src/app.py

import gradio as gr
import os
import sys
import logging
import numpy as np
from PIL import Image

# Configure logging
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)

# Suppress TensorFlow logging
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
try:
    import absl.logging
    absl.logging.set_verbosity(absl.logging.ERROR)
except ImportError:
    pass
logging.getLogger('tensorflow').setLevel(logging.ERROR)

from model import RawImageModel, PrecomputedModel

# Global Model Instances
raw_model = None
precomputed_model = None
pos_emb = None
neg_emb = None

# Optimal Threshold from Kaggle validation
THRESHOLD = -0.1173

def load_models():
    global raw_model, precomputed_model, pos_emb, neg_emb
    if raw_model is None:
        logger.info("Loading models...")
        try:
            precomputed_model = PrecomputedModel()
            raw_model = RawImageModel()
            
            # Pre-fetch text embeddings
            pos_txt = 'small pneumothorax'
            neg_txt = 'no pneumothorax'
            pos_emb, neg_emb = precomputed_model.get_diagnosis_embeddings(pos_txt, neg_txt)
            logger.info("Models loaded.")
        except Exception as e:
            logger.error(f"Failed to load models: {e}")
            raise e

# ZeroGPU compatibility for Hugging Face Spaces
try:
    import spaces
except ImportError:
    # Dummy decorator if running locally without spaces installed
    class spaces:
        @staticmethod
        def GPU(func):
            return func

@spaces.GPU
def predict(image):
    if image is None:
        return "No image uploaded.", 0.0, "Please upload an image."

    try:
        # Save temp image for model ingestion
        temp_path = "temp_gradio_upload.png"
        image.save(temp_path)
        
        # Run Inference
        img_emb = raw_model.compute_embeddings(temp_path)
        score = PrecomputedModel.zero_shot(img_emb, pos_emb, neg_emb)
        score = float(score)
        
        # Binary Classification
        if score >= THRESHOLD:
            prediction = '<p style="color: red; font-size: 24px; font-weight: bold;">PNEUMOTHORAX ⚠️</p>'
        else:
            prediction = '<p style="color: green; font-size: 24px; font-weight: bold;">NORMAL ✅</p>'
        
        return prediction, score, f"Raw Score: {score:.4f} (Threshold: {THRESHOLD})"

    except Exception as e:
        logger.error(f"Prediction failed: {e}")
        return "<p style='color:red'>Error</p>", 0.0, str(e)

# Load models at startup
load_models()

# UI Layout
with gr.Blocks(title="Chest X-Ray Zero-Shot Classifier") as demo:
    gr.Markdown("# 🩻 Zero-Shot Chest X-Ray Classification")
    gr.Markdown("Detect Pneumothorax from raw X-ray images using a pre-trained foundation model.")
    
    with gr.Row():
        with gr.Column():
            gr.Markdown("### 1. Upload X-Ray")
            input_image = gr.Image(type="pil", label="Upload Image (PNG/JPG/DICOM converted)")
            predict_btn = gr.Button("Analyze Image", variant="primary")
            
        with gr.Column():
            gr.Markdown("### 2. Results")
            output_label = gr.HTML(label="Prediction")
            output_score = gr.Number(label="Zero-Shot Score")
            output_msg = gr.Textbox(label="Details")
            
            gr.Markdown("---")
            gr.Markdown("### Performance Context")
            gr.Markdown("This model uses a **zero-shot** approach. The threshold was calibrated using a local Kaggle dataset.")
            
            with gr.Tabs():
                with gr.TabItem("Local Kaggle Benchmark"):
                    gr.Image("results/kaggle_roc_curve.png", label="local ROC Curve")
                    gr.Markdown("**AUC: 0.88** on 250 local samples.")
                with gr.TabItem("Google Benchmark"):
                    gr.Image("results/roc_PNEUMOTHORAX.png", label="Reference ROC")

    predict_btn.click(predict, inputs=input_image, outputs=[output_label, output_score, output_msg])

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