handler.py

from transformers import pipeline
from PIL import Image
import io, base64, requests, sys, traceback

class EndpointHandler:
    def __init__(self, model_dir: str = "", **kwargs):
        print("🔹 [INIT] Loading Swin2SR model ...")
        try:
            self.model = pipeline(task="image-to-image", model="sergeipetrov/swin2SR-classical-sr-x2-64")
            print("✅ [INIT] Model loaded successfully")
        except Exception as e:
            print("❌ [INIT] Model load failed:", str(e))
            traceback.print_exc(file=sys.stdout)
            raise e

    def __call__(self, data):
        print("\n🟢 [CALL] Received request in handler")
        print(f"🔹 [DEBUG] Raw data type: {type(data)}")
        print(f"🔹 [DEBUG] Raw data keys: {list(data.keys()) if isinstance(data, dict) else 'N/A'}")

        try:
            image_input = data.get("inputs")
            print(f"🔹 [DEBUG] image_input type: {type(image_input)}")

            # Handle nested dicts (double-wrapped inputs)
            if isinstance(image_input, dict) and "inputs" in image_input:
                print("⚠️ [DEBUG] Nested 'inputs' dict detected — unwrapping...")
                image_input = image_input["inputs"]

            # Case 1: URL input
            if isinstance(image_input, str) and image_input.startswith("http"):
                print(f"🌐 [INFO] Fetching image from URL: {image_input[:60]}...")
                image = Image.open(requests.get(image_input, stream=True).raw).convert("RGB")

            # Case 2: Base64-encoded string
            elif isinstance(image_input, str):
                print(f"🧬 [INFO] Detected base64 string (len={len(image_input)})")
                try:
                    image_bytes = base64.b64decode(image_input)
                    image = Image.open(io.BytesIO(image_bytes)).convert("RGB")
                except Exception as e:
                    print("❌ [ERROR] Base64 decode failed:", str(e))
                    traceback.print_exc(file=sys.stdout)
                    return {"error": f"Failed to decode base64 image: {str(e)}"}

            # Case 3: Raw bytes
            elif isinstance(image_input, (bytes, bytearray)):
                print(f"📦 [INFO] Detected raw bytes input (len={len(image_input)})")
                image = Image.open(io.BytesIO(image_input)).convert("RGB")

            else:
                print(f"⚠️ [WARN] Unsupported input type: {type(image_input)}")
                return {"error": f"Unsupported input type: {type(image_input)}"}

            print("✅ [INFO] Image successfully loaded and converted to RGB")

            # Run inference
            print("🚀 [INFER] Running Swin2SR model inference...")
            output = self.model(image)
            print("✅ [INFER] Inference complete")

            # Normalize output format
            if isinstance(output, (list, tuple)):
                print("🔄 [DEBUG] Output is list/tuple — taking first element")
                output = output[0]
            elif isinstance(output, dict) and "image" in output:
                print("🔄 [DEBUG] Output is dict with 'image' key")
                output = output["image"]

            # Validate output type
            if not isinstance(output, Image.Image):
                msg = f"Unexpected model output type: {type(output)}"
                print("❌ [ERROR]", msg)
                return {"error": msg}

            # Encode to base64 for API response
            print("💾 [ENCODE] Encoding result image to base64...")
            buffer = io.BytesIO()
            output.save(buffer, format="PNG")
            encoded = base64.b64encode(buffer.getvalue()).decode("utf-8")

            print("✅ [RETURN] Returning base64-encoded image")
            return {"image_base64": encoded}

        except Exception as e:
            print("❌ [FATAL] Inference failed with exception:")
            traceback.print_exc(file=sys.stdout)
            return {"error": f"Inference failed: {str(e)}"}