This repo contains scripts for modifying existing RLDS datasets. By running modify_rlds_dataset.py, you will load an existing RLDS dataset, apply the specified modifications to each sample, reshard the resulting dataset and store it in a new directory. Apart from a number of simple modification functions, this repo implements a parallelized AdhocTFDSBuilder that can perform the data modifications in parallel for increased conversion speed.

First create a conda environment using the provided environment.yml file (use environment_ubuntu.yml or environment_macos.yml depending on the operating system you're using):

conda env create -f environment_ubuntu.yml

Then activate the environment using:

conda activate rlds_env

If you want to manually create an environment, the key packages to install are tensorflow and tensorflow_datasets.

To download datasets from the Open X-Embodiment Dataset Google cloud bucket, please install gsutil using the installation instructions.

The command below resizes all RGB and depth images to a max. size of 336 and encodes RGB images as jpeg. This can e.g. be useful for reducing the storage size of datasets in the Open X-Embodiment Dataset.

python3 modify_rlds_dataset.py --dataset=<name_of_your_tfds_dataset> --mods=resize_and_jpeg_encode --target_dir=<path_where_mod_data_is_written>

This creates a new dataset with smaller, jpeg encoded images in the target_dir.

You can switch out the resize_and_jpeg_encode mod for other functions in mod_functions.py.

The modify_rlds_dataset.py script supports the following command line arguments:

modify_rlds_dataset.py:
  --data_dir: Directory where source data is stored.
  --dataset: Dataset name.
  --max_episodes_in_memory: Number of episodes converted & stored in memory before writing to disk.
    (default: '100')
    (an integer)
  --mods: List of modification functions, applied in order.
    (a comma separated list)
  --n_workers: Number of parallel workers for data conversion.
    (default: '10')
    (an integer)
  --target_dir: Directory where modified data is stored.

You can increase the n_workers and max_episodes_in_memory parameters based on the resources of your machine. The larger the respective value, the faster the dataset conversion.

A list of all supported dataset modifications ("mods") can be found in mod_functions.py.

You can add your own custom modification functions in mod_functions.py by implementing the TfdsModFunction interface. Your mod function needs to provide one function that modifies the dataset feature spec and one map function that modifies an input tf.data.Dataset. You can use the existing mod functions as examples. Make sure to register your new mod in the TFDS_MOD_FUNCTIONS object.

To download the Open X-Embodiment dataset and convert it for training, run bash prepare_open_x.sh. You can specify the download directory at the top of the script.

• No new tempfile could be created: The script stores large datasets in intermediate temporary files in the \tmp directory. Depending on the dataset size it can store up to 1000 such temp files. The default number of files openable in parallel in Ubuntu is 1024, so this limit can lead to the error above. You can increase the limit by running ulimit -n 200000.