-
Notifications
You must be signed in to change notification settings - Fork 0
/
TR_MT.py
316 lines (266 loc) · 12.7 KB
/
TR_MT.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
from torchvision import models
from torch.autograd import Variable
import torch.optim as optim
from os.path import exists, join
import os
import numpy as np
from sklearn.metrics import roc_auc_score
from ops import *
from glob import glob
img_w = 32 # 160 32
img_h = 32 # 192 32
num_epochs = 200 # 200
batch_size = 64 # 128
patience = num_epochs # 100
repeat_num = 5 # 10
learning_rate = 1e-3 # 1e-5
# dropout = 0.5
fold_start = 1
fold_end = 4
data_path = r'/home/public/Documents/hhy/ivim_new/data/fold/mvi2.npz'
save_path = r'/home/public/Documents/hhy/ivim_new/result_new/new_experiment/tfm_base_multi_net_share_2_4_1'
all_bmap_path = r'/home/public/Documents/hhy/ivim_new/data/mvi2'
bmap_save_path = r'/home/public/Documents/hhy/ivim_new/result_new/new_experiment/tfm_base_multi_net_share_b_2_4_1.npz'
is_save_all = True
if not exists(save_path):
os.makedirs(save_path)
# tfm_base_multi_net_1_2: sc + sf; weight of loss_rec set to 1e-2; lr set to 1e-3, decay 5e-4
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
resnet18 = torch.load('./model/resnet18-5c106cde.pth')
# define b values
b_values = np.array([0, 10, 20, 40, 80, 200, 400, 600, 1000]).astype(np.float32)
b_fit = np.expand_dims(b_values, -1) # .repeat(batch_size, axis=0)
b_fit = Variable(torch.from_numpy(b_fit).to(device))
# b_values = np.array([0, 10, 20, 40, 80, 200, 400, 600, 1000])
# b_fit = torch.FloatTensor(b_values)
print(b_values.shape)
def evaluate(pred_choice, target, p_):
TP = ((pred_choice == 1) & (target.data == 1)).cpu().sum()
TN = ((pred_choice == 0) & (target.data == 0)).cpu().sum()
FN = ((pred_choice == 0) & (target.data == 1)).cpu().sum()
FP = ((pred_choice == 1) & (target.data == 0)).cpu().sum()
# p = TP / (TP + FP + 1e-8)
# r = TP / (TP + FN + 1e-8)
# F1 = 2 * r * p / (r + p)
sen = TP / (TP + FN + 1e-8)
spe = TN / (TN + FP + 1e-8)
acc = (TP + TN) / (TP + TN + FP + FN)
pre_, lb_ = [], []
p_ = nn.Softmax(dim=1)(p_)
for ii in range(target.shape[0]):
lb_.append(target[ii].item())
pre_.append(p_[ii, 1].item())
auc = roc_auc_score(lb_, pre_)
return acc, sen, spe, auc
class Mydataset(torch.utils.data.Dataset):
def __init__(self, x, y):
self.x = x
self.y = y
self.idx = list()
for item in x:
self.idx.append(item)
pass
def __getitem__(self, index):
input_data = self.idx[index]
target = self.y[index]
return input_data, target
def __len__(self):
return len(self.idx)
def loaddata():
train_data = np.load(data_path)['train_data'].transpose([0, 3, 1, 2]).astype(np.float32)
train_lb = np.load(data_path)['train_lb'].astype(np.float32).argmax(1)
test_data = np.load(data_path)['test_data'].transpose([0, 3, 1, 2]).astype(np.float32)
test_lb = np.load(data_path)['test_lb'].astype(np.float32).argmax(1)
print(train_data.shape, ' ', test_data.shape)
print(train_lb.shape, ' ', test_lb.shape)
trainset = Mydataset(train_data, train_lb)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size, shuffle=True,
num_workers=2, drop_last=True)
testset = Mydataset(test_data, test_lb)
testloader = torch.utils.data.DataLoader(testset, batch_size=test_lb.shape[0], shuffle=False,
num_workers=2, drop_last=True)
return trainloader, testloader, train_data.shape[0] // batch_size
class TRMT(nn.Module):
def __init__(self):
super(TRMT, self).__init__()
self.model = models.resnet18(pretrained=False)
self.model.load_state_dict(resnet18)
self.model.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.model.conv1 = nn.Conv2d(9, 64, kernel_size=(7, 7), stride=(1, 1), padding=(3, 3), bias=False)
