"""Tiny deterministic stand-ins for the real models.""" from __future__ import annotations import hashlib import numpy as np from PIL import Image from drape.models import l2_normalize DIM = 48 class FakeEmbedder: """Colour-histogram 'embedding'. Same colours -> close vectors.""" name = "fake" dim = DIM def embed_images(self, images): out = [] for im in images: arr = np.asarray(im.convert("RGB").resize((32, 32)), dtype=np.float32) / 255.0 pix = arr.reshape(-1, 3) fg = pix[(pix < 0.94).any(axis=1)] if len(fg) == 0: fg = pix hist, _ = np.histogramdd(fg, bins=(3, 3, 3), range=((0, 1), (0, 1), (0, 1))) vec = np.concatenate([hist.ravel() / len(fg), fg.mean(axis=0), np.zeros(DIM - 30)]) out.append(vec) return l2_normalize(np.stack(out)) if out else np.zeros((0, DIM), np.float32) def embed_texts(self, texts): out = [] for t in texts: seed = int(hashlib.md5(t.encode()).hexdigest()[:8], 16) out.append(np.random.default_rng(seed).normal(size=DIM)) return l2_normalize(np.stack(out)) if out else np.zeros((0, DIM), np.float32) class FakeSegmenter: """Returns a fixed label map, resized to the image.""" def __init__(self, label_map: np.ndarray): self.label_map = label_map def segment(self, image): lm = Image.fromarray(self.label_map.astype(np.uint8)) return np.asarray(lm.resize(image.size, Image.NEAREST)) def solid(colour, size=(64, 64)) -> Image.Image: return Image.new("RGB", size, colour) def product(colour, size=(64, 64)) -> Image.Image: """A coloured square on white, like a product shot.""" im = Image.new("RGB", size, (255, 255, 255)) w, h = size im.paste(Image.new("RGB", (w // 2, h // 2), colour), (w // 4, h // 4)) return im