Upload README.md
Browse files
README.md
ADDED
@@ -0,0 +1,79 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Classifier for Selecting Pathology Images
|
2 |
+
|
3 |
+
|
4 |
+
|
5 |
+
This is a ConvNext-tiny model trained on 30K annotations on if image is belongs to the pathology image or non-pathology image.
|
6 |
+
|
7 |
+
## Usage
|
8 |
+
|
9 |
+
> #### Step1: Download model checkpoint in [convnext-pathology-classifier](https://huggingface.co/jamessyx/convnext-pathology-classifier) .
|
10 |
+
|
11 |
+
|
12 |
+
|
13 |
+
> #### Step2: Load the model
|
14 |
+
|
15 |
+
You can use the following code to load the model.
|
16 |
+
|
17 |
+
```python
|
18 |
+
import timm ##timm version 0.9.7
|
19 |
+
import torch.nn as nn
|
20 |
+
import torch
|
21 |
+
from torchvision import transforms
|
22 |
+
from PIL import Image
|
23 |
+
|
24 |
+
class CT_SINGLE(nn.Module):
|
25 |
+
def __init__(self, model_name):
|
26 |
+
super(CT_SINGLE, self).__init__()
|
27 |
+
print(model_name)
|
28 |
+
self.model_global = timm.create_model(model_name, pretrained=False, num_classes=0)
|
29 |
+
self.fc = nn.Linear(768, 2)
|
30 |
+
|
31 |
+
def forward(self, x_global):
|
32 |
+
features_global = self.model_global(x_global)
|
33 |
+
logits = self.fc(features_global)
|
34 |
+
return logits
|
35 |
+
|
36 |
+
def load_model(checkpoint_path, model):
|
37 |
+
checkpoint = torch.load(checkpoint_path, map_location='cpu')
|
38 |
+
model.load_state_dict(checkpoint['model'])
|
39 |
+
print("Resume checkpoint %s" % checkpoint_path)
|
40 |
+
|
41 |
+
##load the model
|
42 |
+
model = CT_SINGLE('convnext_tiny')
|
43 |
+
model_path = 'Your model path'
|
44 |
+
load_model(model_path, model)
|
45 |
+
model.eval().cuda()
|
46 |
+
|
47 |
+
```
|
48 |
+
|
49 |
+
|
50 |
+
|
51 |
+
> ### Step3: Construct and predict your own data
|
52 |
+
|
53 |
+
In this step, you'll construct your own dataset. Use PIL to load images and employ `transforms` from torchvision for data preprocessing.
|
54 |
+
|
55 |
+
```python
|
56 |
+
def default_loader(path):
|
57 |
+
img = Image.open(path)
|
58 |
+
return img.convert('RGB')
|
59 |
+
|
60 |
+
data_transforms = transforms.Compose([
|
61 |
+
transforms.Resize((224,224)),
|
62 |
+
transforms.ToTensor(),
|
63 |
+
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
|
64 |
+
|
65 |
+
def predict(img_path, model):
|
66 |
+
img = default_loader(img_path)
|
67 |
+
img = data_transforms(img)
|
68 |
+
img = img.unsqueeze(0)
|
69 |
+
img = img.cuda()
|
70 |
+
output = model(img)
|
71 |
+
_, pred = torch.topk(output, 1, dim=-1)
|
72 |
+
pred = pred.data.cpu().numpy()[:, 0]
|
73 |
+
return pred ## 0 indicates non-pathology image and 1 indicates pathology image
|
74 |
+
|
75 |
+
img_path = 'Your image path'
|
76 |
+
pred = predict(img_path, model)
|
77 |
+
print(pred)
|
78 |
+
```
|
79 |
+
|