""" app_core/preprocessing.py Lightweight, deterministic preprocessing for app inference. Deliberately does NOT import from training/ scripts. This is a self-contained preprocessing path for uploaded images. """ from pathlib import Path from typing import Tuple, Union import numpy as np import torch from PIL import Image from app_core.config import IMAGE_SIZE def load_image(source: Union[str, Path, np.ndarray]) -> np.ndarray: """ Load an image from a file path or numpy array. Returns a float32 numpy array of shape [H, W] normalised to [0, 1]. """ if isinstance(source, np.ndarray): img = source.copy().astype(np.float32) if img.ndim == 3: # Take luminance: if RGB/RGBA, convert to grayscale if img.shape[2] == 4: img = img[:, :, :3] img = 0.299*img[:,:,0] + 0.587*img[:,:,1] + 0.114*img[:,:,2] else: pil = Image.open(str(source)).convert("L") # grayscale img = np.array(pil, dtype=np.float32) # Normalise to [0, 1] mn, mx = img.min(), img.max() if mx > mn: img = (img - mn) / (mx - mn) else: img = np.zeros_like(img) return img def resize_image(img: np.ndarray, size: int = IMAGE_SIZE) -> np.ndarray: """Resize a 2D float array to [size, size] using bilinear interpolation.""" pil = Image.fromarray((img * 255).astype(np.uint8)) pil = pil.resize((size, size), Image.BILINEAR) return np.array(pil, dtype=np.float32) / 255.0 def to_display(img: np.ndarray) -> np.ndarray: """Convert normalised float [H,W] to uint8 [H,W,3] for Gradio display.""" img8 = (img * 255).clip(0, 255).astype(np.uint8) return np.stack([img8, img8, img8], axis=2) def to_tensor(img: np.ndarray) -> torch.Tensor: """ Convert normalised float [H,W] or [H,W,C] numpy image to model-ready tensor [1, 1, H, W] on CPU. """ if img.ndim == 3: img = 0.299*img[:,:,0] + 0.587*img[:,:,1] + 0.114*img[:,:,2] return torch.tensor(img, dtype=torch.float32).unsqueeze(0).unsqueeze(0) def prepare( source: Union[str, Path, np.ndarray] ) -> Tuple[np.ndarray, np.ndarray, torch.Tensor]: """ Full preprocessing pipeline. Args: source: file path or numpy array Returns: (display_img, model_img, tensor) display_img : uint8 [H, W, 3] — for Gradio Image component model_img : float32 [H, W] — 256×256 normalised tensor : float32 [1,1,H,W] — model-ready """ raw = load_image(source) resized = resize_image(raw, IMAGE_SIZE) display = to_display(resized) tensor = to_tensor(resized) return display, resized, tensor