File size: 2,094 Bytes
8c2f5a9 4aae487 8c2f5a9 4aae487 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 | ---
license: apache-2.0
library_name: pytorch
pipeline_tag: feature-extraction
tags:
- tactile-sensing
- feature-extraction
- robotics
- pytorch
- convnextv2
---
# SharpaWave Deform Encoder
`DeformEncoder` converts a preprocessed single-channel scalar deformation image
into a compact learned tactile feature with shape `[B, 512, 1, 1]`. The feature
can be flattened to `[B, 512]` for downstream tasks.
The unified checkpoint also includes `DeformDecoder` parameters. The decoder
and `DeformAutoencoder` are provided only to demonstrate deformation
reconstruction from the compact feature.
## Tensor Shapes
| Operation | Input | Output |
| --- | --- | --- |
| Encoder | `[B, 1, 240, 240]` | `[B, 512, 1, 1]` |
| Flatten feature | `[B, 512, 1, 1]` | `[B, 512]` |
| Autoencoder | `[B, 1, 240, 240]` | `[B, 1, 240, 240]` |
The input is a preprocessed scalar deformation image, not a raw RGB camera
image.
## Usage
Download the checkpoint and use `load_encoder()` from the source repository:
```python
import torch
from huggingface_hub import hf_hub_download
from sharpawave_deform_encoder import load_encoder
checkpoint = hf_hub_download(
repo_id="Sharpa-Robotics/sharpawave-deform-encoder",
filename="sharpawave_deform_autoencoder.safetensors",
)
encoder = load_encoder(checkpoint, "cpu")
deform = torch.zeros(1, 1, 240, 240)
with torch.inference_mode():
feature = encoder(deform) # [1, 512, 1, 1]
```
Source code: <https://github.com/sharpa-robotics/sharpawave-deform-encoder>
## Checkpoint
The single SafeTensors file contains both encoder and reconstruction-only
decoder parameters. `load_encoder()` reads only the encoder tensors;
`load_autoencoder()` loads the complete demonstration model.
The published checkpoint was trained from random initialization without
upstream pretrained weights.
The SHA-256 digest is recorded in `SHA256SUMS`.
## Limitations
The encoder expects the documented 240-by-240 scalar input representation.
## License
Developed by Sharpa Group. Licensed under Apache License 2.0. See `LICENSE` and
`THIRD_PARTY_NOTICES.md`.
|