-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
322 lines (250 loc) · 12.7 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
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
import os
import numpy as np
import torch
import torch.utils.data.dataset as Dataset
from skimage.transform import resize
import SimpleITK as sitk
# import test_data as Data
import data_utils as Data
import torch.utils.data as Datas
import Network as Network
import math
import torch.nn.functional as F
import metrics as metrics
from torch.autograd import Variable
device = torch.device("cuda:0")
data = Data.dataset3
dataloder = Datas.DataLoader(dataset=data,batch_size=1,shuffle=True)
Gen = Network.SegNet(inchannel=1).to(device)
Dis1 = Network.high_discriminator(channel=1).to(device)
# Dis2 = Network.low_discriminator(channel=1).to(device)
fake_A_buffer = metrics.ReplayBuffer()
fake_B_buffer = metrics.ReplayBuffer()
# ###
pretrained_dict = torch.load('../pklh/generattor_epoch_78Network.pkl')
model_dict = Gen.state_dict()
pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict}
model_dict.update(pretrained_dict)
Gen.load_state_dict(model_dict)
####
# pretrained_dict = torch.load('../pklh/Dis1_epoch_81Network.pkl')
# model_dict = Dis1.state_dict()
# pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict}
# model_dict.update(pretrained_dict)
# Dis1.load_state_dict(model_dict)
####
# pretrained_dict = torch.load('../pkl/Dis2_epoch_7Network.pkl')
# model_dict = Dis2.state_dict()
# pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict}
# model_dict.update(pretrained_dict)
# Dis2.load_state_dict(model_dict)
criterion_L1 = torch.nn.L1Loss()
criterion_MSE = torch.nn.MSELoss()
criterion_BCE = torch.nn.BCEWithLogitsLoss()
opt_gen = torch.optim.RMSprop(Gen.parameters(),lr=0.0001)
opt_dis1 = torch.optim.RMSprop(Dis1.parameters(),lr=0.0001)
# opt_dis2 = torch.optim.RMSprop(Dis2.parameters(),lr=0.00001)
cuda = torch.cuda.is_available()
Tensor = torch.cuda.FloatTensor if cuda else torch.Tensor
def compute_local_sums(I, J, filt, stride, padding, win):
I2 = I * I
J2 = J * J
IJ = I * J
I_sum = F.conv2d(I, filt, stride=stride, padding=padding)
J_sum = F.conv2d(J, filt, stride=stride, padding=padding)
I2_sum = F.conv2d(I2, filt, stride=stride, padding=padding)
J2_sum = F.conv2d(J2, filt, stride=stride, padding=padding)
IJ_sum = F.conv2d(IJ, filt, stride=stride, padding=padding)
win_size = np.prod(win)
u_I = I_sum / win_size
u_J = J_sum / win_size
cross = IJ_sum - u_J * I_sum - u_I * J_sum + u_I * u_J * win_size
I_var = I2_sum - 2 * u_I * I_sum + u_I * u_I * win_size
J_var = J2_sum - 2 * u_J * J_sum + u_J * u_J * win_size
return I_var, J_var, cross
def ncc_loss(I, J, win=None):
"""
calculate the normalize cross correlation between I and J
assumes I, J are sized [batch_size, *vol_shape, nb_feats]
"""
ndims = len(list(I.size())) - 2
assert ndims in [1, 2, 3], "volumes should be 1 to 3 dimensions. found: %d" % ndims
if win is None:
win = [9] * ndims
conv_fn = getattr(F, 'conv%dd' % ndims)
I2 = I * I
J2 = J * J
IJ = I * J
sum_filt = torch.ones([1, 1, *win]).to("cuda")
pad_no = math.floor(win[0] / 2)
if ndims == 1:
stride = (1)
padding = (pad_no)
elif ndims == 2:
stride = (1, 1)
padding = (pad_no, pad_no)
else:
stride = (1, 1, 1)
padding = (pad_no, pad_no, pad_no)
I_var, J_var, cross = compute_local_sums(I, J, sum_filt, stride, padding, win)
cc = cross * cross / (I_var * J_var + 1e-5)
return -1 * torch.mean(cc)
for epoch in range(200):
for step, (sax_img, two_image, three_image, four_image, seg_labels) in enumerate(dataloder):
flag = 1
sax_img = sax_img.to(device).float()
