-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
191 lines (155 loc) · 7.18 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
import numpy as np
from dataset.brats_data_utils import get_loader_brats
import torch
import torch.nn as nn
from monai.networks.nets.basic_unet import BasicUNet
from monai.networks.nets.unetr import UNETR
from monai.networks.nets.swin_unetr import SwinUNETR
from monai.inferers import SlidingWindowInferer
from light_training.evaluation.metric import dice
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
from models.uent2d import UNet2D
from models.uent3d import UNet3D
from monai.networks.nets.segresnet import SegResNet
from models.transbts.TransBTS_downsample8x_skipconnection import TransBTS
from models.nestedformer.nested_former import NestedFormer
from models.swinunet2d.swinunet import SwinUnet
from einops import rearrange
from monai.networks.nets.vnet import VNet
from models.modelgenesis.unet3d import UNet3DModelGen
from models.transvw.models.ynet3d import UNet3DTransVW
from monai.networks.nets.attentionunet import AttentionUnet
from models.unet_nested.unet_nested_3d import UNet_Nested3D
set_determinism(123)
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "0,2,3"
data_dir = "/home/xingzhaohu/sharefs/datasets/brats2020/MICCAI_BraTS2020_TrainingData/"
# logdir = "./logs_brats/swinunetr_multi_gpu/"
# logdir = "./logs_brats/unet2d/"
# logdir = "./logs_brats/unet3d/"
# logdir = "./logs_brats/unetr/"
# logdir = "./logs_brats/tranbts/"
# logdir = "./logs_brats/segresnet/"
# logdir = "./logs_brats/transvw"
# logdir = "./logs_brats/swinunet2d"
# logdir = "./logs_brats/vnet"
# logdir = "./logs_brats/modelsgenesis"
# logdir = "./logs_brats/transvw"
# logdir = "./logs_brats/attentionUNet"
logdir = "./logs_brats/unet_plus"
model_save_path = os.path.join(logdir, "model")
max_epoch = 300
batch_size = 1
val_every = 10
num_gpus = 2
device = "cuda:0"
class SwinUNETR2D(nn.Module):
def __init__(self) -> None:
super().__init__()
self.model = SwinUNETR([96, 96], 4, 4, spatial_dims=2)
def forward(self, x):
b, c, d, w, h = x.shape
x = rearrange(x, "b c d w h -> (b d) c w h")
# print(x.shape)
x = self.model(x)
x = rearrange(x, "(b d) c w h -> b c d w h", b=b, d=d)
return x
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=2,
overlap=0.25)
# self.model = SwinUNETR([96, 96, 96], 4, 4)
# self.model = UNet2D()
# self.model = UNet3D()
# self.model = UNETR(4, 4, [96, 96, 96])
# _, model = TransBTS(dataset='brats', _conv_repr=True, _pe_type="learned")
# self.model = model
# self.model = SegResNet(3, 16, 4, 4)
from models.swinunet2d.config import get_config
config = get_config()
self.model = SwinUnet(config, img_size=96, in_channels=4, num_classes=4)
# self.model = SwinUNETR2D()
# self.model = VNet(3, 4, 4, bias=True)
# self.model = UNet3DModelGen(4)
# self.model = UNet3DTransVW(4)
# weight_dir = "/home/xingzhaohu/jiuding_code/mutual_learning/logs_brats/Genesis_Chest_CT.pt"
# weight_dir = "/home/xingzhaohu/jiuding_code/mutual_learning/logs_brats/TransVW_chest_ct.pt"
# checkpoint = torch.load(weight_dir)
# state_dict = checkpoint['state_dict']
# unParalled_state_dict = {}
# for key in state_dict.keys():
# if "down_tr64.ops.0.conv1.weight" in key:
# state_dict[key] = state_dict[key].repeat(1, 4, 1, 1, 1)
# if "out_tr" not in key:
# unParalled_state_dict[key.replace("module.", "")] = state_dict[key]
# self.model.load_state_dict(unParalled_state_dict, strict=False)
# self.model = AttentionUnet(3, 4, 4, channels=[32, 64, 128, 256], strides=[2, 2, 2, 2])
self.model = UNet_Nested3D(in_channels=4, n_classes=4)
self.best_mean_dice = 0.0
self.optimizer = torch.optim.AdamW(self.model.parameters(), lr=1e-4, weight_decay=1e-3)
self.loss_func = nn.CrossEntropyLoss()
def training_step(self, batch):
import time
image, label = self.get_input(batch)
pred = self.model(image)
loss = self.loss_func(pred, label)
self.log("train_loss", loss, step=self.global_step)
return loss
# for image, label in data_loader:
def get_input(self, batch):
image = batch["image"]
label = batch["label"]
label[label == 4] = 3
if len(label.shape) == 5:
label = label[:, 0]
label = label.long()
return image, label
def validation_step(self, batch):
image, label = self.get_input(batch)
output = self.window_infer(image, self.model).argmax(dim=1).cpu().numpy()
target = label.cpu().numpy()
o = output > 0; t = target > 0 # ce
wt = dice(o, t)
# core
o = (output == 1) | (output == 3)
t = (target == 1) | (target == 3)
tc = dice(o, t)
# active
o = (output == 3);t = (target == 3)
et = dice(o, t)
return [wt, tc, et]
def validation_end(self, mean_val_outputs):
wt, tc, et = mean_val_outputs
self.log("wt", wt, step=self.epoch)
self.log("tc", tc, step=self.epoch)
self.log("et", et, step=self.epoch)
self.log("mean_dice", (wt+tc+et)/3, step=self.epoch)
mean_dice = (wt + tc + et) / 3
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"wt is {wt}, tc is {tc}, et is {et}, mean_dice is {mean_dice}")
if __name__ == "__main__":
train_ds, val_ds, test_ds = get_loader_brats(data_dir=data_dir, batch_size=batch_size, fold=0)
trainer = BraTSTrainer(env_type="DDP",
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__)
trainer.train(train_dataset=train_ds, val_dataset=val_ds)