-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
245 lines (200 loc) · 8.25 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
"""Code adapted from https://github.com/arpitbansal297/Cold-Diffusion-Models"""
import argparse
import copy
import os
import torch
import json
import matplotlib.pyplot as plt
from monai.transforms import (
Compose,
LoadImaged,
AddChanneld,
RandFlipd,
ScaleIntensityRangePercentilesd,
RandRotate90d,
ToTensord,
RandSpatialCropSamplesd,
)
from monai.data import (
DataLoader, Dataset
)
from model import GaussianDiffusion, TimeEmbUNet3D
def cycle_dl(dl):
while True:
for data in dl:
yield data
class EMA:
def __init__(self, beta):
super().__init__()
self.beta = beta
def update_model_average(self, ma_model, current_model):
for current_params, ma_params in zip(current_model.parameters(), ma_model.parameters()):
old_weight, up_weight = ma_params.data, current_params.data
ma_params.data = self.update_average(old_weight, up_weight)
def update_average(self, old, new):
if old is None:
return new
return old * self.beta + (1 - self.beta) * new
def save_images(noisy_img, recon_img, train_step, val_idx, base_path):
# save cross-sectional images to base path with step
_, _, d, h, w = noisy_img.shape
f, axs = plt.subplots(2, 3, figsize=(8, 5))
# visualize d axis (hw plane)
axs[0, 0].imshow(noisy_img[0, 0, d // 2, :, :])
axs[0, 0].set_title('Noisy Image (2-3 plane)')
axs[1, 0].imshow(recon_img[0, 0, d // 2, :, :])
axs[1, 0].set_title('Reconstructed Image (2-3 plane)')
# visualize h axis (dw plane)
axs[0, 1].imshow(noisy_img[0, 0, :, h // 2, :])
axs[0, 1].set_title('Noisy Image (1-3 plane)')
axs[1, 1].imshow(recon_img[0, 0, :, h // 2, :])
axs[1, 1].set_title('Reconstructed Image (1-3 plane)')
# visualize w axis (dh plane)
axs[0, 2].imshow(noisy_img[0, 0, :, :, w // 2])
axs[0, 2].set_title('Noisy Image (1-2 plane)')
axs[1, 2].imshow(recon_img[0, 0, :, :, w // 2])
axs[1, 2].set_title('Reconstructed Image (1-2 plane)')
plt.savefig(os.path.join(base_path, f'img_{val_idx}_train_step_{train_step}.png'))
def train(args):
# get json datalists
with open(args.train_json, 'r') as f:
train_data = json.load(f)
with open(args.val_json, 'r') as f:
val_data = json.load(f)
# remove train item with only 28px in an axis...
train_data = [dat for dat in train_data if 'IXI014-HH-1236-T2.nii.gz' not in dat['image']]
# make output dir
if not os.path.exists(args.log_path):
os.makedirs(args.log_path)
print(f'train data length: {len(train_data)}')
print(f'val data length: {len(val_data)}')
# define train and val transforms
train_transforms = Compose([
LoadImaged(keys=['image']),
AddChanneld(keys=['image']),
ScaleIntensityRangePercentilesd(keys=['image'], lower=0.05, upper=99.95, b_min=-1, b_max=1),
RandSpatialCropSamplesd(keys=['image'], roi_size=(96, 96, 96), num_samples=4, random_size=False),
RandFlipd(keys=['image'], spatial_axis=[0], prob=0.1),
RandFlipd(keys=['image'], spatial_axis=[1], prob=0.1),
RandFlipd(keys=['image'], spatial_axis=[2], prob=0.1),
RandRotate90d(keys=['image'], prob=0.1, spatial_axes=(0, 1)),
ToTensord(keys=['image'])
])
val_transforms = Compose([
LoadImaged(keys=['image']),
AddChanneld(keys=['image']),
ScaleIntensityRangePercentilesd(keys=['image'], lower=0.05, upper=99.95, b_min=-1, b_max=1),
ToTensord(keys=['image'])
])
# datasets and dataloaders
train_ds = Dataset(data=train_data, transform=train_transforms)
train_loader = DataLoader(train_ds, batch_size=args.batch_size, shuffle=True, num_workers=16, pin_memory=True)
val_ds = Dataset(data=val_data, transform=val_transforms)
