# tactile_vae

ViT-based tactile variational autoencoder.

## Architecture
- Encoder: `PatchEmbed + Transformer blocks + fixed sin-cos positional embedding`
- Latent: `mu/logvar` + reparameterization
- Decoder: latent-conditioned transformer patch decoder + unpatchify

## Usage

```python
import torch
from tactile_vae.model import TactileVAE, VAELoss

model = TactileVAE()
out = model(torch.randn(2, 3, 128, 128))
loss_fn = VAELoss(beta=1.0)
losses = loss_fn(out["x_hat"], torch.randn(2, 3, 128, 128), out["mu"], out["logvar"])
```