app.py

import gradio as gr
from transformers import AutoProcessor, AutoModel, AutoTokenizer, pipeline
from PIL import Image
import torch

# ---------------------------
# Load encoder (MedSigLIP)
# ---------------------------
ENCODER_ID = "fokan/medsiglip-448-int8"
encoder_processor = AutoProcessor.from_pretrained(ENCODER_ID)
encoder_model = AutoModel.from_pretrained(ENCODER_ID).eval()

# ---------------------------
# Load decoder (MedGemma)
# ---------------------------
DECODER_ID = "fokan/medgemma-4b-it-int8"
decoder = pipeline("text-generation", model=DECODER_ID, torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32)

# ---------------------------
# Core Function
# ---------------------------
@torch.no_grad()
def analyze_xray(image):
    # Step 1: Encode the image into embedding
    inputs = encoder_processor(images=image, text=["chest x-ray"], return_tensors="pt", padding=True)
    outputs = encoder_model(**inputs)
    if hasattr(outputs, "image_embeds"):
        embedding = outputs.image_embeds[0]
    elif hasattr(outputs, "last_hidden_state"):
        embedding = outputs.last_hidden_state.mean(dim=1)[0]
    else:
        embedding = list(outputs.values())[0].mean(dim=1)[0]
    embedding = embedding / embedding.norm()

    # Step 2: Generate a short diagnostic report
    prompt = (
        "You are a radiologist. Analyze this chest X-ray embedding vector and describe any possible findings, "
        "anomalies, or impressions as a short professional report.\n"
        f"<embedding>{embedding[:256].tolist()}</embedding>"
    )

    report = decoder(prompt, max_new_tokens=180, temperature=0.8, top_p=0.9)[0]["generated_text"]

    # Return both embedding preview + text
    preview = embedding[:5].tolist()
    return f"✅ Embedding (preview): {preview}\n\n🩺 **AI Radiology Report:**\n{report}"

# ---------------------------
# Gradio UI
# ---------------------------
title = "🩻 MedSigLIP → MedGemma Fusion"
desc = """
Upload an **X-ray image**, and this demo will:
1. Extract its visual embedding using `fokan/medsiglip-448-fp16-pruned20`.
2. Generate a **radiology-style report** using `fokan/medgemma-4b-it-fp16-pruned20`.
"""

demo = gr.Interface(
    fn=analyze_xray,
    inputs=gr.Image(type="pil", label="Upload X-ray Image"),
    outputs=gr.Markdown(label="AI Report"),
    title=title,
    description=desc,
    theme="gradio/soft",
)

if __name__ == "__main__":
    demo.launch()