"""Next Match: search a catalog with a reference image plus a text edit ("bunun kirmizi olani", "same but striped"). Shared by the notebook, inference.py and (ported) web/core.js: * tokenize: Turkish-aware lower-casing, punctuation removal, longest-prefix stem matching against a fixed vocabulary * encode_query: image features + text features -> fused, L2-normalised query vector * encode_item: catalog image -> L2-normalised key vector The catalog is synthetic: procedurally drawn product icons (8 categories x 8 colours x 3 patterns), rendered in the notebook. """ import numpy as np from nlrt import forward CATEGORIES = ["tshirt", "mug", "sneaker", "bag", "hat", "chair", "lamp", "bottle"] COLORS = ["red", "blue", "green", "yellow", "black", "white", "purple", "orange"] PATTERNS = ["plain", "striped", "dotted"] RGB = {"red": (214, 40, 40), "blue": (40, 90, 210), "green": (40, 160, 70), "yellow": (240, 200, 30), "black": (30, 30, 34), "white": (250, 250, 250), "purple": (130, 60, 170), "orange": (245, 130, 30)} TR = {"red": "kırmızı", "blue": "mavi", "green": "yeşil", "yellow": "sarı", "black": "siyah", "white": "beyaz", "purple": "mor", "orange": "turuncu", "plain": "düz", "striped": "çizgili", "dotted": "puantiyeli", "tshirt": "tişört", "mug": "kupa", "sneaker": "ayakkabı", "bag": "çanta", "hat": "şapka", "chair": "sandalye", "lamp": "lamba", "bottle": "şişe"} EN = {"tshirt": "t-shirt", "mug": "mug", "sneaker": "sneaker", "bag": "bag", "hat": "hat", "chair": "chair", "lamp": "lamp", "bottle": "bottle", "plain": "plain", "striped": "striped", "dotted": "dotted"} SYNONYMS = {"tshirt": ["tshirt", "shirt", "tee"], "dotted": ["dotted", "dots", "polka"], "striped": ["striped", "stripes", "stripe"], "sneaker": ["sneaker", "shoe", "shoes"], "plain": ["plain", "solid"]} def lower_tr(s): return s.replace("I", "ı").replace("İ", "i").lower() def tokenize(text, vocab): """Return token ids (unknown words dropped). Matches whole words, then the longest vocabulary prefix of >= 4 letters.""" clean = "".join(ch if (ch.isalnum() or ch == " ") else " " for ch in lower_tr(text).replace("-", "")) ids = [] for w in clean.split(): if w in vocab: ids.append(vocab[w]) continue best = None for k in range(len(w) - 1, 3, -1): if w[:k] in vocab: best = vocab[w[:k]] break if best is not None: ids.append(best) return ids def text_vector(ids, emb): if not ids: return np.zeros(emb.shape[1]) return np.mean(emb[ids], axis=0) def l2n(x): return x / max(float(np.linalg.norm(x)), 1e-12) def image_features(spec, images, net="img"): """images: (N, 3, 48, 48) floats in [0, 1].""" return forward(spec["nets"][net], np.asarray(images, dtype=np.float64)) def encode_item(spec, images, net="img"): f = image_features(spec, images, net) return f / np.maximum(np.linalg.norm(f, axis=1, keepdims=True), 1e-12) def encode_query(spec, image, text, mode="combined"): """mode: combined | image | text. image: (3, 48, 48) in [0, 1].""" vocab = spec["meta"]["vocab"] if mode == "combined": fi = image_features(spec, image[None], "img")[0] ft = forward(spec["nets"]["txt"], text_vector(tokenize(text, vocab), spec["tensors"]["emb"])[None])[0] return l2n(fi + forward(spec["nets"]["fuse"], np.concatenate([fi, ft])[None])[0]) if mode == "image": return l2n(forward(spec["nets"]["img_only_head"], image_features(spec, image[None], "img_only")[0][None])[0]) ft = forward(spec["nets"]["txt_only"], text_vector(tokenize(text, vocab), spec["tensors"]["emb_only"])[None])[0] return l2n(ft)