src/utils.py

from __future__ import annotations
import numpy as np
from PIL import Image

def pil_to_np(img: Image.Image) -> np.ndarray:
    """Convert a PIL image into a float32 NumPy array in the [0, 1] range."""
    if img.mode not in ("RGB", "RGBA", "L"):
        img = img.convert("RGB")
    if img.mode == "L":
        img = img.convert("RGB")
    arr = np.asarray(img).astype(np.float32)
    if arr.ndim == 2:
        arr = np.repeat(arr[..., None], 3, axis=2)
    if arr.shape[2] == 4:
        arr = arr[..., :3]
    return arr / 255.0

def np_to_pil(arr: np.ndarray) -> Image.Image:
    """Convert a float32 NumPy array (0–1) into an RGB PIL image."""
    return Image.fromarray(np.clip(arr * 255.0, 0, 255).astype(np.uint8))

def resize_and_crop_to_grid(img: Image.Image, width: int, height: int, grid: int) -> Image.Image:
    """Resize and center-crop an image so both dimensions are multiples of the grid."""
    img = img.convert("RGB").resize((width, height), Image.LANCZOS)
    H, W = img.height, img.width
    H2, W2 = (H // grid) * grid, (W // grid) * grid
    if H2 != H or W2 != W:
        left = (W - W2) // 2
        top = (H - H2) // 2
        img = img.crop((left, top, left + W2, top + H2))
    return img

def block_view(arr: np.ndarray, bh: int, bw: int) -> np.ndarray:
    """Return a strided view that exposes the image as (grid_h, grid_w, bh, bw, C) blocks."""
    H, W, C = arr.shape
    if H % bh or W % bw:
        raise ValueError("Array dimensions must be divisible by the block size")
    shape = (H // bh, W // bw, bh, bw, C)
    strides = (arr.strides[0] * bh, arr.strides[1] * bw, arr.strides[0], arr.strides[1], arr.strides[2])
    return np.lib.stride_tricks.as_strided(arr, shape=shape, strides=strides)

def cell_means(arr: np.ndarray, grid: int) -> np.ndarray:
    """Return weighted mean RGB values for every grid cell using pure NumPy ops."""
    H, W, _ = arr.shape
    bh, bw = H // grid, W // grid
    blocks = block_view(arr, bh, bw)  # (grid, grid, bh, bw, 3)

    center_h = (bh - 1) / 2.0
    center_w = (bw - 1) / 2.0
    yy, xx = np.meshgrid(np.arange(bh), np.arange(bw), indexing="ij")
    dist = np.sqrt((yy - center_h) ** 2 + (xx - center_w) ** 2)
    max_dist = np.sqrt(center_h**2 + center_w**2) or 1.0
    weights = 1.0 - (dist / max_dist) * 0.5
    weights = weights.astype(np.float32)
    weights /= weights.sum()

    weighted = blocks * weights[None, None, :, :, None]
    return weighted.sum(axis=(2, 3))