Create models.py (#3)
Browse files- Create models.py (0ff329ac3909664a717a4b4aeb09baac94f4143f)
models.py
ADDED
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torchvision.transforms as transforms
|
4 |
+
import timm
|
5 |
+
import PIL.Image as Image
|
6 |
+
|
7 |
+
class ViTClassifier(nn.Module):
|
8 |
+
def __init__(self, config, device='cuda', dtype=torch.float32):
|
9 |
+
super(ViTClassifier, self).__init__()
|
10 |
+
self.config = config
|
11 |
+
self.device = device
|
12 |
+
self.dtype = dtype
|
13 |
+
|
14 |
+
# Create the ViT model without unsupported arguments
|
15 |
+
self.vit = timm.create_model(
|
16 |
+
config['model']['variant'],
|
17 |
+
pretrained=False,
|
18 |
+
num_classes=config['model']['num_classes'],
|
19 |
+
drop_rate=config['model']['hidden_dropout_prob'],
|
20 |
+
attn_drop_rate=config['model']['attention_probs_dropout_prob']
|
21 |
+
).to(device)
|
22 |
+
|
23 |
+
# Replace the head with a custom head
|
24 |
+
self.vit.head = nn.Linear(
|
25 |
+
in_features=config['model']['head']['in_features'],
|
26 |
+
out_features=config['model']['head']['out_features'],
|
27 |
+
bias=config['model']['head']['bias'],
|
28 |
+
device=device,
|
29 |
+
dtype=dtype
|
30 |
+
)
|
31 |
+
|
32 |
+
if config['model']['freeze_backbone']:
|
33 |
+
for param in self.vit.parameters():
|
34 |
+
param.requires_grad = False
|
35 |
+
|
36 |
+
for param in self.vit.head.parameters():
|
37 |
+
assert param.requires_grad == True, "Model head should be trainable."
|
38 |
+
|
39 |
+
def preprocess_input(self, x):
|
40 |
+
norm_mean = self.config['preprocessing']['norm_mean']
|
41 |
+
norm_std = self.config['preprocessing']['norm_std']
|
42 |
+
resize_size = self.config['preprocessing']['resize_size']
|
43 |
+
crop_size = self.config['preprocessing']['crop_size']
|
44 |
+
|
45 |
+
augment_list = [
|
46 |
+
transforms.Resize(resize_size),
|
47 |
+
transforms.CenterCrop(crop_size),
|
48 |
+
transforms.ToTensor(),
|
49 |
+
transforms.Normalize(mean=norm_mean, std=norm_std),
|
50 |
+
transforms.ConvertImageDtype(self.dtype),
|
51 |
+
]
|
52 |
+
|
53 |
+
preprocess = transforms.Compose(augment_list)
|
54 |
+
x = preprocess(x)
|
55 |
+
x = x.unsqueeze(0)
|
56 |
+
return x
|
57 |
+
|
58 |
+
def forward(self, x):
|
59 |
+
x = self.preprocess_input(x).to(self.device)
|
60 |
+
x = self.vit(x)
|
61 |
+
x = torch.nn.functional.sigmoid(x)
|
62 |
+
return x
|