self.model.fc = nn.Linear(512, 2)
self.d1 = DownsampleLayer(9, 64) # 9-64
self.d2 = DownsampleLayer(64, 128) # 64-128
self.d3 = DownsampleLayer(128, 256) # 128-256
self.d4 = DownsampleLayer(256, 512) # 256-512
self.u1 = UpSampleLayer(512, 512) # 512-1024-512
self.u2 = UpSampleLayer(1024, 256) # 1024-512-256
self.u3 = UpSampleLayer(512, 128) # 512-256-128
self.u4 = UpSampleLayer(256, 64) # 256-128-64
self.o = nn.Sequential(
nn.Conv2d(128, 64, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(64),
nn.ReLU(),
nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(64),
nn.ReLU(),
nn.Conv2d(64, 3, 3, 1, 1),
)
self.flattener = nn.Flatten(2, 3)
self.positional_emb_c = Parameter(torch.zeros(1, 4, 512),
requires_grad=True)
init.trunc_normal_(self.positional_emb_c, std=0.2)
self.positional_emb_f = Parameter(torch.zeros(1, 4, 512),
requires_grad=True)
init.trunc_normal_(self.positional_emb_f, std=0.2)
self.drop_e = nn.Dropout(p=0.1)
# self.e1 = TransformerEncoderLayer(d_model=512, nhead=2, dim_feedforward=128)
self.ec = TransformerEncoderLayer(d_model=512, nhead=4, dim_feedforward=128)
self.ef = TransformerEncoderLayer(d_model=512, nhead=4, dim_feedforward=128)
self.e_share = TransformerEncoderLayer(d_model=512, nhead=4, dim_feedforward=128)
self.norm = LayerNorm(512)
self.attention_pool = Linear(512, 1)
# self.fc0 = nn.Linear(1024, 128)
def forward(self, inputs):
x_c = self.model.conv1(inputs)
x_c = self.model.bn1(x_c)
x_c = self.model.relu(x_c)
x_c = self.model.maxpool(x_c)
x_c = self.model.layer1(x_c)
x_c = self.model.layer2(x_c)
x_c = self.model.layer3(x_c)
x_c = self.model.layer4(x_c)
out_c = self.model.avgpool(x_c)
out_c = out_c.view(out_c.size(0), -1)
d_1, d1 = self.d1(inputs)
d_2, d2 = self.d2(d1)
d_3, d3 = self.d3(d2)
d_4, d4 = self.d4(d3)
x_c_flat = self.drop_e(self.flattener(x_c).transpose(-2, -1) + self.positional_emb_c)
d4_flat = self.drop_e(self.flattener(d4).transpose(-2, -1) + self.positional_emb_f)
sc1 = self.ec(x_c_flat, x_c_flat)
sf1 = self.ef(d4_flat, d4_flat)
sc2 = self.norm(self.e_share(sc1, sf1))
sf2 = self.norm(self.e_share(sf1, sc1)).transpose(-1, -2).view(d4.size(0), 512, 2, 2)
# s_ = torch.cat((sc, sf), -2)
# s_ = sc + sf
u1 = self.u1(d4 + sf2, d_4)
u2 = self.u2(u1, d_3)
u3 = self.u3(u2, d_2)
u4 = self.u4(u3, d_1)
x_f = self.o(u4)
mask = (inputs[:, :1] > 0).float()
params = torch.abs(x_f) * mask # 64,32,32
out_rec = torch.clamp(self.ivim_matmul(params) * inputs[:, :1], 0.0, 1.0)
sp = torch.matmul(F.softmax(self.attention_pool(sc2), dim=1).transpose(-1, -2), sc2).squeeze(-2)
# out_c = torch.cat((out_c, sp), 1)
# out_c = self.fc0(out_c)
out_c = out_c + sp
out_c = self.model.fc(out_c)
return out_c, out_rec, params
def ivim_matmul(self, params):
flat = params.view(params.size(0), 3, params.size(2) * params.size(3))
dp = flat[:, 0].unsqueeze(1)
dt = flat[:, 1].unsqueeze(1)
fp = flat[:, 2].unsqueeze(1)
b_fit_ = b_fit.unsqueeze(0).repeat(params.size(0), 1, 1)
outputs = fp * torch.exp(-torch.bmm(b_fit_, dp)) + (1 - fp) * torch.exp(-torch.bmm(b_fit_, dt))
outputs = outputs.view(params.size(0), b_values.shape[0], img_w, img_h)
# print(outputs.shape)
return outputs
def train():
trainloader, testloader, num_batch = loaddata()
test_img, test_lb = None, None
for data in testloader:
test_img, test_lb = data
test_img, test_lb = Variable(test_img.to(device)), Variable(test_lb.to(device))
model = TRMT()
model = model.to(device)
total = sum([param.nelement() for param in model.parameters()])
print('params ', total, ' flops ', 0)
# for name, parameters in model.named_parameters():
# print(name, ':', parameters.size())
# optimizer = optim.SGD(model.parameters(), lr=learning_rate, momentum=0.9, weight_decay=2e-4)
optimizer = optim.Adam(model.parameters(), lr=learning_rate, betas=(0.9, 0.999), weight_decay=2e-4)
