app.py

from fastai.vision.all import *
import gradio as gr
from captum.attr import Saliency
from torchvision import transforms
import matplotlib.pyplot as plt
import numpy as np
import torch

learn = load_learner('animal_model.pkl')

transform = transforms.Compose([
    transforms.Resize((128,128)),
    transforms.ToTensor(),
])

categories = learn.dls.vocab

def generate_saliency(image):
    # Prepare the image for the model
    img = PILImage.create(image)
    
    # Get prediction
    _, pred, probs = learn.predict(img)
    
    # Create Captum interpretation object
    interp = Saliency(learn.model)
    
    # Transform and prepare image for saliency
    tensor_image = transform(img).unsqueeze(0)
    tensor_image = tensor_image.requires_grad_()
    
    # Generate the saliency map
    saliency_map = interp.attribute(tensor_image, target=pred)
    
    # Process saliency map for visualization
    saliency_np = saliency_map.squeeze().cpu().detach().numpy()
    saliency_np = np.abs(saliency_np).sum(axis=0)
    #saliency_np = (saliency_np - saliency_np.min()) / (saliency_np.max() - saliency_np.min())
    
    # Create heatmap
    plt.figure(figsize=(10, 10))
    plt.imshow(saliency_np, cmap='viridis')
    plt.axis('off')
    plt.tight_layout()
    plt.savefig('saliency_heatmap.png', pad_inches=0)
    plt.close()
    
    return (
        dict(zip(categories, map(float, probs))),
        'saliency_heatmap.png',
        'saliency_overlay.png'
    )

# Gradio interface
image = gr.Image(type="pil")
label = gr.Label()
examples = ['polar_bear_real.jpg', 'polar_bear.jpg']

interface = gr.Interface(
    fn=generate_saliency,
    inputs=image,
    outputs=[
        gr.Label(label="Predictions"),
        gr.Image(type="filepath", label="Saliency Heatmap")
    ],
    examples=examples
)

interface.launch()