-
Notifications
You must be signed in to change notification settings - Fork 20
/
train.py
192 lines (149 loc) · 7.31 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
import numpy as np
from dataset.btcv_transunet_datasetings import get_loader_btcv
import torch
import torch.nn as nn
from monai.inferers import SlidingWindowInferer
from light_training.evaluation.metric import dice, hausdorff_distance_95
from light_training.trainer import Trainer
from monai.utils import set_determinism
from light_training.utils.lr_scheduler import LinearWarmupCosineAnnealingLR
from light_training.utils.files_helper import save_new_model_and_delete_last
import argparse
from monai.losses.dice import DiceLoss
import yaml
from unet.basic_unet import BasicUNetEncoder
from unet.basic_unet_denose import BasicUNetDe
from guided_diffusion.gaussian_diffusion import get_named_beta_schedule, ModelMeanType, ModelVarType,LossType
from guided_diffusion.respace import SpacedDiffusion, space_timesteps
from guided_diffusion.resample import UniformSampler
set_determinism(123)
import os
data_dir = "./RawData/Training/"
logdir = "./logs_btcv/diffunet_transunet_datasettings/"
model_save_path = os.path.join(logdir, "model")
max_epoch = 3000
batch_size = 1
val_every = 100
env = "DDP"
num_gpus = 4
# or
# env = "pytorch"
# num_gpus = 1
device = "cuda:0"
class DiffUNet(nn.Module):
def __init__(self) -> None:
super().__init__()
self.embed_model = BasicUNetEncoder(3, 1, 2, [64, 64, 128, 256, 512, 64])
self.model = BasicUNetDe(3, 14, 13, [64, 64, 128, 256, 512, 64],
act = ("LeakyReLU", {"negative_slope": 0.1, "inplace": False}))
betas = get_named_beta_schedule("linear", 1000)
self.diffusion = SpacedDiffusion(use_timesteps=space_timesteps(1000, [1000]),
betas=betas,
model_mean_type=ModelMeanType.START_X,
model_var_type=ModelVarType.FIXED_LARGE,
loss_type=LossType.MSE,
)
self.sample_diffusion = SpacedDiffusion(use_timesteps=space_timesteps(1000, [10]),
betas=betas,
model_mean_type=ModelMeanType.START_X,
model_var_type=ModelVarType.FIXED_LARGE,
loss_type=LossType.MSE,
)
self.sampler = UniformSampler(1000)
def forward(self, image=None, x=None, pred_type=None, step=None, embedding=None):
if pred_type == "q_sample":
noise = torch.randn_like(x).to(x.device)
t, weight = self.sampler.sample(x.shape[0], x.device)
return self.diffusion.q_sample(x, t, noise=noise), t, noise
elif pred_type == "denoise":
embeddings = self.embed_model(image)
return self.model(x, t=step, image=image, embeddings=embeddings)
elif pred_type == "ddim_sample":
embeddings = self.embed_model(image)
sample_out = self.sample_diffusion.ddim_sample_loop(self.model, (1, 13, 96, 96, 96), model_kwargs={"image": image, "embeddings": embeddings})
sample_out = sample_out["pred_xstart"]
return sample_out
class BraTSTrainer(Trainer):
def __init__(self, env_type, max_epochs, batch_size, device="cpu", val_every=1, num_gpus=1, logdir="./logs/", master_ip='localhost', master_port=17750, training_script="train.py"):
super().__init__(env_type, max_epochs, batch_size, device, val_every, num_gpus, logdir, master_ip, master_port, training_script)
self.window_infer = SlidingWindowInferer(roi_size=[96, 96, 96],
sw_batch_size=1,
overlap=0.5)
self.model = DiffUNet()
self.best_mean_dice = 0.0
self.optimizer = torch.optim.AdamW(self.model.parameters(), lr=2e-4, weight_decay=1e-3)
self.ce = nn.CrossEntropyLoss()
self.mse = nn.MSELoss()
self.scheduler = LinearWarmupCosineAnnealingLR(self.optimizer,
warmup_epochs=100,
max_epochs=max_epochs)
self.bce = nn.BCEWithLogitsLoss()
self.dice_loss = DiceLoss(sigmoid=True)
def training_step(self, batch):
image, label = self.get_input(batch)
x_start = label
x_start = (x_start) * 2 - 1
x_t, t, noise = self.model(x=x_start, pred_type="q_sample")
pred_xstart = self.model(x=x_t, step=t, image=image, pred_type="denoise")
loss_dice = self.dice_loss(pred_xstart, label)
loss_bce = self.bce(pred_xstart, label)
pred_xstart = torch.sigmoid(pred_xstart)
loss_mse = self.mse(pred_xstart, label)
loss = loss_dice + loss_bce + loss_mse
self.log("train_loss", loss, step=self.global_step)
return loss
def get_input(self, batch):
image = batch["image"]
label = batch["label"]
label = self.convert_labels(label)
label = label.float()
return image, label
def convert_labels(self, labels):
labels_new = []
for i in range(1, 14):
labels_new.append(labels == i)
labels_new = torch.cat(labels_new, dim=1)
return labels_new
def validation_end(self, mean_val_outputs):
dices = mean_val_outputs
print(dices)
mean_dice = sum(dices) / len(dices)
self.log("mean_dice", mean_dice, step=self.epoch)
if mean_dice > self.best_mean_dice:
self.best_mean_dice = mean_dice
save_new_model_and_delete_last(self.model,
os.path.join(model_save_path,
f"best_model_{mean_dice:.4f}.pt"),
delete_symbol="best_model")
save_new_model_and_delete_last(self.model,
os.path.join(model_save_path,
f"final_model_{mean_dice:.4f}.pt"),
delete_symbol="final_model")
print(f" mean_dice is {mean_dice}")
def validation_step(self, batch):
image, label = self.get_input(batch)
output = self.window_infer(image, self.model, pred_type="ddim_sample")
output = torch.sigmoid(output)
output = (output > 0.5).float().cpu().numpy()
target = label.cpu().numpy()
dices = []
hd = []
c = 13
for i in range(0, c):
pred_c = output[:, i]
target_c = target[:, i]
dices.append(dice(pred_c, target_c))
hd.append(hausdorff_distance_95(pred_c, target_c))
return dices
if __name__ == "__main__":
trainer = BraTSTrainer(env_type=env,
max_epochs=max_epoch,
batch_size=batch_size,
device=device,
logdir=logdir,
val_every=val_every,
num_gpus=num_gpus,
master_port=17751,
training_script=__file__)
train_ds, val_ds, test_ds = get_loader_btcv(data_dir=data_dir)
trainer.train(train_dataset=train_ds, val_dataset=val_ds)