Add demo for dataset generation.

d4rl-generation/.gitignore

@@ -0,0 +1,165 @@
+experiment_logs/
+datasets/
+logs/
+*.hdf5
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/

d4rl-generation/README.md

@@ -0,0 +1,54 @@
+# Kabuki Dataset generation.
+
+This repo serves to demonstrate a workflow for creating
+offline datasets compatible with Kabuki.
+
+It uses dm-acme for online RL training. The data
+is logged via EnvLogger, and the logged data is then post-processed
+into an HDF5 file. However, this is only a proof of concept, and 
+neither dm-acme nor EnvLogger constitutes part of the specification.
+
+## Spec for the new HDF5 dataset in Kabuki
+We describe the standard used in the HDF5 dataset used in D4RL-V2.
+Unlike the old datasets used in previous versions of the D4RL datasets,
+we now formalize a new standard for the new datasets.
+
+The new datasets will continue to use HDF5 as the storage format. Although HDF5 files
+do not naturally have the mechanisms for storing distinct episodes as separate entries,
+it is a widely adopted a standard format that can be easily used in different
+frameworks and languages.
+
+The previous iterations of the D4RL datasets have some outstanding issues.
+One notable issue is that terminal observations are not captured.
+While the omission of terminal observations are not really problematic
+for offline actor-critic algorithms such as CQL/BCQ/IQL etc.,
+they pose issues for researchers 
+who would like to work on offline imitation learning research.
+It is known that proper handling of terminal transitions can significantly influence the 
+performance of imitation learning algorithms.
+Therefore, the terminal observations will be recorded in the dataset in the new version. The new datasets should capture as much information from the original environment as possible whenever possible.
+
+In the new version, the dataset will follow the convention introduced by the RLDS project.
+The agent's experience is stored in the dataset as a sequence of episodes consisting of a variable number of steps. The steps are stored as a flattened dictionary of arrays
+in the state-action-reward (SAR) alignment. Concretely each step consists of
+
+* is_first, is_last: indicating the observation for the step is the first/the last step of an episode.
+* observation: observation for the step
+* action: action taken after observing the `observation` of the step
+* reward: reward obtained after applying the action in the step.
+* is_terminal: indicating whether the observation is terminal (is_terminal = False indicates that the episode is truncated.)
+* discount: discount factor at this step. This may be unfamiliar to gym.Env users but 
+is consistent with the discount used in dm_env. In particular, discount = 0 indicates that
+the *next* step is terminal and 1.0 otherwise.
+
+Refer to https://github.com/google-research/rlds for a more detailed description.
+
+## Generating datasets.
+While HDF5 is used as the final format for storing benchmark datasets in D4RL,
+HDF5 is not used as the format during the data collection process. In this repo,
+we demonstrate using EnvLogger for recording the interactions made by an
+RL agent during online learning. The logged experience will then be post-processed
+(and potentially stitched with other datasets) to produce the final HDF5 files.
+We provide `convert_dataset.py` to show how this can be done by converting
+from EnvLogger's Riegeli file formats to a single HDF5 file.
+Alternatively, we can also use EnvLogger's RLDS backend to generate an RLDS-compatible TensorFlow dataset and convert that to the HDF5 file.

