app.py

from fastapi import FastAPI, File, UploadFile, HTTPException
from fastapi.responses import JSONResponse
from pydantic import BaseModel
import torch
import torch.nn.functional as F
from transformers import AutoImageProcessor, AutoModelForImageClassification
from PIL import Image
import io
import numpy as np
from typing import List, Dict, Any
import logging
import os

# Configure logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

app = FastAPI(
    title="ChatGPT Oasis Model Inference API",
    description="FastAPI inference server for Oasis and ViT models deployed on Hugging Face Spaces with Docker",
    version="1.0.0"
)

# Global variables to store loaded models
oasis_model = None
oasis_processor = None
vit_model = None
vit_processor = None

class InferenceRequest(BaseModel):
    image: str  # Base64 encoded image
    model_name: str = "oasis500m"  # Default to oasis model

class InferenceResponse(BaseModel):
    predictions: List[Dict[str, Any]]
    model_used: str
    confidence_scores: List[float]

def load_models():
    """Load both models from local files"""
    global oasis_model, oasis_processor, vit_model, vit_processor
    
    try:
        logger.info("Loading Oasis 500M model from local files...")
        # Load Oasis model from local files
        oasis_processor = AutoImageProcessor.from_pretrained("microsoft/oasis-500m")
        oasis_model = AutoModelForImageClassification.from_pretrained(
            "microsoft/oasis-500m",
            local_files_only=False  # Will download config but use local weights
        )
        
        # Load local weights if available
        oasis_model_path = "/app/models/oasis500m.safetensors"
        if os.path.exists(oasis_model_path):
            logger.info("Loading Oasis weights from local file...")
            from safetensors.torch import load_file
            state_dict = load_file(oasis_model_path)
            oasis_model.load_state_dict(state_dict, strict=False)
        
        oasis_model.eval()
        
        logger.info("Loading ViT-L-20 model from local files...")
        # Load ViT model from local files
        vit_processor = AutoImageProcessor.from_pretrained("google/vit-large-patch16-224")
        vit_model = AutoModelForImageClassification.from_pretrained(
            "google/vit-large-patch16-224",
            local_files_only=False  # Will download config but use local weights
        )
        
        # Load local weights if available
        vit_model_path = "/app/models/vit-l-20.safetensors"
        if os.path.exists(vit_model_path):
            logger.info("Loading ViT weights from local file...")
            from safetensors.torch import load_file
            state_dict = load_file(vit_model_path)
            vit_model.load_state_dict(state_dict, strict=False)
        
        vit_model.eval()
        
        logger.info("All models loaded successfully!")
        
    except Exception as e:
        logger.error(f"Error loading models: {e}")
        raise e

@app.on_event("startup")
async def startup_event():
    """Load models when the application starts"""
    load_models()

@app.get("/")
async def root():
    """Root endpoint with API information"""
    return {
        "message": "ChatGPT Oasis Model Inference API",
        "version": "1.0.0",
        "deployed_on": "Hugging Face Spaces (Docker)",
        "available_models": ["oasis500m", "vit-l-20"],
        "endpoints": {
            "health": "/health",
            "inference": "/inference",
            "upload_inference": "/upload_inference",
            "predict": "/predict"
        },
        "usage": {
            "base64_inference": "POST /inference with JSON body containing 'image' (base64) and 'model_name'",
            "file_upload": "POST /upload_inference with multipart form containing 'file' and optional 'model_name'",
            "simple_predict": "POST /predict with file upload for quick inference"
        }
    }

@app.get("/health")
async def health_check():
    """Health check endpoint"""
    models_status = {
        "oasis500m": oasis_model is not None,
        "vit-l-20": vit_model is not None
    }
    
    # Check if model files exist
    model_files = {
        "oasis500m": os.path.exists("/app/models/oasis500m.safetensors"),
        "vit-l-20": os.path.exists("/app/models/vit-l-20.safetensors")
    }
    
    return {
        "status": "healthy",
        "models_loaded": models_status,
        "model_files_present": model_files,
        "deployment": "huggingface-spaces-docker"
    }

def process_image_with_model(image: Image.Image, model_name: str):
    """Process image with the specified model"""
    if model_name == "oasis500m":
        if oasis_model is None or oasis_processor is None:
            raise HTTPException(status_code=500, detail="Oasis model not loaded")
        
        inputs = oasis_processor(images=image, return_tensors="pt")
        with torch.no_grad():
            outputs = oasis_model(**inputs)
            logits = outputs.logits
            probabilities = F.softmax(logits, dim=-1)
            