val_loader = DataLoader(val_ds, batch_size=1, shuffle=False, num_workers=16, pin_memory=True)
# cycle train dataloader
train_loader = cycle_dl(train_loader)
# model params mimic those in original code
model = TimeEmbUNet3D(
spatial_dims=3,
in_channels=1,
out_channels=1,
channels=(64, 128, 256, 512),
num_res_units=2,
strides=(2, 2, 2)
).cuda()
diffusion_model = GaussianDiffusion(
model,
image_size=96,
channels=1,
timesteps=args.timesteps,
loss_type='l1',
batch_size=args.batch_size
).cuda()
# TODO: dataparallel here if needed
# ema model
ema = EMA(args.ema_decay)
ema_model = copy.deepcopy(diffusion_model)
ema_model.load_state_dict(diffusion_model.state_dict())
ema_model.eval()
# params for ema and training
update_ema_every = 10
grad_accum_every = 2
val_every = 2000
step_start_ema = 2000
# optimizer
optimizer = torch.optim.Adam(diffusion_model.parameters(), lr=args.lr)
# training loop
accum_loss = 0
step = 0
best_val_loss = 1000000
best_val_step = 0
while step < args.train_steps:
denoise_loss = 0
# gradient accumulation
for i in range(grad_accum_every):
data_1 = next(train_loader)['image']
data_2 = torch.randn_like(data_1)
data_1, data_2 = data_1.cuda(), data_2.cuda()
loss = diffusion_model(data_1, data_2)
if step % 100 == 0:
print(f'step: {step}, loss: {loss.item()}')
denoise_loss += loss.item()
loss = loss / grad_accum_every
loss.backward()
accum_loss += denoise_loss / grad_accum_every
optimizer.step()
optimizer.zero_grad()
# update ema model
if step % update_ema_every == 0:
if step < step_start_ema:
ema_model.load_state_dict(diffusion_model.state_dict())
else:
ema.update_model_average(ema_model, diffusion_model)
# validation
if step % val_every == 0 and step != 0:
print(f'validation at step: {step}')
ema_model.eval()
val_denoise_loss = 0
for val_idx, val_data in enumerate(val_loader):
data_1 = val_data['image']
data_2 = torch.randn_like(data_1)
data_1, data_2 = data_1.cuda(), data_2.cuda()
# visualize reconstructions for the first 3 images in val
if val_idx < 3:
loss, noisy_img, recon_img = ema_model.validate(data_1, data_2, return_imgs=True)
noisy_img = noisy_img.detach().cpu().numpy()
recon_img = recon_img.detach().cpu().numpy()
save_images(noisy_img, recon_img, step, val_idx, args.log_path)
else:
loss = ema_model.validate(data_1, data_2)
val_denoise_loss += loss.item()
val_denoise_loss = val_denoise_loss / len(val_loader)
print(f'validation loss: {val_denoise_loss}')
print(f'mean train loss: {accum_loss / val_every}')
accum_loss = 0
# save best model based on validation reconstruction loss
if val_denoise_loss < best_val_loss:
best_val_loss = val_denoise_loss
best_val_step = step
torch.save(ema_model.state_dict(), os.path.join(args.log_path, 'best_model.pth'))
# save model 10 times during training
if (step + 1) % (args.train_steps // 10) == 0:
torch.save(ema_model.state_dict(), os.path.join(args.log_path, f'model_step_{step}.pth'))
step += 1
print(f'best validation loss: {best_val_loss} at step: {best_val_step}')
print('done training!')
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--train_json', type=str)
parser.add_argument('--val_json', type=str)
parser.add_argument('--test_json', type=str)
parser.add_argument('--log_path', type=str)
parser.add_argument('--batch_size', default=4, type=int)
parser.add_argument('--ema_decay', default=0.995, type=float)
parser.add_argument('--train_steps', default=700000, type=int)
parser.add_argument('--timesteps', default=100, type=int)
parser.add_argument('--lr', default=0.00002, type=float)
args = parser.parse_args()
print('===== arguments for this run =====')
print(vars(args))
train(args)