Update app.py
Browse files
app.py
CHANGED
|
@@ -1,132 +1,132 @@
|
|
| 1 |
-
import torch
|
| 2 |
-
import torch.nn as nn
|
| 3 |
-
from torchvision import models, transforms
|
| 4 |
-
from PIL import Image
|
| 5 |
-
import gradio as gr
|
| 6 |
-
from captum.attr import LayerGradCam
|
| 7 |
-
import numpy as np
|
| 8 |
-
import matplotlib.pyplot as plt
|
| 9 |
-
from io import BytesIO
|
| 10 |
-
|
| 11 |
-
# Configuration
|
| 12 |
-
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 13 |
-
CLASS_NAMES = ['airplane', 'automobile', 'bird', 'cat', 'deer',
|
| 14 |
-
'dog', 'frog', 'horse', 'ship', 'truck']
|
| 15 |
-
|
| 16 |
-
# Load model
|
| 17 |
-
def load_model():
|
| 18 |
-
model = models.resnet18(weights=None)
|
| 19 |
-
model.fc = nn.Linear(512, 10)
|
| 20 |
-
|
| 21 |
-
checkpoint = torch.load('
|
| 22 |
-
model.load_state_dict(checkpoint['model_state_dict'])
|
| 23 |
-
model = model.to(DEVICE)
|
| 24 |
-
model.eval()
|
| 25 |
-
|
| 26 |
-
return model
|
| 27 |
-
|
| 28 |
-
model = load_model()
|
| 29 |
-
target_layer = model.layer4[1].conv2
|
| 30 |
-
gradcam = LayerGradCam(model, target_layer)
|
| 31 |
-
|
| 32 |
-
# Image preprocessing
|
| 33 |
-
transform = transforms.Compose([
|
| 34 |
-
transforms.Resize((224, 224)),
|
| 35 |
-
transforms.ToTensor(),
|
| 36 |
-
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
|
| 37 |
-
])
|
| 38 |
-
|
| 39 |
-
# Prediction function
|
| 40 |
-
def predict_and_explain(image):
|
| 41 |
-
if image is None:
|
| 42 |
-
return "Please upload an image", None
|
| 43 |
-
|
| 44 |
-
# Preprocess
|
| 45 |
-
img_tensor = transform(image).unsqueeze(0).to(DEVICE)
|
| 46 |
-
|
| 47 |
-
# Predict
|
| 48 |
-
with torch.no_grad():
|
| 49 |
-
output = model(img_tensor)
|
| 50 |
-
probabilities = torch.softmax(output, dim=1)
|
| 51 |
-
pred_class = probabilities.argmax(1).item()
|
| 52 |
-
confidence = probabilities[0][pred_class].item()
|
| 53 |
-
|
| 54 |
-
# Generate Grad-CAM
|
| 55 |
-
attributions = gradcam.attribute(img_tensor, target=pred_class)
|
| 56 |
-
attr_np = attributions.squeeze().cpu().detach().numpy()
|
| 57 |
-
attr_np = (attr_np - attr_np.min()) / (attr_np.max() - attr_np.min() + 1e-8)
|
| 58 |
-
|
| 59 |
-
# Create visualization
|
| 60 |
-
fig, axes = plt.subplots(1, 3, figsize=(15, 5))
|
| 61 |
-
|
| 62 |
-
axes[0].imshow(image)
|
| 63 |
-
axes[0].set_title("Original Image", fontsize=14, fontweight='bold')
|
| 64 |
-
axes[0].axis('off')
|
| 65 |
-
|
| 66 |
-
im = axes[1].imshow(attr_np, cmap='jet')
|
| 67 |
-
axes[1].set_title("Grad-CAM Heatmap", fontsize=14, fontweight='bold')
|
| 68 |
-
axes[1].axis('off')
|
| 69 |
-
plt.colorbar(im, ax=axes[1], fraction=0.046)
|
| 70 |
-
|
| 71 |
-
axes[2].imshow(image)
|
| 72 |
-
axes[2].imshow(attr_np, cmap='jet', alpha=0.5)
|
| 73 |
-
axes[2].set_title(f"Overlay\nPrediction: {CLASS_NAMES[pred_class]}",
|
| 74 |
-
fontsize=14, fontweight='bold')
|
| 75 |
-
axes[2].axis('off')
|
| 76 |
-
|
| 77 |
-
plt.tight_layout()
|
| 78 |
-
|
| 79 |
-
buf = BytesIO()
|
| 80 |
-
plt.savefig(buf, format='png', dpi=100, bbox_inches='tight')
|
| 81 |
-
buf.seek(0)
|
| 82 |
-
result_image = Image.open(buf)
|
| 83 |
-
plt.close(fig)
|
| 84 |
-
|
| 85 |
-
# Prediction text
|
| 86 |
-
prediction_text = f"**Prediction:** {CLASS_NAMES[pred_class]}\n\n"
|
| 87 |
-
prediction_text += f"**Confidence:** {confidence*100:.2f}%\n\n"
|
| 88 |
-
prediction_text += "**Top 3 Predictions:**\n"
|
| 89 |
-
|
| 90 |
-
