app.py

import gradio as gr
import torch
from torch import nn
from torchvision import models, transforms
from huggingface_hub import hf_hub_download
from PIL import Image
import requests
import base64
from io import BytesIO

# Define the number of classes
num_classes = 2  # Update with the actual number of classes in your dataset (e.g., 2 for healthy and anomalous)

# Download model from Hugging Face
def download_model():
    model_path = hf_hub_download(repo_id="jays009/Restnet50", filename="pytorch_model.bin")
    return model_path

# Load the model from Hugging Face
def load_model(model_path):
    model = models.resnet50(pretrained=False)  # Set pretrained=False because you're loading custom weights
    model.fc = nn.Linear(model.fc.in_features, num_classes)  # Adjust for the number of classes in your dataset
    model.load_state_dict(torch.load(model_path, map_location=torch.device("cpu")))  # Load model on CPU for compatibility
    model.eval()  # Set to evaluation mode
    return model

# Download the model and load it
model_path = download_model()  # Downloads the model from Hugging Face Hub
model = load_model(model_path)

# Define the transformation for the input image
transform = transforms.Compose([
    transforms.Resize(256),  # Resize the image to 256x256
    transforms.CenterCrop(224),  # Crop the image to 224x224
    transforms.ToTensor(),  # Convert the image to a Tensor
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),  # Normalize the image (ImageNet mean and std)
])


def predict(image):
    # Check if the input contains a base64-encoded string
    if isinstance(image, dict) and image.get("data"):
        # Decode the base64 string into a PIL image
        image_data = base64.b64decode(image["data"])
        image = Image.open(BytesIO(image_data))

    # Apply your existing transformations
    image = transform(image).unsqueeze(0)  # Transform and add batch dimension
    image = image.to(torch.device("cuda" if torch.cuda.is_available() else "cpu"))

    # Perform inference
    with torch.no_grad():
        outputs = model(image)
        predicted_class = torch.argmax(outputs, dim=1).item()
    
    # Create a response based on the predicted class
    if predicted_class == 0:
        return "The photo you've sent is of fall army worm with problem ID 126."
    elif predicted_class == 1:
        return "The photo you've sent is of a healthy maize image."
    else:
        return "Unexpected class prediction."

# Create the Gradio interface
iface = gr.Interface(
    fn=predict,  # Function for prediction
    inputs=gr.Image(type="pil"),  # Image input
    outputs=gr.Textbox(),  # Output: Predicted class
    live=True,  # Updates as the user uploads an image
    title="Maize Anomaly Detection",
    description="Upload an image of maize to detect anomalies like disease or pest infestation. You can provide local paths, URLs, or base64-encoded images."
)

# Launch the Gradio interface
iface.launch(share=True)