Checklist for Maturity

• Fully deterministic
  • As far as possible
• Explicit Versioning
• Farama Notifications
• PyPI
• Full Package Description
• Deploy via GH Actions
• Website
• Linux and MacOS
• Precommit
• Typehinting
• Docstrings
• CI Testing
• Logo
• CoC
• Gymnasium/PZ API
• Python Versions

  • 3.7

• JKTerry and Mark as owners
• TOTP
• Google Analytics
• License
• Donations

Proposal
I was wondering if there's interest in making humanoid part of D4RL. The dataset was introduced in the following paper

Paper: https://arxiv.org/abs/2305.14550

The link for the data is here: https://dl.fbaipublicfiles.com/prajj/rl_paradigm/humanoid_offline_rl_data.tar.gz
Repo : https://github.com/prajjwal1/rl_paradigm

Motivation
Making humanoid available would be helpful for offline RL community for the same reason existing datasets are. Humanoid is more challenging in some ways than existing D4RL datasets such as state space dimension.

We provide medium, medium-expert and expert data for humanoid all in the same format as D4RL. It contains a lot of timestep data than what existing D4RL datasets provide.

Question

I was wondering, is there a separate Bibtex citation that we should use specifically for Minari, or should we use the one given for the D4RL project?

Proposal

I think it would be a great idea to port Antmaze data originally proposed in D4RL, to Minari.

Motivation

Many offline RL papers use antmaze. Many papers will even skip pointmaze and only report numbers on antmaze because the latter is more difficult. I think adding this dataset and environment with Minari's clean and easier to use code will boost Minari's usage amongst offline RL researchers.

Pitch

1. Use the data collecting policy from D4RL to re-collect data in Minari. I don't think the policy can be used directly, because the ant model for the original D4RL (mujoco-py antmaze) was different. I found that you could just divide the control output of the D4RL data collecting policy by some constant factor (like 5 / 10), it seems to work on gymansium antmaze.
2. If point (1)' fix is not appropriate, we would need to train a new data collecting policy for the antmaze and use it to collect new data.

Checklist

• [ YES ] I have checked that there is no similar issue in the repo (required)

pip install git+... (minari install from github) succeeds

1. pip install cython numpy
2. pip install git+https://github.com/Farama-Foundation/Minari.git

Everything works ok

pip install minari fails

Even when the prereqs numpy and cython are installed, pip install minari results in:

Building wheels for collected packages: minari
  Building wheel for minari (setup.py) ... error
  error: subprocess-exited-with-error
  
  × python setup.py bdist_wheel did not run successfully.
  │ exit code: 1
  ╰─> [28 lines of output]
      running bdist_wheel
      running build
      running build_py
      creating build
      creating build/lib.linux-x86_64-cpython-310
      creating build/lib.linux-x86_64-cpython-310/minari
      copying minari/__init__.py -> build/lib.linux-x86_64-cpython-310/minari
      copying minari/logger.py -> build/lib.linux-x86_64-cpython-310/minari
      copying minari/dataset.pyx -> build/lib.linux-x86_64-cpython-310/minari
      copying minari/dataset.pxd -> build/lib.linux-x86_64-cpython-310/minari
      copying minari/dataset.pyi -> build/lib.linux-x86_64-cpython-310/minari
      running build_ext
      building 'minari.dataset' extension
      creating build/temp.linux-x86_64-cpython-310
      creating build/temp.linux-x86_64-cpython-310/minari
      x86_64-linux-gnu-gcc -Wno-unused-result -Wsign-compare -DNDEBUG -g -fwrapv -O2 -Wall -g -fstack-protector-strong -Wformat -Werror=format-security -g -fwrapv -O2 -std=c++11 -fPIC -I/home/will/testminari/venv/lib/python3.10/site-packages/numpy/core/include -Iminari/cpp/include -I/home/will/testminari/venv/include -I/usr/include/python3.10 -c minari/dataset.cpp -o build/temp.linux-x86_64-cpython-310/minari/dataset.o -std=c++11 -O3 -ffast-math
      In file included from /home/will/testminari/venv/lib/python3.10/site-packages/numpy/core/include/numpy/ndarraytypes.h:1940,
                       from /home/will/testminari/venv/lib/python3.10/site-packages/numpy/core/include/numpy/ndarrayobject.h:12,
                       from /home/will/testminari/venv/lib/python3.10/site-packages/numpy/core/include/numpy/arrayobject.h:5,
                       from minari/dataset.cpp:792:
      /home/will/testminari/venv/lib/python3.10/site-packages/numpy/core/include/numpy/npy_1_7_deprecated_api.h:17:2: warning: #warning "Using deprecated NumPy API, disable it with " "#define NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION" [-Wcpp]
         17 | #warning "Using deprecated NumPy API, disable it with " \
            |  ^~~~~~~
      minari/dataset.cpp:806:10: fatal error: minari/dataset.h: No such file or directory
        806 | #include "minari/dataset.h"
            |          ^~~~~~~~~~~~~~~~~~
      compilation terminated.
      error: command '/usr/bin/x86_64-linux-gnu-gcc' failed with exit code 1
      [end of output]
  
