app.py

"""
Grounded‑SAM Flask API (CPU only)

POST /segment
Body (multipart/form-data):
  - image: the house photo
  - prompt: text prompt, e.g. "roof sheet"
Query params:
  - overlay (bool, default=false): if true, returns a PNG overlay instead
Returns:
  - image/png mask (single channel)  OR  overlay
"""
import io
import os
import argparse
import numpy as np
from PIL import Image
from flask import Flask, request, send_file
from flask_cors import CORS
import logging

import torch
from groundingdino.util.inference import load_model, predict
from segment_anything import sam_model_registry, SamPredictor
import groundingdino.datasets.transforms as T

# ─── Load models once ───────────────────────────────────────────────────────────
device = torch.device("cpu")

DINO_CONFIG = "GroundingDINO/groundingdino/config/GroundingDINO_SwinT_OGC.py"
DINO_CKPT = "weights/groundingdino_swint_ogc.pth"
SAM_CKPT  = "weights/sam_vit_h_4b8939.pth"

# Load GroundingDINO model
grounder = load_model(DINO_CONFIG, DINO_CKPT)
sam = sam_model_registry["vit_h"](checkpoint=SAM_CKPT).to(device)
predictor = SamPredictor(sam)

# ─── Flask app ──────────────────────────────────────────────────────────────────
app = Flask(__name__)
CORS(app)
app.config["DEBUG"] = True

def segment(image_pil: Image.Image, prompt: str):
    # Use the proper image preprocessing for GroundingDINO
    transform = T.Compose([
        T.RandomResize([800], max_size=1333),
        T.ToTensor(),
        T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
    ])
    
    image_transformed, _ = transform(image_pil, None)
    image_transformed = image_transformed.to(device)

    # Run GroundingDINO to get boxes for the prompt
    boxes, _, _ = predict(
        model=grounder,
        image=image_transformed,
        caption=prompt,
        box_threshold=0.3,
        text_threshold=0.25,
        device="cpu"
    )
    if boxes.size == 0:
        raise ValueError("No boxes found for prompt.")

    # 2) Largest box → mask via SAM
    box = boxes[np.argmax((boxes[:,2]-boxes[:,0])*(boxes[:,3]-boxes[:,1]))]
    box_np = box.cpu().numpy() if hasattr(box, "cpu") else np.array(box)
    predictor.set_image(np.array(image_pil))
    masks, _, _ = predictor.predict(box=box_np)
    mask = masks[0]  # boolean HxW

    return mask

@app.route("/segment", methods=["POST"])
def segment_endpoint():
    if "image" not in request.files or "prompt" not in request.form:
        return {"error": "image file and prompt are required."}, 400

    prompt = request.form["prompt"]
    image = request.files["image"]

    # Check for unsupported file types
    if not image.content_type.startswith("image/"):
        return {"error": "unsupported file type. Only image files are allowed."}, 415

    try:
        image_pil = Image.open(image.stream).convert("RGB")
        mask = segment(image_pil, prompt)
    except ValueError as e:
        app.logger.error(f"ValueError: {e}")
        return {"error": str(e)}, 422
    except Exception as e:
        app.logger.exception("Exception in /segment endpoint")
        return {"error": str(e)}, 500  # Return actual exception message

    overlay = request.args.get("overlay", "false").lower() == "true"
    if overlay:
        colored = np.array(image_pil).copy()
        mask_np = mask.cpu().numpy() if hasattr(mask, "cpu") else np.array(mask)
        colored[mask_np] = [255, 0, 0]  # red overlay
        out_img = Image.fromarray(colored)
    else:
        out_img = Image.fromarray((mask.astype(np.uint8) * 255) if hasattr(mask, "astype") else (np.array(mask, dtype=np.uint8) * 255))

    buf = io.BytesIO()
    out_img.save(buf, format="PNG")
    buf.seek(0)
    return send_file(buf, mimetype="image/png")

# ─── CLI ────────────────────────────────────────────────────────────────────────
if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--host", default="127.0.0.1")
    parser.add_argument("--port", default=7860, type=int)
    args = parser.parse_args()
    app.run(host=args.host, port=args.port, debug=True)