two_image = two_image.to(device).float()
three_image = three_image.to(device).float()
four_image = four_image.to(device).float()
seg_labels = seg_labels.to(device).float()
# dense_sax_img = dense_sax_img.to(device).float()
# dense_sax_img2 = dense_sax_img2.to(device).float()
# dense_sax_img[:, :, 48:, :] = 0
# print(dense_sax_img.size())
valid1 = Variable(Tensor(np.ones((1, *Dis1.output_shape))), requires_grad=False)
fake1 = Variable(Tensor(np.zeros((1, *Dis1.output_shape))), requires_grad=False)
# valid2 = Variable(Tensor(np.ones((1, *Dis2.output_shape))), requires_grad=False)
# fake2 = Variable(Tensor(np.zeros((1, *Dis2.output_shape))), requires_grad=False)
#####Seg Training
for p in Dis1.parameters(): # reset requires_grad
p.requires_grad = False # they are set to False below in netG update
# for p in Dis2.parameters(): # reset requires_grad
# p.requires_grad = False # they are set to False below in netG update
opt_gen.zero_grad()
gen_out = Gen(sax_img, two_image, three_image, four_image)
dense_sax = gen_out[0]
coarse_dense_sax = gen_out[1]
coarse_suoxiao = gen_out[2]
suoxiao = gen_out[3]
seg_result = gen_out[4]
# print(seg_labels.shape)
lax1 = dense_sax[:,:,:,:,64]
lax2 = dense_sax[:, :, :, 64, :]
loss_seg = metrics.DiceMeanLoss()(seg_result, seg_labels)
loss_suoxiao = criterion_L1(suoxiao, sax_img)
# loss_suoxiao1 = criterion_L1(coarse_suoxiao, sax_img)
dz = torch.abs(dense_sax[:, :, 1:, :, :] - dense_sax[:, :, :-1, :, :])
dx = torch.abs(dense_sax[:, :, :, 1:, :] - dense_sax[:, :, :, :-1, :])
dy = torch.abs(dense_sax[:, :, :, :, 1:] - dense_sax[:, :, :, :, :-1])
# lossflow = (torch.mean(dz.mul(dz))+torch.mean(dx.mul(dx))+torch.mean(dy.mul(dy)))/3
lossflow = torch.mean(dz.mul(dz))
dz1 = torch.abs(coarse_dense_sax[:, :, 1:, :, :] - coarse_dense_sax[:, :, :-1, :, :])
lossflow1 = torch.mean(dz1.mul(dz1))
loss_GAN_1_2 = criterion_MSE(Dis1(dense_sax[:,:,:64,:,64]), fake1)
loss_GAN_1_4 = criterion_MSE(Dis1(dense_sax[:, :, :64, 64, :]), fake1)
loss_GAN_2_2 = criterion_MSE(Dis1(dense_sax[:,:,:64,:,64]), valid1)
loss_GAN_2_4 = criterion_MSE(Dis1(dense_sax[:, :, :64, 64, :]), valid1)
# loss_GAN_1 = (loss_GAN_1_2+loss_GAN_1_4)*0.5
# loss_GAN_2 = (loss_GAN_2_2+loss_GAN_2_4)*0.5
# loss_GAN_1 = (10*loss_GAN_1+loss_GAN_2)*0.5
# if loss_GAN_1 >=0.5:
#
# if loss_GAN_2 >=0.5:
# loss_GAN_1 = loss_GAN_1 - 0.5
# loss_GAN_2 = loss_GAN_2-0.5
# else:
# loss_GAN_2 = loss_GAN_2+0.5
# loss_GAN_1 = loss_GAN_1 - 0.5
# else:
#
# if loss_GAN_2 >=0.5:
# loss_GAN_1 = loss_GAN_1 + 0.5
# loss_GAN_2 = loss_GAN_2-0.5
# else:
# loss_GAN_2 = loss_GAN_2+0.5
# loss_GAN_1 = loss_GAN_1 + 0.5
loss_GAN_1 = (loss_GAN_1_2+loss_GAN_1_4+loss_GAN_2_2+loss_GAN_2_4)/4
# loss_Lax = criterion_L1(lax1[:,:,60:,:], dense_sax_img[:,:,60:,:])
# loss_Lax2 = criterion_L1(lax1, dense_sax_img_mirror)
# loss_Lax = min(loss_Lax1,loss_Lax2)
# loss_Lax = (1+ncc_loss(lax1, dense_sax_img)+1+ncc_loss(lax2, dense_sax_img2))/8
# print(1+ncc_loss(dense_sax_img2, dense_sax_img2))
# loss_GAN_lax1 = criterion_MSE(Dis2(lax1), valid2)
# loss_GAN_lax2 = criterion_MSE(Dis2(lax2), valid2)
# loss_GAN_lax = loss_GAN_lax1
# if loss_GAN_2>1:
# loss_GAN_2 = loss_GAN_2-loss_GAN_2
#loss_GAN = torch.abs(loss_GAN_1-0.5)#+loss_GAN_lax
# loss_g =20*lossflow + 80*loss_suoxiao+10*loss_seg+80*loss_loss_GANLax+
# = 4*lossflow + 5*loss_suoxiao + 5*loss_suoxiao1 + 5*loss_seg + 1*loss_Lax + loss_GAN
loss_g = 5*lossflow + 5*loss_suoxiao + 1*loss_seg + loss_GAN_1+5*lossflow1