  note: This error originates from a subprocess, and is likely not a problem with pip.
  ERROR: Failed building wheel for minari
  Running setup.py clean for minari
Failed to build minari
Installing collected packages: pyasn1, gymnasium-notices, urllib3, typing_extensions, structlog, six, rsa, pyasn1-modules, protobuf, jax-jumpy, idna, h5py, google-crc32c, cloudpickle, charset-normalizer, certifi, cachetools, tensorboardX, requests, googleapis-common-protos, google-resumable-media, google-auth, google-api-core, google-cloud-core, google-cloud-storage, shimmy, gymnasium, minari
  Running setup.py install for minari ... error
  error: subprocess-exited-with-error
  
  × Running setup.py install for minari did not run successfully.
  │ exit code: 1
  ╰─> [30 lines of output]
      running install
      /home/will/testminari/venv/lib/python3.10/site-packages/setuptools/command/install.py:34: SetuptoolsDeprecationWarning: setup.py install is deprecated. Use build and pip and other standards-based tools.
        warnings.warn(
      running build
      running build_py
      creating build
      creating build/lib.linux-x86_64-cpython-310
      creating build/lib.linux-x86_64-cpython-310/minari
      copying minari/__init__.py -> build/lib.linux-x86_64-cpython-310/minari
      copying minari/logger.py -> build/lib.linux-x86_64-cpython-310/minari
      copying minari/dataset.pyx -> build/lib.linux-x86_64-cpython-310/minari
      copying minari/dataset.pxd -> build/lib.linux-x86_64-cpython-310/minari
      copying minari/dataset.pyi -> build/lib.linux-x86_64-cpython-310/minari
      running build_ext
      building 'minari.dataset' extension
      creating build/temp.linux-x86_64-cpython-310
      creating build/temp.linux-x86_64-cpython-310/minari
      x86_64-linux-gnu-gcc -Wno-unused-result -Wsign-compare -DNDEBUG -g -fwrapv -O2 -Wall -g -fstack-protector-strong -Wformat -Werror=format-security -g -fwrapv -O2 -std=c++11 -fPIC -I/home/will/testminari/venv/lib/python3.10/site-packages/numpy/core/include -Iminari/cpp/include -I/home/will/testminari/venv/include -I/usr/include/python3.10 -c minari/dataset.cpp -o build/temp.linux-x86_64-cpython-310/minari/dataset.o -std=c++11 -O3 -ffast-math
      In file included from /home/will/testminari/venv/lib/python3.10/site-packages/numpy/core/include/numpy/ndarraytypes.h:1940,
                       from /home/will/testminari/venv/lib/python3.10/site-packages/numpy/core/include/numpy/ndarrayobject.h:12,
                       from /home/will/testminari/venv/lib/python3.10/site-packages/numpy/core/include/numpy/arrayobject.h:5,
                       from minari/dataset.cpp:792:
      /home/will/testminari/venv/lib/python3.10/site-packages/numpy/core/include/numpy/npy_1_7_deprecated_api.h:17:2: warning: #warning "Using deprecated NumPy API, disable it with " "#define NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION" [-Wcpp]
         17 | #warning "Using deprecated NumPy API, disable it with " \
            |  ^~~~~~~
      minari/dataset.cpp:806:10: fatal error: minari/dataset.h: No such file or directory
        806 | #include "minari/dataset.h"
            |          ^~~~~~~~~~~~~~~~~~
      compilation terminated.
      error: command '/usr/bin/x86_64-linux-gnu-gcc' failed with exit code 1
      [end of output]
  
  note: This error originates from a subprocess, and is likely not a problem with pip.
error: legacy-install-failure

× Encountered error while trying to install package.
╰─> minari

note: This is an issue with the package mentioned above, not pip.

