src/scripts/test_local_embedding.py

from pathlib import Path

import numpy as np
from PIL import Image
from sentence_transformers import SentenceTransformer

from utils.utils import SRC_PATH
from config import cnf

IMAGE_DIR = SRC_PATH / "data" / "images"

image_paths = sorted(IMAGE_DIR.glob("*.jpg"))

if len(image_paths) < 2:
    raise FileNotFoundError(f"Need at least 2 JPG images in {IMAGE_DIR}")

model = SentenceTransformer(
    str(cnf.local_emb_path),
    device="cpu",
    local_files_only=True,
)

images = [Image.open(path).convert("RGB") for path in image_paths]

embeddings = model.encode(
    images,
    batch_size=16,
    normalize_embeddings=True,
    convert_to_numpy=True,
    show_progress_bar=True,
)

# Normalized embeddings: dot product == cosine similarity
similarities = embeddings @ embeddings.T

names = [path.stem[:8] for path in image_paths]
heat_chars = " .:-=+*#%@"

print("\nImages:")
for i, path in enumerate(image_paths):
    print(f"{i:2}: {path.name}")

print("\nCLI similarity heatmap:")
print("    " + " ".join(f"{i:2}" for i in range(len(image_paths))))

for i, row in enumerate(similarities):
    cells = []

    for score in row:
        # Map typical cosine range 0–1 to an ASCII character
        normalized = float(np.clip(score, 0, 1))
        index = round(normalized * (len(heat_chars) - 1))
        cells.append(f" {heat_chars[index]}")

    print(f"{i:2} |" + "".join(cells))

print("\nNearest neighbour for each image:")

for i, row in enumerate(similarities):
    row = row.copy()
    row[i] = -np.inf

    best_index = int(np.argmax(row))

    print(
        f"{image_paths[i].name:16} -> "
        f"{image_paths[best_index].name:16} "
        f"{row[best_index]:.3f}"
    )