d4rl-generation/convert_dataset.py

@@ -0,0 +1,77 @@
+"""Convert dataset logged by the EnvLogger Riegelli backend to HDF5."""
+from envlogger import reader
+import numpy as np
+import h5py
+import tree
+from absl import flags
+from absl import app
+
+_DATASET_DIR = flags.DEFINE_string("dataset_dir", None, "")
+_OUTPUT_FILE = flags.DEFINE_string("output_file", "dataset.hdf5", "")
+flags.mark_flag_as_required("dataset_dir")
+
+
+def _convert_envlogger_episode_to_rlds_steps(episode):
+    """Convert an episode of envlogger.StepData to RLDS compatible steps."""
+    observations = np.stack([step.timestep.observation for step in episode])
+    # RLDS uses the SAR alignment while envlogger uses ARS.
+    # The following lines handle converting from the ARS to SAR alignment.
+    actions = np.stack([step.action for step in episode[1:]])
+    # Add dummy action to the last step containing the terminal observation.
+    actions = np.concatenate(
+        [actions, np.expand_dims(np.zeros_like(actions[0]), axis=0)]
+    )
+    # Add dummy reward to the last step containing the terminal observation.
+    rewards = np.stack([step.timestep.reward for step in episode[1:]])
+    rewards = np.concatenate(
+        [rewards, np.expand_dims(np.zeros_like(rewards[0]), axis=0)]
+    )
+    # Add dummy discounts to the last step containing the terminal observation.
+    discounts = np.stack([step.timestep.reward for step in episode[1:]])
+    discounts = np.concatenate(
+        [discounts, np.expand_dims(np.zeros_like(discounts[0]), axis=0)]
+    )
+    # the is_first/last/terminal flags are already aligned in ARS alignment.
+    is_first = np.array([step.timestep.first() for step in episode])
+    is_last = np.array([step.timestep.last() for step in episode])
+    is_terminal = np.array(
+        [step.timestep.last() and step.timestep.discount == 0.0 for step in episode]
+    )
+    return {
+        "observation": observations,
+        "action": actions,
+        "reward": rewards,
+        "discounnts": discounts,
+        "is_first": is_first,
+        "is_last": is_last,
+        "is_terminal": is_terminal,
+    }
+
+
+def write_to_hdf5_file(episodes, filename):
+    """Write episodes in EnvLogger format to an HDF5 file."""
+    all_steps = []
+    for episode in episodes:
+        all_steps.append(_convert_envlogger_episode_to_rlds_steps(episode))
+    all_steps = tree.map_structure(lambda *xs: np.concatenate(xs), *all_steps)
+    f = h5py.File(filename, "w")
+    for key in all_steps.keys():
+        f.create_dataset(key, data=all_steps[key], compression="gzip")
+    f.close()
+
+
+def main(_):
+    output_file = _OUTPUT_FILE.value
+    with reader.Reader(data_directory=_DATASET_DIR.value) as r:
+        print(r.observation_spec())
+        print(r.metadata())
+        write_to_hdf5_file(r.episodes, output_file)
+    # Inspecting the created HDF5 file
+    f = h5py.File(output_file, "r")
+    for k in f:
+        print(k, f[k].shape)
+    f.close()
+
+
+if __name__ == "__main__":
+    app.run(main)

d4rl-generation/helpers.py

@@ -0,0 +1,69 @@
+"""Shared helpers for rl_continuous experiments."""
+from typing import Optional
+from acme import wrappers
+import dm_env
+import gym
+from acme.utils import loggers as acme_loggers
+from absl import logging
+
+_VALID_TASK_SUITES = ("gym", "control")
+
+
+def make_environment(suite: str, task: str, seed=None) -> dm_env.Environment:
+    """Makes the requested continuous control environment.
+    Args:
+      suite: One of 'gym' or 'control'.
+      task: Task to load. If `suite` is 'control', the task must be formatted as
+        f'{domain_name}:{task_name}'
+    Returns:
+      An environment satisfying the dm_env interface expected by Acme agents.
+    """
+
+    if suite not in _VALID_TASK_SUITES:
+        raise ValueError(
+            f"Unsupported suite: {suite}. Expected one of {_VALID_TASK_SUITES}"
+        )
+
+    if suite == "gym":
+        env = gym.make(task)
+        env.seed(seed)
+        # Make sure the environment obeys the dm_env.Environment interface.
+        env = wrappers.GymWrapper(env)
+
+    elif suite == "control":
+        # Load dm_suite lazily not require Mujoco license when not using it.
+        from dm_control import suite as dm_suite  # pylint: disable=g-import-not-at-top
+
+        domain_name, task_name = task.split(":")
+        env = dm_suite.load(domain_name, task_name, task_kwargs={'random': seed})
+        env = wrappers.ConcatObservationWrapper(env)
+
+    # Wrap the environment so the expected continuous action spec is [-1, 1].
+    # Note: this is a no-op on 'control' tasks.
+    env = wrappers.CanonicalSpecWrapper(env, clip=True)
+    env = wrappers.SinglePrecisionWrapper(env)
+    return env
+
+
+def get_default_logger_factory(workdir: str, save_data=True, time_delta: float = 1.0):
+    """Create a custom logger factory for use in the experiment."""
+
+    def logger_factory(label: str, steps_key: Optional[str] = None, task_id: int = 0):
+        del steps_key, task_id
+
+        print_fn = logging.info
+        terminal_logger = acme_loggers.TerminalLogger(label=label, print_fn=print_fn)
+
+        loggers = [terminal_logger]
+
+        if save_data:
+            loggers.append(acme_loggers.CSVLogger(workdir, label=label))
+
+        # Dispatch to all writers and filter Nones and by time.
+        logger = acme_loggers.Dispatcher(loggers, acme_loggers.to_numpy)
+        logger = acme_loggers.NoneFilter(logger)
+        logger = acme_loggers.TimeFilter(logger, time_delta)
+
+        return logger
+
+    return logger_factory