Hi,

would it be better to let "ref_max_score" and "ref_min_score" be attribute of environment but not dataset? So that different datasets have a uniform normalization metric?

in d4RL implementation it also actually do so, for different dataset's corresponding environemnt, like "env-xxx-xxx", their ref_max_score and ref_min_score comes from same macro definition. https://github.com/Farama-Foundation/D4RL/blob/71a9549f2091accff93eeff68f1f3ab2c0e0a288/d4rl/gym_mujoco/__init__.py#L23-L31

So why not we directly make it as attribute of environemnt or dictionary of minari package?

Tests for the file minari_dataset should be added.

Proposal

Always remove the data generated by tests, whenever they fail or not.

Motivation

Currently, tests leave leftovers. This can make all the tests fail after one failed, because of minari.list_local_datasets()

Pitch

We can use pytest.fixture and a temporary directory to be sure everything is cleaned after.
See here for an example.

Checklist

• I have checked that there is no similar issue in the repo (required)

Describe the bug
The pointmaze-open-v1 dataset is not found in the list of remote datasets and fails to download.

Code example

minari.download_dataset('pointmaze-open-v1')
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File ".../lib/python3.8/site-packages/minari/storage/hosting.py", line 117, in download_dataset
    raise ValueError(
ValueError: Couldn't find any compatible version of dataset pointmaze-open with the local installed version of Minari, 0.4.1.

System Info
Describe the characteristic of your environment:

• Describe how Gymnasium was installed: Pip
• Operating system: MacOS
• Python version: 3.8.12

Checklist

• I have checked that there is no similar issue in the repo (required)

Hey, I am trying to make dataset of Ant-v5 and I am getting this error:

$ py create_dataset.py  # expert-v0
Traceback (most recent call last):
  File "/home/master-andreas/gym/rl/project/create_dataset.py", line 56, in <module>
    obs, rew, terminated, truncated, info = collector_env.step(action)
                                            ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/master-andreas/gym/rl/project/temp_env/lib/python3.11/site-packages/minari/data_collector/data_collector.py", line 222, in step
    self._add_step_data(self._buffer[-1], step_data)
  File "/home/master-andreas/gym/rl/project/temp_env/lib/python3.11/site-packages/minari/data_collector/data_collector.py", line 162, in _add_step_data
    raise ValueError(
ValueError: Info structure inconsistent with info structure returned by original reset.

for reference:
self._reference_info is set using info from reset

which is then compared to the info from step

in general, step can have additional info compared to reset such as transition information

Hi, isn't it too hard to have an assertion error when observations or action are not contained in its defined spaces?
Perhaps an option to disable or set error-level to ignore/warn/error?

Question

While hdf5 and h5py is the most popular approach for multi-dimensional arrays storage, is has some major limitations. For example, the inability to read data in multiple processes / threads simultaneously, which can be important for the implementation of efficient data loading.

There is an alternative - Zarr, which is very similar, but a bit more capable. I think a discussion on this would be useful to the community.

Hey, I am trying to make dataset of Ant-v5 and i am getting this error:

$ py create_dataset.py  # expert-v0
Traceback (most recent call last):
  File "/home/master-andreas/gym/rl/project/create_dataset.py", line 56, in <module>
    obs, rew, terminated, truncated, info = collector_env.step(action)
                                            ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/master-andreas/gym/rl/project/temp_env/lib/python3.11/site-packages/minari/data_collector/data_collector.py", line 222, in step
    self._add_step_data(self._buffer[-1], step_data)
  File "/home/master-andreas/gym/rl/project/temp_env/lib/python3.11/site-packages/minari/data_collector/data_collector.py", line 162, in _add_step_data
    raise ValueError(
ValueError: Info structure inconsistent with info structure returned by original reset.

for reference:
self._reference_info is set using info from reset

which is then compared to the info from reset

Also in general, it is not in Env's specification that info s need to have the same structure each time
https://gymnasium.farama.org/main/api/env/#gymnasium.Env.step

