app.py

import torch
import gradio as gr
from PIL import Image
import requests
from io import BytesIO

# Load the model
model_url = "https://huggingface.co/facebook/sapiens/resolve/main/sapiens_lite_host/torchscript/normal/checkpoints/sapiens_0.3b/sapiens_0.3b.pt"
model = torch.jit.load(model_url, map_location=torch.device('cpu'))

# Define inference function
def predict(image):
    # Preprocess image
    image = image.convert("RGB")
    input_tensor = torch.from_numpy(np.array(image)).permute(2, 0, 1).unsqueeze(0).float() / 255.0
    
    # Run model
    with torch.no_grad():
        output = model(input_tensor)
    
    # Postprocess output
    output_image = output.squeeze().permute(1, 2, 0).numpy()
    output_image = (output_image * 255).astype(np.uint8)
    return Image.fromarray(output_image)

# Gradio Interface
iface = gr.Interface(
    fn=predict, 
    inputs=gr.Image(type="pil"),
    outputs=gr.Image(type="pil"),
    title="Sapiens Body Part Segmentation",
    description="Upload an image to segment body parts using the Sapiens model."
)

# Launch the interface
if __name__ == "__main__":
    iface.launch()