirs():
if not os.path.exists(setting['model_dir']):
os.makedirs(setting['model_dir'])
if not os.path.exists(setting['log_dir']):
os.makedirs(setting['log_dir'])
def save_image(img, ref_img, name, if_flow = False):
if if_flow == False:
img = sitk.GetImageFromArray(img[0, 0, ...].cpu().detach().numpy())
ref_img = sitk.GetImageFromArray(ref_img[0, 0, ...].cpu().detach().numpy())
img.SetOrigin(ref_img.GetOrigin())
img.SetDirection(ref_img.GetDirection())
img.SetSpacing(ref_img.GetSpacing())
sitk.WriteImage(img, os.path.join(setting['result_dir'], name))
else:
img = sitk.GetImageFromArray(img[0, ...].cpu().detach().numpy())
sitk.WriteImage(img, os.path.join(setting['result_dir'], name))
def train():
# 创建需要的文件夹并指定gpu
make_dirs()
device = torch.device('cuda:{}'.format(setting['gpu']) if torch.cuda.is_available() else 'cpu')
# 日志文件
time_str = time.strftime("%Y-%m-%d-%H_%M_%S", time.localtime())
log_name = str(setting['n_iter'])+"_"+str(setting['lr_unet'])+"_"+str(setting['lr_ext'])+'_'+str(setting['alpha'])+'_'+str(time_str)
print("log_name: ", log_name)
f = open(os.path.join(setting['log_dir'], log_name + ".txt"), "w")
vol_size = (240, 240)
# 创建配准网络(UNet_m2f)和STN
nf_enc = [16, 32, 32, 32]
if setting['model']== "vm1":
nf_dec = [32, 32, 32, 32, 8, 8]
else:
nf_dec = [32, 32, 32, 32, 32, 16, 16]
UNet_m2f = U_Network(len(vol_size), nf_enc, nf_dec).to(device)
UNet_f2m = U_Network(len(vol_size), nf_enc, nf_dec).to(device)
STN1 = SpatialTransformer(vol_size).to(device)
STN2 = SpatialTransformer(vol_size).to(device)
Extractor_m1 = Extractor().to(device)
Extractor_m2 = Extractor().to(device)
UNet_m2f.train()
UNet_f2m.train()
STN1.train()
STN2.train()
Extractor_m1.train()
Extractor_m2.train()
# 模型参数个数
print("UNet_m2f: ", count_parameters(UNet_m2f))
print("Extractor: ", count_parameters(Extractor_m1))
print("STN: ", count_parameters(STN1))
print("==========>")
# Set opt_unetimizer and losses
opt_unet = Adam(itertools.chain(UNet_m2f.parameters(), UNet_f2m.parameters()),lr = setting['lr_unet'])
opt_ext = Adam(itertools.chain(Extractor_m1.parameters(), Extractor_m2.parameters()), lr = setting['lr_ext'])
grad_loss_fn = losses.gradient_loss
# Get all the names of the training data
DS_train = Dataset(["../../Dataset/MICCAI_BraTS/Deformed2D/HGG/train/*/t1.nii.gz",
"../../Dataset/MICCAI_BraTS/Deformed2D/HGG/train/*/t2.nii.gz"])
print("Number of training images: ", len(DS_train))
DL_train = Data.DataLoader(DS_train, batch_size = setting['batch_size'], shuffle = True, num_workers = 4, drop_last = True)
# DS_val = Dataset(["../../Dataset/MICCAI_BraTS/Deformed2D/HGG/validation/*/t1.nii.gz",
# "../../Dataset/MICCAI_BraTS/Deformed2D/HGG/validation/*/t2.nii.gz"])
# print("Number of training images: ", len(DS_val))
# DL_val = Data.DataLoader(DS_val, batch_size = setting['batch_size'], shuffle = True, num_workers = 0, drop_last = True)
mod1 = slice(0, 240)
mod2 = slice(240, 480)
# Training loop.
for i in range(1, setting['n_iter'] + 1):
# Generate the moving images and convert them to tensors.
imgs = iter(DL_train).next()
# [B, C, D, W, H]
input_moving = imgs[:, :, :, mod1].to(device).float()
input_fixed = imgs[:, :, :,mod2].to(device).float()
# Run the data through the model to produce warp and flow field
torch.autograd.set_detect_anomaly(True)
flow_m2f = UNet_m2f(input_moving, input_fixed)
flow_f2m = UNet_f2m(input_fixed, input_moving)
m2f = STN1(input_moving, flow_m2f) # wraped_moving. M1. vs fixed
f2m = STN2(input_fixed, flow_f2m) # wraped_fixed. M2. vs moving
""" 提取公共表示"""
feature_m = Extractor_m1(input_moving, layer=6)
feature_f = Extractor_m2(input_fixed, layer=6)
feature_f2m = Extractor_m1(f2m, layer=6)
feature_m2f = Extractor_m2(m2f, layer=6)
""" 计算 Loss """
sim_loss = - sim_loss_fn(feature_m, feature_f2m) - sim_loss_fn(feature_f, feature_m2f)
grad_loss = grad_loss_fn(flow_m2f, dim) + grad_loss_fn(flow_f2m, dim)
loss = sim_loss + setting['alpha']* grad_loss
print("i: %d loss: %f sim: %f grad: %f" % (i, loss.item(), sim_loss.item(), grad_loss.item()), flush=True)
print("%d, %f, %f, %f" % (i, loss.item(), sim_loss.item(), grad_loss.item()), file=f)
""" Backward """
opt_unet.zero_grad()
opt_ext.zero_grad()
with torch.autograd.set_detect_anomaly(True):
loss.backward()
opt_unet.step()
opt_ext.step()
if i % setting['n_save_iter'] == 0:
# Save model checkpoint
save_file_name_unet1 = os.path.join(setting['model_dir'], '%d_unet1.pth' % i)
save_file_name_unet2 = os.path.join(setting['model_dir'], '%d_unet2.pth' % i)
save_file_name_ext1 = os.path.join(setting['model_dir'], '%d_ext1.pth' % i)
save_file_name_ext2 = os.path.join(setting['model_dir'], '%d_ext2.pth' % i)
torch.save(UNet_m2f.state_dict(), save_file_name_unet1)
torch.save(UNet_f2m.state_dict(), save_file_name_unet2)
torch.save(Extractor_m1.state_dict(), save_file_name_ext1)
torch.save(Extractor_m2.state_dict(), save_file_name_ext2)
# Save images
m_name = str(i) + "_m.nii.gz"
f_name = str(i) + '_f.nii.gz'
m2f_name = str(i) + "_m2f.nii.gz"
f2m_name = str(i) + "_f2m.nii.gz"
flow_m2f_name = str(i) + '_flow_m2f.nii.gz'
save_image(input_moving, input_fixed, m_name)
save_image(input_fixed, input_fixed, f_name)
save_image(m2f, input_fixed, m2f_name)
save_image(f2m, input_fixed, f2m_name)
save_image(flow_m2f, input_fixed, flow_m2f_name)
print("warped images have saved.")
f.close()
if __name__ == "__main__":
with warnings.catch_warnings():
warnings.filterwarnings("ignore", category = DeprecationWarning)
train()