@alexdavey @younik

I am trying load a dataset created from a Minigrid-type environment using the following code:

import os

import gymnasium as gym
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from gymnasium import spaces
from stable_baselines3 import PPO
from torch.utils.data import DataLoader
from tqdm.auto import tqdm

import minari
from minari import DataCollectorV0

def collate_fn(batch):
    return {
        "id": torch.Tensor([x.id for x in batch]),
        "seed": torch.Tensor([x.seed for x in batch]),
        "total_timesteps": torch.Tensor([x.total_timesteps for x in batch]),
        "observations": torch.nn.utils.rnn.pad_sequence(
            [torch.as_tensor(x.observations) for x in batch],
            batch_first=True
        ),
        "actions": torch.nn.utils.rnn.pad_sequence(
            [torch.as_tensor(x.actions) for x in batch],
            batch_first=True
        ),
        "rewards": torch.nn.utils.rnn.pad_sequence(
            [torch.as_tensor(x.rewards) for x in batch],
            batch_first=True
        ),
        "terminations": torch.nn.utils.rnn.pad_sequence(
            [torch.as_tensor(x.terminations) for x in batch],
            batch_first=True
        ),
        "truncations": torch.nn.utils.rnn.pad_sequence(
            [torch.as_tensor(x.truncations) for x in batch],
            batch_first=True
        )
    }


torch.manual_seed(42)

minari_testset = minari.load_dataset("MinigridRandomWall-6Spots-v0")
dataloader = DataLoader(minari_dataset, batch_size=64, shuffle=True, collate_fn=collate_fn)

for batch in dataloader:
    print("Observation shape: " + str(batch['observations'].shape))
    print("Action shape: " + str(batch['actions'].shape))
    print("Reward shape: " + str(batch['rewards'].shape))
    print("Timestep shape " + str(batch["infos"]["timestep"].shape))

When I run the code, I get this error:

    minari_testset = minari.load_dataset("MinigridRandomWall-6Spots-v0")
  File "/home/justin/Minari/minari/storage/local.py", line 22, in load_dataset
    return MinariDataset(data_path)
  File "/home/justin/Minari/minari/dataset/minari_dataset.py", line 133, in __init__
    self._data = MinariStorage(data)
  File "/home/justin/Minari/minari/dataset/minari_storage.py", line 22, in __init__
    flatten_observations = f.attrs["flatten_observation"].item()
  File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper
  File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper
  File "/usr/local/lib/python3.8/dist-packages/h5py/_hl/attrs.py", line 56, in __getitem__
    attr = h5a.open(self._id, self._e(name))
  File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper
  File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper
  File "h5py/h5a.pyx", line 80, in h5py.h5a.open
KeyError: "Can't open attribute (can't locate attribute in name index)"

When I created the dataset, I used an ImgObsWrapper for the environment. Could that be the source of the problem?

Describe the bug
According to the documentation, the shape of the observation should be num_steps + 1. However, for a handful of episodes, the length is num_steps.

Code example

import minari

dataset = minari.load_dataset("antmaze-umaze-v0", download=True)

for episode_data in dataset.iterate_episodes():
    if episode_data.observations["observation"].shape[0] == episode_data.total_timesteps:
        print("Warning: bad episode")

Output:

Warning: bad episode
Warning: bad episode
Warning: bad episode
Warning: bad episode

System Info
Describe the characteristic of your environment:

• Installed via PIP
• Operating system: Ubuntu 20.04.6 LTS
• Python version: 3.8

Additional context
I haven't checked the other Ant environments.

Checklist

• I have checked that there is no similar issue in the repo (required)

Question

Is this compatible with PettingZoo? I've adapted a Gym environment into a custom PettingZoo environment, and I wanted to use Minari for some Offline RL. However, it seems DataCollector needs an environment instantiated by "gym.make" and whenever I try to make my environment "registrable" I get errors because it seems that the interface required is not really compatible with that of PettingZoo.

Is this a problem of PettingZoo or Minari? Or mine?
Thanks

Question

Hi, please state clearly in the documentation and dataset definition if in a time step "r_0" is consequence of "a_0"

With previous Offline RL libs, there has been some confusion with this respect.

With the standar in RL being (s,a,r,s') one assume that r is a consequence of applying action a in state s.

If r is not, please state it clearly, because then, the r(s,a) should be r_1 and not r_0

Thanks !

Parallel episode sampling

I have a use case where we have a dataset consisting of image-based observations, and I notice that sampling speed seems to be slower than with 1D observations. I checked out how sampling is working internally, and noticed that Minari samples episodes serially, instead of sampling in parallel. I thought that parallelizing this call may have been thought about already, so I was curious for any recommendations on the best way to do this. I was also wondering if this was something that will be added in the future.

I have one more layer of complexity on top of this, where instead of 1 dataset, I have say 10 datasets from different envs, each have image-based observations. Think multi-task Atari. I have 10 minari datasets, and then say want 30 episodes from each for each gradient update. Also want to do this in parallel, and will experiment with different parallelization techniques but curious if others had intuition about this.

