ta-jepa โ Phase 1 causal audio JEPA (FMA + FSD50K, grounded)
A causal, latent world model for general audio. From past context only it predicts future audio representations in embedding space โ the audio analogue of a causal V-JEPA, not a static representation learner.
- Input: EnCodec 24 kHz continuous pre-quantizer embeddings (75 Hz, dim 128).
- Architecture: causal transformer encoder
f_ฮธ+ EMA target encoderf_ฮธฬ+ causal multi-offset predictorg_ฯ; loss = latent smooth-L1 vs stop-grad EMA target + VICReg variance/covariance (anti-collapse). Offsets 1/2/4/8. - Training data: FMA-small (music) + FSD50K (general audio).
- Result: on held-out ESC-50 it beats latent persistence at every horizon and
matches APC on transfer; no collapse (effective rank ~226/256). The forecasting-vs-horizon
metric โ not the std-pooling linear probe โ is the right yardstick for a world model
(the causal objective makes the latent temporally smooth, which the probe penalizes).
See the repo's
RESULTS.md.
This file: jepa_fma_grounded.ckpt โ a PyTorch Lightning checkpoint (state dict +
hyper-parameters). ~103 MB.
Usage (Lightning-free load)
import copy, torch
from huggingface_hub import hf_hub_download
from tajepa.models.jepa import JEPA # pip install "tajepa @ git+https://github.com/thias42/ta-jepa.git"
ck = torch.load(hf_hub_download("Maeich/ta-jepa-anticipation", "jepa_fma_grounded.ckpt"),
map_location="cpu", weights_only=False)
hp, sd = ck["hyper_parameters"], ck["state_dict"]
jepa = JEPA(in_dim=hp["in_dim"], dim=hp["dim"], enc_depth=hp["enc_depth"],
pred_depth=hp["pred_depth"], heads=hp["heads"],
offsets=tuple(hp["offsets"]), dropout=hp.get("dropout", 0.0))
jepa.load_state_dict({k[5:]: v for k, v in sd.items() if k.startswith("jepa.")}, strict=False)
target = copy.deepcopy(jepa.encoder) # EMA target encoder (prediction targets)
target.load_state_dict({k[7:]: v for k, v in sd.items() if k.startswith("target.")})
jepa.eval(); target.eval()
Links
- Code: https://github.com/thias42/ta-jepa
- Interactive demo (Space): https://huggingface.co/spaces/Maeich/ta-jepa-anticipation-demo
License
MIT.