top3_probs, top3_indices = torch.topk(probabilities[0], 3)
|
| 91 |
-
for prob, idx in zip(top3_probs, top3_indices):
|
| 92 |
-
prediction_text += f"- {CLASS_NAMES[idx]}: {prob.item()*100:.2f}%\n"
|
| 93 |
-
|
| 94 |
-
return prediction_text, result_image
|
| 95 |
-
|
| 96 |
-
# Gradio Interface
|
| 97 |
-
with gr.Blocks(title="π Explainable Image Classifier", theme=gr.themes.Soft()) as demo:
|
| 98 |
-
gr.Markdown("""
|
| 99 |
-
# π Explainable Image Classifier with Grad-CAM
|
| 100 |
-
|
| 101 |
-
Upload an image and see:
|
| 102 |
-
- **What** the AI predicts (classification)
|
| 103 |
-
- **Why** it made that decision (Grad-CAM visualization)
|
| 104 |
-
|
| 105 |
-
**Supported categories:** airplane, car, bird, cat, deer, dog, frog, horse, ship, truck
|
| 106 |
-
""")
|
| 107 |
-
|
| 108 |
-
with gr.Row():
|
| 109 |
-
with gr.Column():
|
| 110 |
-
input_image = gr.Image(type="pil", label="Upload Image")
|
| 111 |
-
predict_btn = gr.Button("π Analyze Image", variant="primary", size="lg")
|
| 112 |
-
|
| 113 |
-
with gr.Column():
|
| 114 |
-
output_text = gr.Markdown(label="Prediction Results")
|
| 115 |
-
output_image = gr.Image(label="Grad-CAM Visualization", type="pil")
|
| 116 |
-
|
| 117 |
-
predict_btn.click(
|
| 118 |
-
fn=predict_and_explain,
|
| 119 |
-
inputs=input_image,
|
| 120 |
-
outputs=[output_text, output_image]
|
| 121 |
-
)
|
| 122 |
-
|
| 123 |
-
gr.Markdown("""
|
| 124 |
-
---
|
| 125 |
-
### π§ About This Model
|
| 126 |
-
- **Architecture:** ResNet18 (transfer learning)
|
| 127 |
-
- **Training Data:** CIFAR-10 (60,000 images)
|
| 128 |
-
- **Explainability:** Grad-CAM visualization
|
| 129 |
-
""")
|
| 130 |
-
|
| 131 |
-
if __name__ == "__main__":
|
| 132 |
demo.launch()
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
import torch.nn as nn
|
| 3 |
+
from torchvision import models, transforms
|
| 4 |
+
from PIL import Image
|
| 5 |
+
import gradio as gr
|
| 6 |
+
from captum.attr import LayerGradCam
|
| 7 |
+
import numpy as np
|
| 8 |
+
import matplotlib.pyplot as plt
|
| 9 |
+
from io import BytesIO
|
| 10 |
+
|
| 11 |
+
# Configuration
|
| 12 |
+
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 13 |
+
CLASS_NAMES = ['airplane', 'automobile', 'bird', 'cat', 'deer',
|
| 14 |
+
'dog', 'frog', 'horse', 'ship', 'truck']
|
| 15 |
+
|
| 16 |
+
# Load model
|
| 17 |
+
def load_model():
|
| 18 |
+
model = models.resnet18(weights=None)
|
| 19 |
+
model.fc = nn.Linear(512, 10)
|
| 20 |
+
|
| 21 |
+
checkpoint = torch.load('xai_resnet18.pth', map_location=DEVICE)
|
| 22 |
+
model.load_state_dict(checkpoint['model_state_dict'])
|
| 23 |
+
model = model.to(DEVICE)
|
| 24 |
+
model.eval()
|
| 25 |
+
|
| 26 |
+
return model
|
| 27 |
+
|
| 28 |
+
model = load_model()
|
| 29 |
+
target_layer = model.layer4[1].conv2
|
| 30 |
+
gradcam = LayerGradCam(model, target_layer)
|
| 31 |
+
|
| 32 |
+
# Image preprocessing
|
| 33 |
+
transform = transforms.Compose([
|
| 34 |
+
transforms.Resize((224, 224)),
|
| 35 |
+
transforms.ToTensor(),
|
| 36 |
+
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
|
| 37 |
+
])
|
| 38 |
+
|
| 39 |
+
# Prediction function
|
| 40 |
+
def predict_and_explain(image):
|
| 41 |
+
if image is None:
|
| 42 |
+
return "Please upload an image", None
|
| 43 |
+
|
| 44 |
+
# Preprocess
|
| 45 |
+
img_tensor = transform(image).unsqueeze(0).to(DEVICE)
|
| 46 |
+
|
| 47 |
+
# Predict
|
| 48 |
+
with torch.no_grad():
|
| 49 |
+
output = model(img_tensor)
