-
Notifications
You must be signed in to change notification settings - Fork 0
/
3DIVIMNET_train.py
244 lines (200 loc) · 8.95 KB
/
3DIVIMNET_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
# %%
from pytorch3dunet.unet3d.model import *
import nibabel as nib
import torch
from torch.utils.data import Dataset, DataLoader
from tqdm import tqdm
import numpy as np
from utils import prepare_path
from datetime import datetime
import os
# %%
data_files = []
bvalues_files = []
params_files = []
for i in range(1,17):
for j in range(1,3):
path = 'synthetic/{}.{}/data.nii.gz'.format(i,j)
if os.path.isfile(path):
data_files.append(path)
path = 'synthetic/{}.{}/bvalues.bval'.format(i,j)
if os.path.isfile(path):
bvalues_files.append(path)
path = 'synthetic/{}.{}/params.nii.gz'.format(i,j)
if os.path.isfile(path):
params_files.append(path)
# %%
# Make sure all files have the same sequence of bvalues
bvalues = []
for i in bvalues_files:
text_file = np.genfromtxt(i)
bvalues.append(np.array2string(text_file))
assert len(set(bvalues)) == 1
bvalues = np.array(np.genfromtxt(bvalues_files[0]))
# %%
class IVIMDataset(Dataset):
def __init__(self, data_files, bvalues, params_files, snr=None) -> None:
self.data_files = data_files
self.bvalues = bvalues
self.params_files = params_files
self.snr = snr
super().__init__()
def __len__(self):
return len(self.data_files)
#TODO: do all this preprocessing outside and save a lot of time
#TODO: alternative: load only D, f, Dp and create augmented data on-the-fly
def __getitem__(self, idx):
# load and init b-values
selsb = np.array(self.bvalues) == 0
# load nifti
data = nib.load(self.data_files[idx])
datas = data.get_fdata()
# reshape image for fitting
sx, sy, sz, n_b_values = datas.shape
X_dw = np.reshape(datas, (sx * sy * sz, n_b_values))
### select only relevant values, delete background and noise, and normalise data
S0 = np.nanmean(X_dw[:, selsb], axis=1)
S0[S0 != S0] = 0
S0 = np.squeeze(S0)
valid_id = (S0 > (0.5 * np.median(S0[S0 > 0])))
datatot = X_dw[valid_id, :]
#add noise
meansig=np.nanmean(datatot[:,selsb])
SNR = self.snr
if SNR is not None:
meansig=np.nanmean(X_dw[:,selsb])
if type(SNR) is tuple:
SNR = SNR[0]+np.random.rand(1)*(SNR[1]-SNR[0])
noise=meansig/SNR
datatot=datatot+np.random.randn(np.shape(datatot)[0],np.shape(datatot)[1])*noise
datatot[datatot<0]=0
# normalise data
S0 = np.nanmean(datatot[:, selsb], axis=1).astype('<f')
S0[S0<0.1]=0.1
datatot = datatot / S0[:, None]
datatot = np.clip(datatot,0,2.5)
data_norm = np.zeros(sx*sy*sz*n_b_values).reshape((-1, n_b_values))
data_norm[valid_id,:] = datatot
data_norm = data_norm.reshape((sx,sy,sz,n_b_values))
data_norm = data_norm.transpose(-1,0,1,2)
# load ground truth parameter maps
params = nib.load(self.params_files[idx])
params = params.get_fdata()
params = params.transpose(-1,0,1,2)
params = params[:3]
return data_norm, valid_id, params
# %%
train_data = IVIMDataset(data_files[:25], bvalues, params_files[:25], snr=(10, 30))
valid_data = IVIMDataset(data_files[25:28], bvalues, params_files[25:28], snr=10)
test_data = IVIMDataset(data_files[-2:], bvalues, params_files[-2:], snr=15)
# %%
train_dataloader = DataLoader(train_data, batch_size=5, shuffle=True)
valid_dataloader = DataLoader(valid_data, batch_size=2, shuffle=False)
test_dataloader = DataLoader(test_data, batch_size=2, shuffle=False)
# %%
n_b_values = len(bvalues)
model = UNet3D(n_b_values, 3, num_groups=1, is_segmentation=False, f_maps=(128)).to('cuda', dtype=torch.float32)
# %%
optimizer = torch.optim.Adam(model.parameters(), lr=5e-6)
cons_min = [-0.002, -0.3, -0.06] # D, f, Dp
cons_max = [0.007, 1, 0.3] # D, f, Dp
# %%
def get_snet(D, f, Dp, b):
Snet = (f * torch.exp(-Dp*b) + (1-f) * torch.exp(-D*b))
return torch.squeeze(Snet)
# %%
def predict_batch(model, x_batch, mask_batch):
pred = model.forward(x_batch)
pred = torch.sigmoid(pred)
D_ = cons_min[0] + pred[:,0,:,:,:].flatten(1)[mask_batch] * (cons_max[0] - cons_min[0])
f_ = cons_min[1] + pred[:,1,:,:,:].flatten(1)[mask_batch] * (cons_max[1] - cons_min[1])
Dp_ = cons_min[2] + pred[:,2,:,:,:].flatten(1)[mask_batch] * (cons_max[2] - cons_min[2])