        # Get top predictions
        top_probs, top_indices = torch.topk(probabilities, 5)
        
        predictions = []
        for i in range(top_indices.shape[1]):
            pred = {
                "label": oasis_model.config.id2label[top_indices[0][i].item()],
                "confidence": top_probs[0][i].item()
            }
            predictions.append(pred)
            
        return predictions
        
    elif model_name == "vit-l-20":
        if vit_model is None or vit_processor is None:
            raise HTTPException(status_code=500, detail="ViT model not loaded")
        
        inputs = vit_processor(images=image, return_tensors="pt")
        with torch.no_grad():
            outputs = vit_model(**inputs)
            logits = outputs.logits
            probabilities = F.softmax(logits, dim=-1)
            
        # Get top predictions
        top_probs, top_indices = torch.topk(probabilities, 5)
        
        predictions = []
        for i in range(top_indices.shape[1]):
            pred = {
                "label": vit_model.config.id2label[top_indices[0][i].item()],
                "confidence": top_probs[0][i].item()
            }
            predictions.append(pred)
            
        return predictions
    
    else:
        raise HTTPException(status_code=400, detail=f"Unknown model: {model_name}")

@app.post("/inference", response_model=InferenceResponse)
async def inference(request: InferenceRequest):
    """Inference endpoint using base64 encoded image"""
    try:
        import base64
        
        # Decode base64 image
        image_data = base64.b64decode(request.image)
        image = Image.open(io.BytesIO(image_data)).convert('RGB')
        
        # Process with model
        predictions = process_image_with_model(image, request.model_name)
        
        # Extract confidence scores
        confidence_scores = [pred["confidence"] for pred in predictions]
        
        return InferenceResponse(
            predictions=predictions,
            model_used=request.model_name,
            confidence_scores=confidence_scores
        )
        
    except Exception as e:
        logger.error(f"Inference error: {e}")
        raise HTTPException(status_code=500, detail=str(e))

@app.post("/upload_inference", response_model=InferenceResponse)
async def upload_inference(
    file: UploadFile = File(...),
    model_name: str = "oasis500m"
):
    """Inference endpoint using file upload"""
    try:
        # Validate file type
        if not file.content_type.startswith('image/'):
            raise HTTPException(status_code=400, detail="File must be an image")
        
        # Read and process image
        image_data = await file.read()
        image = Image.open(io.BytesIO(image_data)).convert('RGB')
        
        # Process with model
        predictions = process_image_with_model(image, model_name)
        
        # Extract confidence scores
        confidence_scores = [pred["confidence"] for pred in predictions]
        
        return InferenceResponse(
            predictions=predictions,
            model_used=model_name,
            confidence_scores=confidence_scores
        )
        
    except Exception as e:
        logger.error(f"Upload inference error: {e}")
        raise HTTPException(status_code=500, detail=str(e))

@app.get("/models")
async def list_models():
    """List available models and their status"""
    return {
        "available_models": [
            {
                "name": "oasis500m",
                "description": "Oasis 500M vision model",
                "loaded": oasis_model is not None,
                "file_present": os.path.exists("/app/models/oasis500m.safetensors")
            },
            {
                "name": "vit-l-20",
                "description": "Vision Transformer Large model",
                "loaded": vit_model is not None,
                "file_present": os.path.exists("/app/models/vit-l-20.safetensors")
            }
        ]
    }

# Hugging Face Spaces specific endpoint for Gradio compatibility
@app.post("/predict")
async def predict(file: UploadFile = File(...)):
    """Simple prediction endpoint for Hugging Face Spaces integration"""
    try:
        # Validate file type
        if not file.content_type.startswith('image/'):
            raise HTTPException(status_code=400, detail="File must be an image")
        
        # Read and process image
        image_data = await file.read()
        image = Image.open(io.BytesIO(image_data)).convert('RGB')
        
        # Process with default model (oasis500m)
        predictions = process_image_with_model(image, "oasis500m")
        
        # Return simplified format for Gradio
        return {
            "predictions": predictions[:3],  # Top 3 predictions
            "model_used": "oasis500m"
        }
        
    except Exception as e:
        logger.error(f"Predict error: {e}")
        raise HTTPException(status_code=500, detail=str(e))