Proposal

If download True fetch dataset name from remote Farama server.

A clear and concise description of the proposal.

Motivation

This functionality is similar to Pytorch dataset loading

Checklist

• I have checked that there is no similar issue in the repo (required)

I tried to set the render mode in recover_environment as is allowed in gym.make:

data_complete = minari.load_dataset("kitchen-complete-v1")
env = data_complete.recover_environment(render_mode="rgb_array")

But I got the following error:

TypeError: MinariDataset.recover_environment() got an unexpected keyword argument 'render_mode'

Question

Hello, the majority of data collections from simulated environments comes from parallel data collectors such as VecEnvs and methods that return either a step data with a batch dimension or even at the end of the collector a list like collection of episode data.

By now the simpler scenario is to sample using VecEnv so is there any simple recipe for adapting the current DataCollector for VecEnv format?

Describe the bug

I have a custom gym env that inherits MujocoRobotEnv and wrap it with DataCollector. However, I found all image observations that are stored in DataCollector are black. When not using DataCollector, a custom environment wrapped by PixelObservationV0 can successfully obtain a colored image from the step function. Besides, the ValueError is also triggered when the info dict returned by the step function is not empty. Below is a minimal working example to reproduce these errors. The panda_mujoco_gym env was created by myself and can be required from here.

Thanks in advance for your precious time in checking this issue!

Code example

import gymnasium as gym
from gymnasium.experimental.wrappers import PixelObservationV0
import matplotlib.pyplot as plt
from minari import DataCollector
import panda_mujoco_gym

env = gym.make("FrankaPickAndPlaceSparse-v0", render_mode="rgb_array")
env = PixelObservationV0(env, pixels_only=True)
obs, info = env.reset(seed=seed) # rgb image can be observed

env = DataCollector(env, record_infos=True, max_buffer_steps=max_buffer_steps)
obs, info = env.reset(seed=seed)
plt.imshow(obs) # rgb image can not be observed
plt.show()

System Info
Describe the characteristic of your environment:

• Describe how Minari was installed (pip, docker, source, ...): pip
• Operating system: Ubuntu 20.04
• Python version: 3.9.18
• Minari version: 0.4.3

Checklist

• I have checked that there is no similar issue in the repo (required)

Describe the bug
Setting the author email to None, then saving the dataset will lead to the following exception:

    dataset.save()
  File "minari/dataset.pyx", line 576, in minari.dataset.MinariDataset.save
  File "minari/dataset.pyx", line 582, in minari.dataset.MinariDataset.save
  File "/Users/markus/OpenSource/venvs/gymnasium_env/lib/python3.10/site-packages/h5py/_hl/group.py", line 161, in create_dataset
    dsid = dataset.make_new_dset(group, shape, dtype, data, name, **kwds)
  File "/Users/markus/OpenSource/venvs/gymnasium_env/lib/python3.10/site-packages/h5py/_hl/dataset.py", line 54, in make_new_dset
    raise TypeError("One of data, shape or dtype must be specified")
TypeError: One of data, shape or dtype must be specified

I would expect None to be a valid choice.

Code example
Just set the author email to None in tutorials/LocalStorage/local_storage.py

System Info
Describe the characteristic of your environment:

• Gymnasium installed from source
• Operating system: MacOS
• Python version: Python 3.10

Checklist

• I have checked that there is no similar issue in the repo (required)

The line:

This currently causes an import error in gymnasium 1.0.0rc

Should be changed to:

from gymnasium.wrappers import RecordEpisodeStatistics

Describe the bug
Have you tried combining more than two dataset at a time? When I try to do that, I get an error, while combining only two datasets works fine. I think this might be related to the h5py/h5py#1385.

