Tr0612's picture
Add files using upload-large-folder tool
271ac99 verified
# Button-Press Mini Demo
This folder is a tiny, self-contained test set for `button-press-topdown-v3`.
## Contents
- `samples.jsonl`: 5 demo samples (image path, instruction, state, gt_action)
- `images/`: copied JPEG frames for those samples
- `manifest.json`: dataset metadata
- `inference_results.jsonl`: created after running the demo script
## Run Inference on All 4 Shared Checkpoints
From `vla_project/` root:
```bash
./.venv/bin/python demo/run_button_press_demo.py
```
By default this runs on `1` sample for quick verification.
This runs `infer.py` on:
- `checkpoints/checkpoints_stage2_no_moe_unfreeze/best.pt`
- `checkpoints/checkpoints_stage2_moe_text_unfreeze/best.pt`
- `checkpoints/checkpoints_stage2_moe_full_unfreeze/best.pt`
- `checkpoints/checkpoints_stage2_moe_full_peg_only/best.pt`
and writes:
- `demo/button_press_mini/inference_results.jsonl`
## Optional: Run All 5 Samples
```bash
./.venv/bin/python demo/run_button_press_demo.py --max-samples 0
```