loopback-twotower

loopback — two-tower music recommender

Open-source two-tower neural recommender for music, trained from scratch on the Last.fm 1K users dataset. Repo: https://github.com/DanielRegaladoUMiami/loopback.

Architecture

User tower:  user_id  ──► Embedding(64) ──► MLP(256→128) ──► L2-norm ──► user_vec
Track tower: track_id ──► Embedding(64) ┐
             artist_id ─► Embedding(64) ┴► MLP(256→128) ──► L2-norm ──► track_vec
                                          score = u · t * exp(temp)

Loss: symmetric InfoNCE (CLIP-style) with in-batch negatives and a learnable temperature.

Training

• 3 epochs, batch size 4096, AdamW lr=1e-3, weight decay 1e-5
• 15.3 M training interactions (992 users × 1.5 M unique tracks)
• Apple M-series MPS, ~9 min / epoch
• Final loss: 5.6 (random baseline at this batch size: ln(4096) ≈ 8.32)

Results

Evaluated on 847 held-out users with seen-track filtering against the full 1.5 M-track catalog:

Usage

import torch
from huggingface_hub import hf_hub_download
from loopback.model import TwoTower  # from github.com/DanielRegaladoUMiami/loopback

ckpt = torch.load(hf_hub_download("DanielRegaladoCardoso/loopback-twotower", "two_tower_epoch3.pt"),
                  map_location="cpu", weights_only=False)
model = TwoTower(992, 1_500_661, 174_091, out_dim=ckpt["embed_dim"])
model.load_state_dict(ckpt["model"])
model.eval()

License

Apache 2.0