Error

Traceback (most recent call last):
  File "rllib_mini/train_ppo.py", line 171, in <module>
    train()
  File "/usr/local/lib/python3.8/dist-packages/pyrallis/argparsing.py", line 158, in wrapper_inner
    response = fn(cfg, *args, **kwargs)
  File "rllib_mini/train_ppo.py", line 164, in train
    comdined_dataset = minari.combine_datasets(datasets, new_dataset_id=f"{config.env_name}-dataset-v{config.version}")
  File "/usr/local/lib/python3.8/dist-packages/minari/utils.py", line 132, in combine_datasets
    combined_data_file.attrs["author"] = dataset_file.attrs["author"]
  File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper
  File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper
  File "/usr/local/lib/python3.8/dist-packages/h5py/_hl/attrs.py", line 103, in __setitem__
    self.create(name, data=value)
  File "/usr/local/lib/python3.8/dist-packages/h5py/_hl/attrs.py", line 212, in create
    h5a.delete(self._id, self._e(name))
  File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper
  File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper
  File "h5py/h5a.pyx", line 145, in h5py.h5a.delete
KeyError: 'Unable to delete attribute (record is not in B-tree)'

Checklist

• I have checked that there is no similar issue in the repo (required)

Ths discord link does not work

Describe the bug
When creating a BabyAI dataset, I get the following NotImplementedError No serialization method available for MissionSpace(<function BabyAIMissionSpace._gen_mission at 0x7f7a93fcdb90>, None)

Code example

import gymnasium as gym
import minari
from minari import DataCollectorV0

dataset_id = 'BabyAI-GoToLocal-v0'
env = gym.make(dataset_id)
env = DataCollectorV0(env, record_infos=True, max_buffer_steps=100000)

total_episodes = 100

for _ in range(total_episodes):
    env.reset(seed=123)
    while True:
        # random action policy
        action = env.action_space.sample()
        obs, rew, terminated, truncated, info = env.step(action)
        
        if terminated or truncated:
            break

dataset = minari.create_dataset_from_collector_env(dataset_id=dataset_id, 
                                               collector_env=env,
                                               algorithm_name="Random-Policy"
                                               )

System Info
Describe the characteristic of your environment:

• Describe how Gymnasium was installed: pip
• Operating system: MacOs
• Python version: 3.7.16

Additional context
BabyAI is part of MiniGrid: BabyAI GoToLocal

Checklist

• I have checked that there is no similar issue in the repo (required)

Describe the bug
Total steps for "pen-human-v0" dataset is none or doesn't exist. Seems like this is true for all adroit datasets.

Code example

import minari

minari.download_dataset("pen-human-v0")
dataset = minari.load_dataset("pen-human-v0")
print(dataset.total_steps)

Additional context
However, this will work

print(dataset.spec.total_steps)

Why doesn't the code just take it out of spec? Why does it either None or re-read from self._data?

Checklist

• I have checked that there is no similar issue in the repo (required)

Proposal

Add docs to the webpage for each dataset in the remote GCP bucket (https://minari.farama.org/main/datasets/coming_soon/). Generate the docs automatically wen building the website, this can be done similarly to the Gymnasium environment docs.

The side menu in the docs webpage should have the names of the datasets classified by environment id:

For the docs of each dataset include the following:

Dataset Specs

These specs can be retrieved from the minari.list_remote_datasets() function

Checklist

• I have checked that there is no similar issue in the repo (required)

>>> python pointmaze_dataset.py
/home/graham/micromamba/envs/ctx/lib/python3.10/site-packages/gymnasium/core.py:297: UserWarning: WARN: env.maze to get variables from other wrappers is deprecated and will be removed in v1.0, to get this variable you can do `env.unwrapped.maze` for environment variables or `env.get_attr('maze')` that will search the reminding wrappers.
  logger.warn(
Traceback (most recent call last):
  File "/home/graham/code/inctxdt/compare/pointmaze/pointmaze_dataset.py", line 369, in <module>
    obs, rew, terminated, truncated, info = collector_env.step(action)
  File "/home/graham/micromamba/envs/ctx/lib/python3.10/site-packages/minari/data_collector/data_collector.py", line 216, in step
    assert self.dataset_action_space.contains(
AssertionError: Actions are not in action space.

also seems like there is an issue with the spaces .contain method whereby the assert in data_collector.py can fail if the dtype is float64 and the space dtype is float32

Describe the bug

Hi, glad to know that the info in dataset is also iteratable now. However, I just find that the existing dataset just lost the info when every time the environment is reset.

Acoording to the current data collection standard, https://minari.farama.org/content/dataset_standards/#additional-information-formatting, the collector will collect info while env.reset() and every time env.step(), i.e. it should satisfy: len(observations) == len(info[xxx]), however, I tested some datasets, it only satisifes len(actions) == len(info[xxx]), I guess it may related to d4rl, while d4rl released, env.reset() doesn't return any infos.

We may need to fix this issue.

