-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathmain_gan_vit.py
144 lines (125 loc) · 7.1 KB
/
main_gan_vit.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
import sys; sys.path.append('./')
import argparse
from accelerate import Accelerator
from tqdm import tqdm
from pytorch3dunet.unet3d.model import Residual_mid_UNet3D_vit
from generative.networks.nets import PatchDiscriminator
from generative.losses import PerceptualLoss, PatchAdversarialLoss
from utils.common import load_config, copy_yaml_to_folder_auto, get_parameters, plt_mri_pet, see_mri_pet
from dataloader.threed_loader import form_dataloader
import torch
from accelerate.utils import DistributedDataParallelKwargs
from torch.nn import functional as F
from os.path import join as j
import os
from torchvision.utils import save_image
def main(args):
cf = load_config(args.config_path)
if not cf['is_debug']:
dir = copy_yaml_to_folder_auto(args.config_path, cf['project_dir'])
cf['project_dir'] = dir
kwargs = DistributedDataParallelKwargs(broadcast_buffers=False)
train_dataloader = form_dataloader(cf['train_path'],
cf['img_sz'],
cf['train_bc'],
True)
val_dataloader = form_dataloader(cf['eval_path'],
cf['img_sz'],
cf['eval_bc'],
False)
accelerator = Accelerator(**get_parameters(Accelerator, cf), kwargs_handlers=[kwargs])
model = Residual_mid_UNet3D_vit(1, 1, is_segmentation=False, f_maps=(64, 128, 256))
discriminator = PatchDiscriminator(
spatial_dims=3,
num_layers_d=3,
num_channels=32,
in_channels=1,
out_channels=1,
kernel_size=4,
padding=1,)
perceptual_loss = PerceptualLoss(spatial_dims=3, network_type="squeeze", fake_3d_ratio=0.25).to(accelerator.device)
adv_loss = PatchAdversarialLoss(criterion="least_squares")
adv_weight = 0.01
perceptual_weight = 0.001
optimizer_g = torch.optim.Adam(model.parameters(), 1e-4)
optimizer_d = torch.optim.Adam(discriminator.parameters(), lr=5e-4)
val_interval = cf['val_inter']
save_interval = cf['save_inter']
autoencoder_warm_up_n_epochs = 10
if len(cf['log_with']):
accelerator.init_trackers('train_example')
model, discriminator, optimizer_g, optimizer_d, train_dataloader, val_dataloader = accelerator.prepare(
model, discriminator, optimizer_g, optimizer_d, train_dataloader, val_dataloader
)
global_step = 0
for epoch in range(cf['num_epochs']):
progress_bar = tqdm(total=len(train_dataloader), disable=not accelerator.is_local_main_process)
progress_bar.set_description(f"Epoch {epoch+1}")
for step, batch in enumerate(train_dataloader):
condition, target = batch['image'], batch['label']
if condition.shape[1] != 1:
print("Wrong! Got the first channel!", batch['name'])
condition = condition[:,:1,...]
with accelerator.accumulate(model):
reconstruction = model(condition)
recons_loss = F.l1_loss(reconstruction.float(), target.float())
p_loss = perceptual_loss(reconstruction.float(), target.float())
loss_g = recons_loss + (perceptual_weight * p_loss)
if epoch+1 > autoencoder_warm_up_n_epochs:
logits_fake = discriminator(reconstruction.contiguous())[-1]
generator_loss = adv_loss(logits_fake, target_is_real=True, for_discriminator=False)
loss_g += adv_weight * generator_loss
accelerator.backward(loss_g)
accelerator.clip_grad_norm_(model.parameters(), 1.0)
optimizer_g.step()
optimizer_g.zero_grad()
if epoch+1 > autoencoder_warm_up_n_epochs:
with accelerator.accumulate(discriminator):
with torch.no_grad():
recon = model(condition)
logits_fake = discriminator(recon.contiguous().detach())[-1]
loss_d_fake = adv_loss(logits_fake, target_is_real=False, for_discriminator=True)
logits_real = discriminator(target.contiguous().detach())[-1]
loss_d_real = adv_loss(logits_real, target_is_real=True, for_discriminator=True)
discriminator_loss = (loss_d_fake + loss_d_real) * 0.5
loss_d = adv_weight * discriminator_loss
accelerator.backward(loss_d)
optimizer_d.step()
optimizer_d.zero_grad()
progress_bar.update(1)
logs = {"g_loss": loss_g.detach().item()}
progress_bar.set_postfix(**logs)
accelerator.log(logs, step=global_step)
global_step += 1
if accelerator.is_main_process:
val_model = accelerator.unwrap_model(model)
if (epoch + 1) % val_interval == 0 or epoch == cf['num_epochs'] - 1:
for i, batch in enumerate(val_dataloader):
with torch.no_grad():
condition, target = batch['image'], batch['label']
val_recon = val_model(condition)
images = torch.cat([condition, target, val_recon], dim=-2)
save_pic_dir = j(cf['project_dir'], f'results_save/{epoch+1}')
os.makedirs(save_pic_dir, exist_ok=True)
save_image(see_mri_pet(images), j(save_pic_dir, f'{i+1}.png'))
# with torch.no_grad():
# batch = next(iter(val_dataloader))
# val_recon = val_model(batch["image"])
# image = batch['image'][:, 0, ...]
# label = batch["label"][:, 0, ...]
# val_recon = val_recon[:, 0, ...]
# val_recon = torch.cat([image, label, val_recon], dim=-2)
# for i in range(len(val_recon)):
# tensor = val_recon[i]
# save_pic_dir = j(cf['project_dir'], f'results_save/{epoch+1}/{i}.png')
# os.makedirs(os.path.dirname(save_pic_dir), exist_ok=True)
# plt_mri_pet(tensor.cpu(), save_pic_dir)
if (epoch + 1) % save_interval == 0 or epoch == cf['num_epochs'] - 1:
model_save_dir = j(dir, 'model_save/model.pt')
os.makedirs(os.path.dirname(model_save_dir), exist_ok=True)
torch.save(val_model.state_dict(), model_save_dir)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--config_path', type=str, default='config/main_gan_vit_config.yaml')
args = parser.parse_args()
main(args)