`data_features_name_mapping` doesn't work

[shuheng-liu]

System Info

N/A

Information

• One of the scripts in the src/opentau/scripts/ folder of OpenTau
• My own task or dataset (give details below)

Reproduction

Remove the following section from standard_data_format_mapping.py

"physical-intelligence/libero": {
    "camera0": "image",
    "camera1": "wrist_image",
    "state": "state",
    "actions": "actions",
    "prompt": "task",
    "response": "response",
},

and add the following to pi05_training_config.json

"data_features_name_mapping": {
    "camera0": "image",
    "camera1": "wrist_image",
    "state": "state",
    "actions": "actions",
    "prompt": "task",
    "response": "response"
}

and train with

accelerate launch src/opentau/scripts/train.py --config_path=configs/examples/pi05_training_config.json

This results in

Traceback (most recent call last):
  File "/path/to/OpenTau/src/opentau/scripts/train.py", line 456, in <module>
    train()
  File "/path/to/OpenTau/src/opentau/configs/parser.py", line 388, in wrapper_inner
    response = fn(cfg, *args, **kwargs)
  File "/path/to/OpenTau/src/opentau/scripts/train.py", line 183, in train
    train_dataset, val_dataset = make_dataset_mixture(cfg)
  File "/path/to/OpenTau/src/opentau/datasets/factory.py", line 247, in make_dataset_mixture
    res = make_dataset(dataset_cfg, cfg, return_advantage_input=return_advantage_input)
  File "/path/to/OpenTau/src/opentau/datasets/factory.py", line 193, in make_dataset
    (dt_mean, dt_std, dt_lower, dt_upper), f2g = resolve_delta_timestamps(train_cfg, cfg, ds_meta)
  File "/path/to/OpenTau/src/opentau/datasets/factory.py", line 128, in resolve_delta_timestamps
    name_map = DATA_FEATURES_NAME_MAPPING[dataset_cfg.repo_id]
KeyError: 'physical-intelligence/libero'

Expected behavior

It should work without an error as dataset config should update the preset DATA_FEATURES_NAME_MAPPING registry.

[FishofJLlake]

I also named the data format based on my own dataset, and there were no errors. Please check if the repo_id is configured correctly in the configuration file.

[shuheng-liu]

The idea is that one should be able to set any repo_id (as long as hosted by and accessible via huggingface) in the training config.json, even those that are not included in opentau.datasets.standard_data_format_mapping.DATA_FEATURES_NAME_MAPPING, provided the data feature mapping is specified.

[FishofJLlake]

i get it, thanks