# S0_ = cons_min[3] + pred[:,3,:,:,:].flatten(1)[batch_valid_id] * (cons_max[3] - cons_min[3])
return D_, f_, Dp_
def eval_batch(x_batch, D_, f_, Dp_, y_batch, mask_batch, bvalues):
criterion = torch.nn.MSELoss()
# ugly non-vectorised version:
physics_loss = torch.tensor(0).to('cuda', dtype=torch.float32)
for j, b in enumerate(bvalues):
Snet = get_snet(D_, f_, Dp_, b)
physics_loss += criterion(Snet, x_batch[:,j,:,:,:].flatten(1)[mask_batch])
physics_loss /= n_b_values
D_loss = np.sqrt(criterion(D_, y_batch[:,0,:,:,:].flatten(1)[mask_batch]).item())
f_loss = np.sqrt(criterion(f_, y_batch[:,1,:,:,:].flatten(1)[mask_batch]).item())
Dp_loss = np.sqrt(criterion(Dp_, y_batch[:,2,:,:,:].flatten(1)[mask_batch]).item())
return physics_loss, D_loss, f_loss, Dp_loss
def eval_dataloader(model, dataloader):
epoch_loss = np.zeros(4) #losses per epoch: [total, D , f, Dp]
for batch_data, batch_valid_id, batch_y in dataloader:
batch_data = batch_data.to('cuda', dtype=torch.float32)
batch_y = batch_y.to('cuda', dtype=torch.float32)
D_, f_, Dp_ = predict_batch(model, batch_data, batch_valid_id)
batch_loss, D_loss, f_loss, Dp_loss = eval_batch(batch_data, D_, f_, Dp_, batch_y, batch_valid_id, bvalues)
epoch_loss[0] += batch_loss.item()
epoch_loss[1] += D_loss
epoch_loss[2] += f_loss
epoch_loss[3] += Dp_loss
del batch_data, batch_valid_id, batch_y, D_, f_, Dp_, D_loss, f_loss, Dp_loss
epoch_loss /= len(dataloader)
return epoch_loss
# %%
def train(model, train_dataloader, valid_dataloader, optimizer, patience=10):
timestamp = datetime.now()
model.train()
epoch_n = 0
n_bad_epochs = 0
best_loss = np.inf
while n_bad_epochs < patience:
epoch_n +=1
epoch_loss = np.zeros(4) #losses per epoch: [total, D , f, Dp]
tqdm_epoch = tqdm(train_dataloader)
tqdm_epoch.set_description(desc='Epoch#{}'.format(epoch_n))
for batch_data, batch_valid_id, batch_y in tqdm_epoch:
batch_data = batch_data.to('cuda', dtype=torch.float32)
batch_y = batch_y.to('cuda', dtype=torch.float32)
D_, f_, Dp_ = predict_batch(model, batch_data, batch_valid_id)
batch_loss, D_loss, f_loss, Dp_loss = eval_batch(batch_data, D_, f_, Dp_, batch_y, batch_valid_id, bvalues)
optimizer.zero_grad()
batch_loss.backward()
optimizer.step()
epoch_loss[0] += batch_loss.item()
epoch_loss[1] += D_loss
epoch_loss[2] += f_loss
epoch_loss[3] += Dp_loss
tqdm_epoch.set_postfix({'physics': ('%.3e '% batch_loss.item()), 'D': ('%.3e '% D_loss), 'f': ('%.3e '% f_loss), 'Dp': ('%.3e '% Dp_loss)})
del batch_loss, batch_data, batch_valid_id, batch_y, D_, f_, Dp_
epoch_loss /= len(train_dataloader)
tqdm_epoch.write('Training - Physics:{:.3e}, D: {:.3e}, f: {:.3e}, Dp: {:.3e}'.format(*epoch_loss))
torch.cuda.empty_cache()
model.eval()
val_loss = eval_dataloader(model, valid_dataloader)
if val_loss[0] < best_loss:
best_loss = val_loss[0]
n_bad_epochs = 0
model_path = prepare_path('saved_models', '3DIVIMNET', timestamp)
torch.save(model.state_dict(), model_path + '/model.pt')
else:
n_bad_epochs += 1
tqdm_epoch.write('Eval (#bad: {}) - Physics:{:.3e}, D: {:.3e}, f: {:.3e}, Dp: {:.3e}'.format(n_bad_epochs, *val_loss))
return model_path + '/model.pt'
# %%
saved_model = train(model, train_dataloader, valid_dataloader, optimizer, patience=10)
# %%
model.load_state_dict(torch.load(saved_model))
model.eval()
batch_data, batch_valid_id, batch_y = next(iter(test_dataloader))
batch_data = batch_data.to('cuda', dtype=torch.float32)
batch_valid_id = batch_valid_id.to('cuda', dtype=torch.float32)
batch_y = batch_y.to('cuda', dtype=torch.float32)
pred = model.forward(batch_data)
pred = torch.sigmoid(pred)
D_ = cons_min[0] + pred[:,0] * (cons_max[0] - cons_min[0])
f_ = cons_min[1] + pred[:,1] * (cons_max[1] - cons_min[1])
Dp_ = cons_min[2] + pred[:,2] * (cons_max[2] - cons_min[2])
path = prepare_path('saved_preds',16,'3DIVIMNET')
D_valid = (D_ * batch_valid_id.reshape(D_.shape)).detach().cpu().numpy()
f_valid = (f_ * batch_valid_id.reshape(f_.shape)).detach().cpu().numpy()
Dp_valid = (Dp_ * batch_valid_id.reshape(Dp_.shape)).detach().cpu().numpy()
np.save(path + '/D.npy', D_valid)
np.save(path + '/f.npy', f_valid)
np.save(path + '/Dp.npy', Dp_valid)