-
Notifications
You must be signed in to change notification settings - Fork 138
/
train_one_gpu.py
180 lines (158 loc) · 6.24 KB
/
train_one_gpu.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
import os
import cv2
import numpy as np
import torch
from torchmetrics import Dice
from PIL import Image
from torch.nn import Conv2d, LayerNorm, ReLU, Upsample
from torch.optim import AdamW
from torch.optim.lr_scheduler import LinearLR
from torch.utils.data.dataloader import DataLoader
from vit import load_model_and_transform
os.makedirs("output", exist_ok=True)
def main():
my_model, transform = load_model_and_transform("ViT-B/32")
my_model.cuda().train()
my_model: torch.nn.Module
if os.path.exists("FP16-ViT-B-32.pt"):
my_model.load_state_dict(torch.load("FP16-ViT-B-32.pt", "cpu"), strict=True)
my_dataset = MyDataset(transform)
my_dataloader = DataLoader(my_dataset, 32, True, num_workers=8, drop_last=True)
my_model_seg = MyModelSeg().cuda().train()
my_optimizer = AdamW([
{"params": my_model_seg.parameters(), "lr": 0.0005},
{"params": my_model.parameters(), "lr": 0.00001},
], weight_decay=0.1)
my_linear_lr = LinearLR(my_optimizer, start_factor=1, end_factor=0, total_iters=20000)
my_loss = torch.nn.CrossEntropyLoss().cuda()
step = 0
while True:
for image, label in my_dataloader:
image = image.cuda()
label = label.cuda()
predict = my_model_seg(my_model(image))
loss = my_loss(predict.reshape(-1, 5), label.reshape(-1))
loss.backward()
my_optimizer.step()
my_linear_lr.step()
my_optimizer.zero_grad()
step += 1
if step % 100 == 0:
output, metric = evaluate(my_model, my_model_seg)
output = output.cpu().numpy()
color = np.ones([output.shape[0], output.shape[1], 3])
color[output==0] = [255, 255, 255] #其他，白色，0
color[output==1] = [0, 255, 0] #植被，绿色，1
color[output==2] = [0, 0, 0] #道路，黑色，2
color[output==3] = [131, 139, 139] #建筑，黄色，3
color[output==4] = [139, 69, 19] #水体，蓝色，4
cv2.imwrite(f"output/{step :07d}.jpg", color)
with torch.no_grad():
print(f"step: {step :07d} loss: {loss.item() :.4f} dice: {metric.item() :.4f}")
if step > 20000:
torch.save(my_model_seg.state_dict(), "model.pt")
break
class MyDataset(torch.utils.data.Dataset):
def __init__(self, transform):
if os.path.exists("dataset/path_list.csv"):
with open("dataset/path_list.csv", "r") as f:
lines = f.readlines()
lines = [x.strip() for x in lines]
lines = lines[1:]
self.lines = lines
else:
raise "Please run <python preprocess.py first!>"
self.transform = transform
def __getitem__(self, index):
path_image, path_label = self.lines[index].split(",")
image = self.transform(Image.open(path_image))
return image, torch.from_numpy(cv2.imread(path_label, cv2.IMREAD_GRAYSCALE))
def __len__(self):
return len(self.lines)
class MyModelSeg(torch.nn.Module):
def __init__(self,) -> None:
super().__init__()
self.layers = torch.nn.Sequential(
# B x 768 x 7 x 7
Upsample(scale_factor=2),
Conv2d(768, 768, (1, 1), 1, 0, bias=False),
ReLU(),
Conv2d(768, 768, (3, 3), 1, (1, 1), bias=False),
LayerNorm([768, 14, 14]),
ReLU(),
# B x 768 x 14 x 14
Upsample(scale_factor=2),
Conv2d(768, 384, (1, 1), 1, 0, bias=False),
ReLU(),
Conv2d(384, 384, (3, 3), 1, (1, 1), bias=False),
LayerNorm([384, 28, 28]),
ReLU(),
# B x 384 x 28 x 28
Upsample(scale_factor=2),
Conv2d(384, 192, (1, 1), 1, 0, bias=False),
ReLU(),
Conv2d(192, 192, (3, 3), 1, (1, 1), bias=False),
LayerNorm([192, 56, 56]),
ReLU(),
# B x 192 x 56 x 56
Upsample(scale_factor=2),
Conv2d(192, 96, (1, 1), 1, 0, bias=False),
ReLU(),
Conv2d(96, 96, (3, 3), 1, (1, 1), bias=False),
LayerNorm([96, 112, 112]),
ReLU(),
# B x 96 x 112 x 112
Upsample(scale_factor=2),
Conv2d(96, 48, (1, 1), 1, 0, bias=False),
ReLU(),
Conv2d(48, 48, (3, 3), 1, (1, 1), bias=False),
LayerNorm([48, 224, 224]),
ReLU(),
# B x 48 x 224 x 224
Conv2d(48, 5, (1, 1), 1, 0, bias=False))
def forward(self, x):
B, S, D = x.size()
x = torch.reshape(x, (B, 7, 7, 768))
x: torch.Tensor
x = x.permute(0, 3, 1, 2)
x = self.layers(x)
x = x.permute(0, 2, 3, 1)
return x
def evaluate(model, model_seg, image_size=224):
my_dice = Dice().cuda()
model.eval()
model_seg.eval()
image = cv2.imread("dataset/origin/5.png")
gt = torch.from_numpy(cv2.imread("dataset/origin/5_class.png", cv2.IMREAD_GRAYSCALE))
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
h, w, c = image.shape
image = torch.from_numpy(image).float()
image = image.permute(2, 0, 1).unsqueeze(0)
output = torch.zeros([h, w]).cuda().long()
idx_h = 0
while idx_h * image_size < h:
h_e = min(h, idx_h * image_size + image_size)
h_s = h_e - image_size
idx_w = 0
while idx_w * image_size < w:
w_e = min(w, idx_w * image_size + image_size)
w_s = w_e - image_size
idx_w += 1
img = image[:, :, h_s: h_s + image_size, w_s: w_s + image_size].cuda()
img = img / 255
img[:, 0, :, :] -= 0.485
img[:, 1, :, :] -= 0.456
img[:, 2, :, :] -= 0.406
img[:, 0, :, :] /= 0.229
img[:, 1, :, :] /= 0.224
img[:, 2, :, :] /= 0.225
predict = model_seg(model(img))
predict = torch.argmax(predict, dim=3).squeeze()
output[h_s: h_s + image_size, w_s: w_s + image_size] = predict
idx_h += 1
model.train()
model_seg.train()
metric = my_dice(output, gt.cuda())
return output, metric
if __name__ == "__main__":
main()