|
| 50 |
+
probabilities = torch.softmax(output, dim=1)
|
| 51 |
+
pred_class = probabilities.argmax(1).item()
|
| 52 |
+
confidence = probabilities[0][pred_class].item()
|
| 53 |
+
|
| 54 |
+
# Generate Grad-CAM
|
| 55 |
+
attributions = gradcam.attribute(img_tensor, target=pred_class)
|
| 56 |
+
attr_np = attributions.squeeze().cpu().detach().numpy()
|
| 57 |
+
attr_np = (attr_np - attr_np.min()) / (attr_np.max() - attr_np.min() + 1e-8)
|
| 58 |
+
|
| 59 |
+
# Create visualization
|
| 60 |
+
fig, axes = plt.subplots(1, 3, figsize=(15, 5))
|
| 61 |
+
|
| 62 |
+
axes[0].imshow(image)
|
| 63 |
+
axes[0].set_title("Original Image", fontsize=14, fontweight='bold')
|
| 64 |
+
axes[0].axis('off')
|
| 65 |
+
|
| 66 |
+
im = axes[1].imshow(attr_np, cmap='jet')
|
| 67 |
+
axes[1].set_title("Grad-CAM Heatmap", fontsize=14, fontweight='bold')
|
| 68 |
+
axes[1].axis('off')
|
| 69 |
+
plt.colorbar(im, ax=axes[1], fraction=0.046)
|
| 70 |
+
|
| 71 |
+
axes[2].imshow(image)
|
| 72 |
+
axes[2].imshow(attr_np, cmap='jet', alpha=0.5)
|
| 73 |
+
axes[2].set_title(f"Overlay\nPrediction: {CLASS_NAMES[pred_class]}",
|
| 74 |
+
fontsize=14, fontweight='bold')
|
| 75 |
+
axes[2].axis('off')
|
| 76 |
+
|
| 77 |
+
plt.tight_layout()
|
| 78 |
+
|
| 79 |
+
buf = BytesIO()
|
| 80 |
+
plt.savefig(buf, format='png', dpi=100, bbox_inches='tight')
|
| 81 |
+
buf.seek(0)
|
| 82 |
+
result_image = Image.open(buf)
|
| 83 |
+
plt.close(fig)
|
| 84 |
+
|
| 85 |
+
# Prediction text
|
| 86 |
+
prediction_text = f"**Prediction:** {CLASS_NAMES[pred_class]}\n\n"
|
| 87 |
+
prediction_text += f"**Confidence:** {confidence*100:.2f}%\n\n"
|
| 88 |
+
prediction_text += "**Top 3 Predictions:**\n"
|
| 89 |
+
|
| 90 |
+
top3_probs, top3_indices = torch.topk(probabilities[0], 3)
|
| 91 |
+
for prob, idx in zip(top3_probs, top3_indices):
|
| 92 |
+
prediction_text += f"- {CLASS_NAMES[idx]}: {prob.item()*100:.2f}%\n"
|
| 93 |
+
|
| 94 |
+
return prediction_text, result_image
|
| 95 |
+
|
| 96 |
+
# Gradio Interface
|
| 97 |
+
with gr.Blocks(title="π Explainable Image Classifier", theme=gr.themes.Soft()) as demo:
|
| 98 |
+
gr.Markdown("""
|
| 99 |
+
# π Explainable Image Classifier with Grad-CAM
|
| 100 |
+
|
| 101 |
+
Upload an image and see:
|
| 102 |
+
- **What** the AI predicts (classification)
|
| 103 |
+
- **Why** it made that decision (Grad-CAM visualization)
|
| 104 |
+
|
| 105 |
+
**Supported categories:** airplane, car, bird, cat, deer, dog, frog, horse, ship, truck
|
| 106 |
+
""")
|
| 107 |
+
|
| 108 |
+
with gr.Row():
|
| 109 |
+
with gr.Column():
|
| 110 |
+
input_image = gr.Image(type="pil", label="Upload Image")
|
| 111 |
+
predict_btn = gr.Button("π Analyze Image", variant="primary", size="lg")
|
| 112 |
+
|
| 113 |
+
with gr.Column():
|
| 114 |
+
output_text = gr.Markdown(label="Prediction Results")
|
| 115 |
+
output_image = gr.Image(label="Grad-CAM Visualization", type="pil")
|
| 116 |
+
|
| 117 |
+
predict_btn.click(
|
| 118 |
+
fn=predict_and_explain,
|
| 119 |
+
inputs=input_image,
|
| 120 |
+
outputs=[output_text, output_image]
|
| 121 |
+
)
|
| 122 |
+
|
| 123 |
+
gr.Markdown("""
|
| 124 |
+
---
|
| 125 |
+
### π§ About This Model
|
| 126 |
+
- **Architecture:** ResNet18 (transfer learning)
|
| 127 |
+
- **Training Data:** CIFAR-10 (60,000 images)
|
| 128 |
+
- **Explainability:** Grad-CAM visualization
|
| 129 |
+
""")
|
| 130 |
+
|
| 131 |
+
if __name__ == "__main__":
|
| 132 |
demo.launch()
|