Code example
I tested two datasets, but other should be same:

for name in ['door-human-v1', 'hammer-expert-v1']:
  dt = minari.load_dataset(name)
  obs =  dt[1].observations
  act = dt[1].observations
  info = dt[1].infos
  info_key = list(info.keys()][0]
  info_i = info[info_key]
  print(len(info_i) == len(act))
  print(len(info_i) == len(obs))

System Info
I thinks this is not important to report system info for this issue

Checklist

• I have checked that there is no similar issue in the repo (required)

Question

I am trying to build a dataset for the MiniHack environment. The environment has a fairly large dictionary space, with many keys, but only one level of nesting. I don't want to save all the keys, because it bloats the size of the dataset very much. I thought it could be done with StepDataCallback, like this:

class TTYStepDataCallback(minari.StepDataCallback):
    def __call__(self, env, obs, info, action=None, rew=None, terminated=None, truncated=None):
        tty_filter_keys = ["tty_chars", "tty_colors", "tty_cursor"]
        obs = {k: v for k, v in obs.items() if k in tty_filter_keys}

        step_data = super().__call__(env, obs, info, action, rew, terminated, truncated)
        
        return step_data

However, this will raise an error, as some keys from original observation space are missing. Why can't I just use the ones I want to save right away? I am logging data during distributed PPO training, which uses the other keys too, so they are needed for online training :(

Question

Hi,

it looks strange that the installing requires 4.4.0, since it could cause conflict frequently.

I checked the usage of typing-extensions, it only used for two places

• Use TypedDict as StepData, here

  Minari/minari/data_collector/callbacks/step_data.py

  Line 2 in ea40978

  from typing_extensions import TypedDict

  TypedDict is officially support since python 3.8 and minari requires also at least python3.8.

• Use Annotated here,

  Minari/minari/data_collector/callbacks/step_data.py

  Line 2 in ea40978

  from typing_extensions import TypedDict

  
  Looks it's mainly designed for cooperating with typer, but typer requires typing-extensions>3.7 . Here (https://github.com/tiangolo/typer/blob/e0b207f3f577cb2e59fdd60da39686a2f5ed0e77/pyproject.toml#L34 )

So, shall the requirement of typing-extensions==4.4.0 be removed?

Question

I have just created a dataset from the deployment of an agent on the Farama Minigrid environment, and I am wondering how to ensure that when data is loaded from the dataset, it is done in the original order, without any steps being loaded redundantly. Is there a way to do that with the pytorch DataLoader class? For instance, given the example shown here https://minari.farama.org/content/basic_usage/, how would I go about getting batches from the DataLoader such that the steps given in the batches were in the same order as that in which they were created when the agent was deployed on the environment? Also, is there a way that I can do that without any of the steps being repeated?

Missing info property in EpisodeData

The DataCollectorV0 wrapper can optionally record the info from each step. However, when creating a dataset from the collector the info data is lost.
Shouldn't EpisodeData also optionally keep the infos? Else, what is the purpose of recording the info on the wrapper?

Proposal

Add gymasium env id, file size and dataset group to the displayed table when running command minari list remote or minari list local.

So this would mean adding new columns named something like "env_id", "size on disk" and something like "dataset group."

Right now, the datasets do not have a dataset_group value, so for backwards compatibility, the PR should check for a dataset_group attribute, and if there is none, it should use the string "Unknown" as a placeholder value.

This should be a useful hint for getting started with getting the file size for remote datasets: https://stackoverflow.com/questions/50875461/google-cloud-storage-get-object-size-api

To get started with this, it would be useful to look at the code in cli.py local.py and hosting.py

The doc will also need to be updated to reflect the existence of the new field dataset_group. Definitely on this page, https://minari.farama.org/main/content/dataset_standards/, and probably also on the individual dataset pages.

Motivation

This is partially to address #79, and also it's useful to know how large each dataset is to get an idea of how long it will take to download or process a particular dataset.

Checklist

• I have checked that there is no similar issue in the repo (required)

Question

Does not iterating over episodes should produce "all" the episode data including infos and other stores keys?

After that we can use more specific iterators to just select a subset of the keys.

I.e. for what reason we are able to store in MuJoCo Hopper the x-position if when iterating we can't read it? And we cannot access Infos per episode?

Question

Hi,

I am curious is there any time plan to migrates d4rl mujoco environments to minari? E.g. ant-medium/expert/...-v2/v4, hopper-medium/expert/...-v2/v4/?

Though mujoco's -v2 environments are no longer maintained, but it's still popular baseline for research.

Thanks your great work!

