mwnuk's picture
Add sequence-to-sequence checkpoints
dd3c601 verified
|
Raw History Blame Contribute Delete
1.19 kB
---
license: mit
tags:
- tutorial
- robotics
- pytorch
---
# Robot learning guide models
Small checkpoints trained by the notebooks at
[github.com/wnukmat/robot-learning-notebooks](https://github.com/wnukmat/robot-learning-notebooks), which accompany the guide at
[matthewwnuk.com/guide.html](https://www.matthewwnuk.com/guide.html).
They exist so that the notebooks run in seconds without training. Each notebook has a `TRAIN = True` switch that reproduces
its checkpoints from scratch. These are teaching models on toy problems, not models for use elsewhere.
| Files | Trained by | What they are |
|---|---|---|
| `seq2seq/bottleneck.pt`, `attention.pt`, `transformer.pt` | `1_foundations/05_sequence_to_sequence` | Three sequence-reversal models: a GRU encoder-decoder (125k parameters), the same with attention (155k) and a small transformer (530k) |
| `generative/flow.pt`, `diffusion.pt` | `2_models/04_flow_matching_vs_diffusion` | Two small MLPs trained on a 2D two-crescent dataset, one with a flow-matching loss and one with a diffusion loss |
Load them through the helper package:
```python
from guide_utils import data
path = data.fetch_model("seq2seq/attention.pt")
```