# loss_g = 20 * loss_suoxiao + 10 * loss_seg + 10 * loss_Lax + loss_GAN####withoutFLOW
# loss_g = 80 * lossflow + 20 * loss_suoxiao + 10 * loss_seg + 10 * loss_Lax###withoutGAN
# loss_g = 80 * lossflow + 10 * loss_seg + 10 * loss_Lax + loss_GAN###withoutsuoxiao
# loss_g = 80 * lossflow + 20 * loss_suoxiao + 10 * loss_seg + loss_GAN####withou lax
loss_g.backward(retain_graph=True)
opt_gen.step()
if epoch % 1 == 0:
######High Dis
for p in Dis1.parameters(): # reset requires_grad
p.requires_grad = True # they are set to False below in netG update
opt_dis1.zero_grad()
loss_real = criterion_MSE(Dis1(coarse_dense_sax[:, :, :64, :, 64]), valid1)
loss_two = criterion_MSE(Dis1(two_image[:, :, :64, :, 64]), fake1)
loss_three = criterion_MSE(Dis1(three_image[:, :, :64, :, 64]), fake1)
loss_four = criterion_MSE(Dis1(four_image[:, :, :64, :, 64]), fake1)
loss_dis1 = (loss_real + (loss_two+loss_three+loss_four)/3)/2
# fake_A_ = fake_A_buffer.push_and_pop(dense_sax[:, :, :64, :, :])
# loss_fake = criterion_MSE(Dis1(fake_A_.detach()), valid1)
# loss_dis1 = (loss_real + loss_fake) / 2
loss_dis1.backward()
opt_dis1.step()
#
# #######Low Dis
# for p in Dis2.parameters(): # reset requires_grad
# p.requires_grad = True # they are set to False below in netG update
# opt_dis2.zero_grad()
#
#
# loss_real = criterion_MSE(Dis2(dense_sax_img), valid2)
# fake_B_ = fake_B_buffer.push_and_pop(lax1)
# loss_fake1 = criterion_MSE(Dis2(fake_B_.detach()), fake2)
# # fake_B_ = fake_B_buffer.push_and_pop(lax2)
# # loss_fake2 = criterion_MSE(Dis2(fake_B_.detach()), fake2)
# loss_dis2 = (loss_real + loss_fake1) / 2
# if loss_dis2<0.1:
# loss_dis2 = loss_dis2+1
# loss_dis2.backward()
# opt_dis2.step()
#####Log
if step % 2 == 0:
torch.save(Gen.state_dict(), '../pklh/generattor_epoch_' + str(epoch) + 'Network.pkl')
torch.save(Dis1.state_dict(), '../pklh/Dis1_epoch_' + str(epoch) + 'Network.pkl')
# torch.save(Dis2.state_dict(), '../pkl/Dis2_epoch_' + str(epoch) + 'Network.pkl')
if step % 1 == 0:
print('EPOCH:', epoch, '|Step:', step, '|loss_seg: %.3e' % loss_seg.data.cpu().numpy(), '|loss_suoxiao: %.3e' % loss_suoxiao.data.cpu().numpy(), '|lossflow: %.3e' % lossflow1.data.cpu().numpy(), '|loss_GAN: %.3e' % loss_GAN_1.data.cpu().numpy(), '|loss_dis: %.3e' % loss_dis1.data.cpu().numpy())
# torch.cuda.empty_cache()
if step % 1 == 0:
with torch.no_grad():
pt = dense_sax[0, 0, :, :, :].data.cpu().numpy()
# pt = np.transpose(pt, (2, 1, 0))
out = sitk.GetImageFromArray(pt)
sitk.WriteImage(out, './constructed_dense_sax'+str(step)+'.nii')
pt = coarse_dense_sax[0, 0, :, :, :].data.cpu().numpy()
# pt = np.transpose(pt, (2, 1, 0))
out = sitk.GetImageFromArray(pt)
sitk.WriteImage(out, './constructed_coarse_sax'+str(step)+'.nii')
mm = seg_result[0, 1, :, :, :] * 1
# pt = np.transpose(mm.data.cpu().numpy(), (2, 1, 0))
out = sitk.GetImageFromArray(mm.data.cpu().numpy())
sitk.WriteImage(out, './seg_result'+str(step)+'.nii')
mm = seg_labels[0, 1, :, :, :] * 1
# pt = np.transpose(mm.data.cpu().numpy(), (2, 1, 0))
out = sitk.GetImageFromArray(mm.data.cpu().numpy())
sitk.WriteImage(out, './label'+str(step)+'.nii')
pt = sax_img[0, 0, :, :, :].data.cpu().numpy()
# pt = np.transpose(pt, (2, 1, 0))
out = sitk.GetImageFromArray(pt)
sitk.WriteImage(out, './sax'+str(step)+'.nii')
# pt = dense_sax_img[0, 0, :, :].data.cpu().numpy()
# pt = np.transpose(pt, (2, 1, 0))
# out = sitk.GetImageFromArray(pt)
# sitk.WriteImage(out, './dense_sax.nii.gz')
# pt = coarse_suoxiao[0, 0, :, :,:].data.cpu().numpy()
# # pt = np.transpose(pt, (2, 1, 0))
# out = sitk.GetImageFromArray(pt)
# sitk.WriteImage(out, './lax1.nii.gz')
pt = two_image[0, 0, :64, :,64].data.cpu().numpy()
# pt = np.transpose(pt, (2, 1, 0))
out = sitk.GetImageFromArray(pt)
sitk.WriteImage(out, './lax1'+str(step)+'.nii')