---
license: mit
tags:
- sign-language
- motion
- rvq-vae
- momask
library_name: pytorch
---

# Signvrse MoMask Motion RVQ-VAE

MoMask-style residual vector-quantized VAE trained on Signvrse motion features.

## Model

| Field | Value |
|-------|-------|
| Training step | 200,000 |
| Input dim | 668 |
| Latent dim | 512 |
| Codebooks | 6 × 512 |
| Window length | 64 frames |
| Downsample factor | 4 |

## Files

- `model.safetensors` — model weights only (for inference)
- `checkpoint.pt` — full training checkpoint (optimizer + scheduler + args)
- `config.json` — architecture hyperparameters
- `training_metadata.json` — training step and CLI args

## Loading

```python
import json
import torch
from safetensors.torch import load_file
from huggingface_hub import hf_hub_download
from rvqvae import MotionResidualRVQVAE

repo_id = "Signvrse/signvrse-momask-rvqvae"
config_path = hf_hub_download(repo_id, "config.json")
weights_path = hf_hub_download(repo_id, "model.safetensors")

with open(config_path) as f:
    cfg = json.load(f)

model = MotionResidualRVQVAE(
    input_dim=cfg["input_dim"],
    latent_dim=cfg["latent_dim"],
    num_codebooks=cfg["num_codebooks"],
    codebook_size=cfg["codebook_size"],
    quant_dropout_prob=cfg["quant_dropout_prob"],
    ema_decay=cfg["ema_decay"],
)
model.load_state_dict(load_file(weights_path))
model.eval()

# x: [batch, 668, seq_len] float32 motion features
# out = model(x)
```

Install the model code from this repository's `rvqvae` package, or copy `rvqvae/momask_rvqvae.py` into your project.