Download code/runner/check_overlay.py from Noctis12138/dexhand-train-kit: direct link, hf CLI and curl.
- Browser
- Download file 1.28 kB
-
https://huggingface.co/Noctis12138/dexhand-train-kit/resolve/main/code/runner/check_overlay.py
- Command line
-
hf download hf://Noctis12138/dexhand-train-kit/code/runner/check_overlay.py
-
curl -L -o check_overlay.py https://huggingface.co/Noctis12138/dexhand-train-kit/resolve/main/code/runner/check_overlay.py
1.28 kB
| """Check that sub-task samples carry exact masks and that MaskOverlay changes the images.""" | |
| import os, sys, argparse, pathlib | |
| import numpy as np | |
| SYT = pathlib.Path(os.environ["SYT"]); sys.path.insert(0, str(SYT / "code/aug")); sys.path.insert(0, str(SYT / "code/runner")) | |
| import run as R | |
| from openpi.training import data_loader as dl | |
| cfg = R.build_config(argparse.Namespace(kind="subtask", repo_id=sys.argv[1]), batch_size=8, num_workers=0) | |
| dc = cfg.data.create(cfg.assets_dirs, cfg.model) | |
| ds = dl.create_torch_dataset(dc, cfg.model.action_horizon, cfg.model) | |
| overlay = dc.repack_transforms.inputs[0] | |
| print("overlay transform:", type(overlay).__name__) | |
| for i in (100, 5000, 20000): | |
| r = ds[i] | |
| print(f"sample {i}: mask keys {[k for k in r if 'mask' in k]} prompt={r.get('prompt')!r}") | |
| before = {k: np.asarray(r[k]).copy() for k in r if k.startswith("observation.images.")} | |
| msum = {k: float(np.asarray(r[k]).astype(np.float32).sum()) for k in r if "mask" in k} | |
| out = overlay(dict(r)) | |
| for k, b in before.items(): | |
| cam = k.split(".")[-1] | |
| d = np.abs(np.asarray(out[k]).astype(np.float32) - b.astype(np.float32)).mean() | |
| print(f" {cam}: target in exact mask={msum.get('observation.masks.' + cam, 0) > 0} mean|diff| after overlay={d:.3f}") | |