Update app.py
Browse files
app.py
CHANGED
@@ -8,8 +8,13 @@ from pytorch_grad_cam import GradCAM
|
|
8 |
from pytorch_grad_cam.utils.model_targets import ClassifierOutputTarget
|
9 |
from pytorch_grad_cam.utils.image import show_cam_on_image
|
10 |
|
11 |
-
|
|
|
|
|
|
|
12 |
device = torch.device("cpu")
|
|
|
|
|
13 |
|
14 |
# Load model
|
15 |
model = models.resnet50(weights=None)
|
@@ -18,11 +23,11 @@ model.load_state_dict(torch.load("resnet50_dr_classifier.pth", map_location=devi
|
|
18 |
model.to(device)
|
19 |
model.eval()
|
20 |
|
21 |
-
# Grad-CAM
|
22 |
target_layer = model.layer4[-1]
|
23 |
cam = GradCAM(model=model, target_layers=[target_layer])
|
24 |
|
25 |
-
#
|
26 |
transform = transforms.Compose([
|
27 |
transforms.Resize((224, 224)),
|
28 |
transforms.ToTensor(),
|
@@ -30,8 +35,9 @@ transform = transforms.Compose([
|
|
30 |
[0.229, 0.224, 0.225])
|
31 |
])
|
32 |
|
33 |
-
#
|
34 |
def predict_retinopathy(image):
|
|
|
35 |
img = image.convert("RGB").resize((224, 224))
|
36 |
img_tensor = transform(img).unsqueeze(0).to(device)
|
37 |
|
@@ -41,18 +47,22 @@ def predict_retinopathy(image):
|
|
41 |
pred = torch.argmax(probs, dim=1).item()
|
42 |
confidence = probs[0][pred].item()
|
43 |
|
44 |
-
label = "
|
45 |
|
46 |
# Grad-CAM
|
47 |
rgb_img_np = np.array(img).astype(np.float32) / 255.0
|
48 |
rgb_img_np = np.ascontiguousarray(rgb_img_np)
|
49 |
grayscale_cam = cam(input_tensor=img_tensor, targets=[ClassifierOutputTarget(pred)])[0]
|
50 |
cam_image = show_cam_on_image(rgb_img_np, grayscale_cam, use_rgb=True)
|
51 |
-
|
52 |
cam_pil = Image.fromarray(cam_image)
|
|
|
|
|
|
|
|
|
|
|
53 |
return cam_pil, f"{label} (Confidence: {confidence:.2f})"
|
54 |
|
55 |
-
# Gradio
|
56 |
gr.Interface(
|
57 |
fn=predict_retinopathy,
|
58 |
inputs=gr.Image(type="pil"),
|
@@ -61,5 +71,5 @@ gr.Interface(
|
|
61 |
gr.Text(label="Prediction")
|
62 |
],
|
63 |
title="Diabetic Retinopathy Detection",
|
64 |
-
description="Upload a retinal image to classify DR and view Grad-CAM heatmap."
|
65 |
).launch()
|
|
|
8 |
from pytorch_grad_cam.utils.model_targets import ClassifierOutputTarget
|
9 |
from pytorch_grad_cam.utils.image import show_cam_on_image
|
10 |
|
11 |
+
import os
|
12 |
+
import datetime
|
13 |
+
|
14 |
+
# Setup
|
15 |
device = torch.device("cpu")
|
16 |
+
save_dir = "saved_predictions"
|
17 |
+
os.makedirs(save_dir, exist_ok=True)
|
18 |
|
19 |
# Load model
|
20 |
model = models.resnet50(weights=None)
|
|
|
23 |
model.to(device)
|
24 |
model.eval()
|
25 |
|
26 |
+
# Grad-CAM
|
27 |
target_layer = model.layer4[-1]
|
28 |
cam = GradCAM(model=model, target_layers=[target_layer])
|
29 |
|
30 |
+
# Preprocessing
|
31 |
transform = transforms.Compose([
|
32 |
transforms.Resize((224, 224)),
|
33 |
transforms.ToTensor(),
|
|
|
35 |
[0.229, 0.224, 0.225])
|
36 |
])
|
37 |
|
38 |
+
# Predict and save
|
39 |
def predict_retinopathy(image):
|
40 |
+
timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
|
41 |
img = image.convert("RGB").resize((224, 224))
|
42 |
img_tensor = transform(img).unsqueeze(0).to(device)
|
43 |
|
|
|
47 |
pred = torch.argmax(probs, dim=1).item()
|
48 |
confidence = probs[0][pred].item()
|
49 |
|
50 |
+
label = "DR" if pred == 0 else "NoDR"
|
51 |
|
52 |
# Grad-CAM
|
53 |
rgb_img_np = np.array(img).astype(np.float32) / 255.0
|
54 |
rgb_img_np = np.ascontiguousarray(rgb_img_np)
|
55 |
grayscale_cam = cam(input_tensor=img_tensor, targets=[ClassifierOutputTarget(pred)])[0]
|
56 |
cam_image = show_cam_on_image(rgb_img_np, grayscale_cam, use_rgb=True)
|
|
|
57 |
cam_pil = Image.fromarray(cam_image)
|
58 |
+
|
59 |
+
# Save image with label and confidence
|
60 |
+
filename = f"{timestamp}_{label}_{confidence:.2f}.png"
|
61 |
+
cam_pil.save(os.path.join(save_dir, filename))
|
62 |
+
|
63 |
return cam_pil, f"{label} (Confidence: {confidence:.2f})"
|
64 |
|
65 |
+
# Gradio app
|
66 |
gr.Interface(
|
67 |
fn=predict_retinopathy,
|
68 |
inputs=gr.Image(type="pil"),
|
|
|
71 |
gr.Text(label="Prediction")
|
72 |
],
|
73 |
title="Diabetic Retinopathy Detection",
|
74 |
+
description="Upload a retinal image to classify DR and view Grad-CAM heatmap. All predictions are auto-saved with label and confidence."
|
75 |
).launch()
|