app.py

import gradio as gr
import torch
from PIL import Image
import numpy as np
import subprocess
import sys
import io
import matplotlib.pyplot as plt
import traceback

# --- CRITICAL PATCH: Fix for 'container_abcs' not found in torch._six ---
# This makes older code compatible with PyTorch 1.13.1 by providing the correct import.
try:
    from torch._six import container_abcs
except ImportError:
    import collections.abc as container_abcs

# --- 0. FORCE INSTALL: Install pre-built mmcv-full for _ext modules ---
try:
    print("INFO: Attempting to install pre-built mmcv-full...")
    # This installs the mmcv-full wheel pre-built for PyTorch 1.13, which includes the necessary compiled _ext modules.
    subprocess.check_call([
        sys.executable, '-m', 'pip', 'install', 
        'mmcv-full==1.7.1', 
        '-f', 'https://download.openmmlab.com/mmcv/dist/cpu/torch1.13/index.html'
    ])
    print("INFO: Successfully installed pre-built mmcv-full.")
except subprocess.CalledProcessError as e:
    print(f"FATAL ERROR: Failed to install pre-built mmcv-full via subprocess. {e}")
    # We allow the code to continue execution but the model will likely fail to load later
    pass # Continue execution, but model will likely fail to load


# --- 1. Load Custom Model Utilities (Must come after mmcv is installed) ---
try:
    from mmseg.apis import init_segmentor, inference_segmentor 
except Exception as e:
    print(f"FATAL ERROR: Failed to import mmseg utilities: {e}")
    # Returning None here will trigger the "Error: Model failed to load" message in the app.
    init_segmentor = None
    inference_segmentor = None


# --- 2. CONFIGURATION ---
WEIGHTS_PATH = "R50_ReLeM.pth"
CONFIG_FILE = "configs/foodnet/SETR_Naive_768x768_80k_base_RM.py" 
DEVICE = 'cuda:0' if torch.cuda.is_available() else 'cpu'

# --- 3. Model Loading Function ---
@torch.no_grad()
def load_relem_model():
    """Initializes the segmentation model and loads the pre-trained weights."""
    if init_segmentor is None:
        return None # Skip if imports failed

    try:
        model = init_segmentor(
            CONFIG_FILE,             
            checkpoint=WEIGHTS_PATH,             
            device=DEVICE
        )
        model.eval()
        print(f"ReLeM Model loaded successfully onto {DEVICE}!")
        return model
    except Exception as e:
        print(f"CRITICAL ERROR: Model failed to load weights or config: {e}")
        traceback.print_exc()
        return None

RELEM_MODEL = load_relem_model()


# --- 4. Inference Function for Gradio ---
def segment_food(input_image: Image.Image):
    """Takes a PIL Image, runs inference, and returns a colorful segmentation mask."""
        
    if RELEM_MODEL is None:
        return "Error: Model failed to load at startup. Check build logs for reason."

    try:
        # Step 1: Save input image temporarily (Required by mmseg's inference pipeline)
        temp_path = "/tmp/input_img.png"
        input_image.save(temp_path)
        
        # Step 2: Run Inference (Produces the raw class ID map)
        result = inference_segmentor(RELEM_MODEL, temp_path)
        
        # Step 3: Post-process the result into a COLORFUL image
        seg_mask_array = result[0]
        
        # --- MATPLOTLIB VISUALIZATION (Robust Color Mask) ---
        fig, ax = plt.subplots(figsize=(8, 8))
        ax.imshow(seg_mask_array, cmap='nipy_spectral', interpolation='nearest')
        ax.axis('off')
        
        # Save the figure to an in-memory buffer
        buf = io.BytesIO()
        plt.savefig(buf, format='png', bbox_inches='tight', pad_inches=0)
        plt.close(fig)
        buf.seek(0)
        
        return Image.open(buf)
    
    except Exception as e:
        print(f"RUNTIME CRASH: Inference failed with error: {e}")
        traceback.print_exc()
        return f"Inference failed at runtime. Error: {e}. Try a smaller image."

# --- 5. GRADIO INTERFACE ---
gr.Interface(
    fn=segment_food,
    inputs=gr.Image(type="pil", label="Upload Food Image"),
    outputs=gr.Image(type="pil", label="ReLeM Segmentation Mask"),
    title="ReLeM (FoodSeg103) Deployment Final Attempt",
    description="Custom deployment of the ReLeM PyTorch model. Check logs for deployment status.",
    allow_flagging="never"
).launch()