hc-mlx house classifier

A tiny MLX model that looks at a house photo and guesses one of five types: apartment, bungalow, villa, mansion, or haveli. Validation accuracy is 0.634 on 5,200 images (random guessing would get 0.20). A toy, built for fun.

Run it

You need a Mac with Apple silicon, Python 3.14, and the mlx, numpy, pillow, and huggingface_hub packages.

import numpy as np
from huggingface_hub import hf_hub_download
from PIL import Image
import mlx.core as mx
from model import HouseCNN  # model.py in this repo

ckpt = hf_hub_download("RanveerSoni98/hc-mlx", "house_cnn_best.safetensors")
model = HouseCNN(num_classes=5)
model.load_weights(ckpt)
model.eval()

img = Image.open("photo.jpg").convert("RGB").resize((128, 128))
x = (np.asarray(img, dtype=np.float32) / 255.0 - np.array([0.485, 0.456, 0.406])) / np.array([0.229, 0.224, 0.225])
probs = mx.softmax(model(mx.array(x[None])), axis=-1)[0]
mx.eval(probs)
labels = ["apartment", "bungalow", "villa", "mansion", "haveli"]
print(sorted(zip(labels, probs.tolist()), key=lambda t: -t[1])[:3])

Or with the project CLI: hc-mlx predict photo.jpg --repo RanveerSoni98/hc-mlx.

Training data

Places365 images from the Hub, mapped onto five house types (apartment, bungalow from house and cottage, villa from beach house, mansion, haveli from palace). See the hc-mlx project README for the full story.