criterion = nn.CrossEntropyLoss()
l1_loss = nn.L1Loss(reduce=True, size_average=True)
lc_, acc_ = [], []
sen_, spe_, auc_ = [], [], []
auc_max = 0
for epoch in range(num_epochs):
running_loss, running_loss_rec = 0.0, 0.0
for i, data in enumerate(trainloader, 0):
model.train()
inputs, labels = data # data是从enumerate返回的data，包含数据和标签信息，分别赋值给inputs和labels
inputs, labels = Variable(inputs.to(device)), Variable(labels.to(device))
optimizer.zero_grad() # 梯度置零，因为反向传播过程中梯度会累加上一次循环的梯度
# forward + backward + optimize
outputs, outputs_rec, ivim_pre = model(inputs)
_, predicted = torch.max(outputs, 1)
acc = torch.eq(predicted, labels).sum().float().item() / labels.shape[0]
loss_c = criterion(outputs, labels)
loss_rec = l1_loss(outputs_rec, inputs)
loss = loss_c + 1e-2 * loss_rec
loss.backward()
optimizer.step()
running_loss += loss_c.item()
running_loss_rec += loss_rec.item()
if i == num_batch - 1:
model.eval()
with torch.no_grad():
outputs_t, _, _ = model(test_img)
loss_t = criterion(outputs_t, test_lb).item()
_, predicted_t = torch.max(outputs_t, 1)
# acc_t = torch.eq(predicted_t, test_lb).sum().float().item() / test_lb.shape[0]
acc_t, sen_t, spe_t, auc_t = evaluate(predicted_t, test_lb, outputs_t)
lc_.append(loss_t), acc_.append(acc_t)
sen_.append(sen_t), spe_.append(spe_t), auc_.append(auc_t)
print('[%3d, %3d] lc %.4f lrec %.4f acc %.4f\ttest: loss %.4f acc %.4f sen %.4f spe %.4f auc %.4f' %
(epoch + 1, i + 1, running_loss / num_batch, running_loss_rec / num_batch,
acc, loss_t, acc_t, sen_t, spe_t, auc_t))
running_loss, running_loss_rec = 0.0, 0.0
print('Finished Training')
model.eval()
with torch.no_grad():
b_files = glob(join(all_bmap_path, '*'))
b_files.sort(key=lambda x: (int(x.split('/')[-1].split('.')[0].split('_')[-1])))
test_data_org = []
for idx, bf in enumerate(b_files):
img = np.load(bf)['x']
test_data_org.append(img)
test_data_org = np.array(test_data_org).transpose([0, 3, 1, 2]).astype(np.float32)
# b0 = test_data_org[..., :1]
test_data = test_data_org
dataset = Mydataset(test_data, np.ones((test_data.shape[0], 2)))
dataloader = torch.utils.data.DataLoader(dataset, batch_size=test_data.shape[0], shuffle=False,
num_workers=2, drop_last=True)
ivim_pre_all, x_fit_pre_all = [], []
for data in dataloader:
test_img, test_lb = data
test_img = Variable(test_img.to(device))
_, outputs_rec, ivim_p = model(test_img)
ivim_pre_all = ivim_p.cpu().detach().numpy().transpose([0, 2, 3, 1]).astype(np.float32)
x_fit_pre_all = outputs_rec.cpu().detach().numpy().transpose([0, 2, 3, 1]).astype(np.float32)
np.savez(bmap_save_path, ivim=ivim_pre_all, x=x_fit_pre_all)
return lc_, acc_, sen_, spe_, auc_
# print(torch.__version__)
# print(torch.cuda.device_count())
# print(torch.cuda.is_available())
lc_fold, acc_fold = [], []
sen_fold, spe_fold, auc_fold = [], [], []
temp_path = data_path
for fold_ in range(fold_start - 1, fold_end):
data_path = data_path.split('.')[0] + '_%d' % (fold_ + 1) + '.npz'
lc_all, acc_all = [], []
sen_all, spe_all, auc_all = [], [], []
for _ in range(repeat_num):
loss_c, acc_c, sen_c, spe_c, auc_c = train()
lc_all.append(loss_c), acc_all.append(acc_c)
sen_all.append(sen_c), spe_all.append(spe_c), auc_all.append(auc_c)
if fold_end - fold_start + 1 == 1:
acc_fold, lc_fold, sen_fold, spe_fold, auc_fold = acc_all, lc_all, sen_all, spe_all, auc_all
else:
lc_fold.append(lc_all), acc_fold.append(acc_all)
sen_fold.append(sen_all), spe_fold.append(spe_all), auc_fold.append(auc_all)
data_path = temp_path
if is_save_all:
np.savez(join(save_path, r'result.npz'), acc=acc_fold, loss=lc_fold, sen=sen_fold, spe=spe_fold, auc=auc_fold)