app.py

import gradio as gr
import torch
from PIL import Image
import numpy as np

# --- 1. Load Custom Model Utilities ---
# NOTE: These imports MUST match the files you copied from the GitHub repo.
# Example imports - adjust these if the model files are deeper in subfolders!
try:
    from mmseg.apis import init_segmentor, inference_segmentor # Core MMSeg functions
    from mmseg.datasets import build_dataloader, build_dataset # Utilities
    # You might also need to copy config files, e.g., to 'configs/relem/'
except ImportError:
    print("MMSegmentation utilities not found. Ensure files were copied correctly.")


# --- 2. CONFIGURATION ---
# Define the paths for the files you placed in the repository
WEIGHTS_PATH = "R50_ReLeM.pth"
CONFIG_FILE = "configs/foodnet/SETR_Naive_768x768_80k_base_RM.py" # Replace with actual config file from the repo

# --- 3. Model Loading Function ---
@torch.no_grad()
def load_relem_model():
    """Initializes the segmentation model and loads the pre-trained weights."""
    try:
        # 1. Initialize the segmentor using MMSegmentation's utility
        # This requires the config file and the checkpoint path
        model = init_segmentor(
            CONFIG_FILE, 
            checkpoint=WEIGHTS_PATH, 
            device='cuda:0' if torch.cuda.is_available() else 'cpu'
        )
        model.eval()
        print("ReLeM Model loaded successfully!")
        return model
    except Exception as e:
        print(f"Error loading model: {e}")
        # Return a flag if loading fails
        return None

# Load the model once when the Space starts
RELEM_MODEL = load_relem_model()


# --- 4. Inference Function for Gradio ---
def segment_food(input_image: Image.Image):
    """Takes a PIL Image and returns a segmentation mask image."""
    
    if RELEM_MODEL is None:
        return "Error: Model failed to load. Check logs for details."

    try:
        # Use MMSegmentation's inference pipeline 
        # The input is usually a filepath, so we need to save and then load
        
        # 1. Save input image temporarily
        temp_path = "/tmp/input_img.png"
        input_image.save(temp_path)
        
        # 2. Run Inference
        result = inference_segmentor(RELEM_MODEL, temp_path)
        
        # 3. Post-process the result (usually a numpy array) into a color mask image
        # The result is a segmentation map (array of class IDs).
        # We use a simple utility to convert the ID map to a visible color mask.
        seg_mask_array = result[0] 
        color_mask = Image.fromarray(seg_mask_array.astype(np.uint8)).convert("L") 
        # NOTE: Full color mapping requires the class labels/palette, which you must also copy from the repo.
        
        return color_mask 
    
    except Exception as e:
        return f"Inference failed: {e}"

# --- 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) Segmentation Demo",
    description="Custom deployment of the ReLeM PyTorch model. **NOTE:** Model loading requires the full code/config structure from the GitHub repo.",
    allow_flagging="never"
).launch()