Describe the bug
Looks if we wanna to make pre-commit check, it always throws error about typing issue that relates to gymnasium. If the gymnasium is installed, when we do pre-commit, pyright check will give error relates to typing hint of env_spec or like those. If we uninstall gymnasium then the error will dispear.

For more detail, you could check the code example and attached output. I think it's very clear that the error is caused by type hint of Optional, the env attribute could be None. E.g. environment's env_spec could be None, then env_spec.make could be None.make() so pyright thorws error.

Also, I believe in github workflows, if we add step pip install gymnasium(or install dependencies of minari) into precommit action, pre-commit will thorow error anyway.

Code example
To recover, just do so:

$ conda create -n test3.11 python=3.11
$ conda activate test3.11
$ git clone [email protected]:Farama-Foundation/Minari.git
$ cd Minari
$ pip install -e . # or run pip install gymnasium
$ pip install pre-commit
$ pre-commit run --all-files

Then you could find there is four errors and quite some warnings If we run $ pip uninstall gymnasium then the error and warning will disapear.

xxxxx/Minari/minari/utils.py
xxxxx/Minari/minari/utils.py:177:13 - error: Argument of type "Generator[int | None, None, None]" cannot be assigned to parameter "__iterable" of type "Iterable[SupportsRichComparisonT@max]" in function "max"
  "Generator[int | None, None, None]" is incompatible with "Iterable[SupportsRichComparisonT@max]"
    TypeVar "_T_co@Iterable" is covariant
      Type "int | None" cannot be assigned to type "SupportsRichComparison"
        Type "int | None" cannot be assigned to type "SupportsRichComparison"
          Type "None" cannot be assigned to type "SupportsRichComparison" (reportGeneralTypeIssues)
xxxxx/Minari/minari/utils.py:344:60 - error: Argument of type "ActType@get_average_reference_score" cannot be assigned to parameter "action" of type "WrapperActType@Wrapper" in function "step"
  Type "ActType@get_average_reference_score" cannot be assigned to type "WrapperActType@Wrapper" (reportGeneralTypeIssues)
xxxxx/Minari/minari/dataset/minari_dataset.py
xxxxx/Minari/minari/dataset/minari_dataset.py:279:41 - warning: "_total_episodes" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/minari/dataset/minari_dataset.py:314:41 - warning: "_total_episodes" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/common.py
xxxxx/Minari/tests/common.py:566:32 - warning: "_buffer" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/common.py:568:28 - warning: "_buffer" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/data_collector/callbacks/test_step_data_callback.py
xxxxx/Minari/tests/data_collector/callbacks/test_step_data_callback.py:104:34 - warning: "_data" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/dataset/test_dataset_download.py
xxxxx/Minari/tests/dataset/test_dataset_download.py:59:34 - warning: "_data" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/dataset/test_minari_dataset.py
xxxxx/Minari/tests/dataset/test_minari_dataset.py:94:32 - warning: "_buffer" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/dataset/test_minari_dataset.py:96:28 - warning: "_buffer" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/dataset/test_minari_dataset.py:107:34 - warning: "_data" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/dataset/test_minari_dataset.py:154:26 - warning: "_data" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/dataset/test_minari_dataset.py:168:32 - warning: "_buffer" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/dataset/test_minari_dataset.py:170:28 - warning: "_buffer" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/dataset/test_minari_dataset.py:199:19 - warning: "_episode_indices" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/dataset/test_minari_dataset.py:200:29 - warning: "_data" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/dataset/test_minari_dataset.py:201:20 - warning: "_data" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/dataset/test_minari_dataset.py:288:19 - warning: "_episode_indices" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/dataset/test_minari_dataset.py:289:29 - warning: "_data" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/dataset/test_minari_dataset.py:290:20 - warning: "_data" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/dataset/test_minari_dataset.py:466:34 - warning: "_data" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/utils/test_dataset_combine.py
xxxxx/Minari/tests/utils/test_dataset_combine.py:74:18 - error: "max_episode_steps" is not a known member of "None" (reportOptionalMemberAccess)
xxxxx/Minari/tests/utils/test_dataset_combine.py:75:24 - error: "make" is not a known member of "None" (reportOptionalMemberAccess)
xxxxx/Minari/tests/utils/test_dataset_creation.py
xxxxx/Minari/tests/utils/test_dataset_creation.py:59:32 - warning: "_buffer" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/utils/test_dataset_creation.py:61:28 - warning: "_buffer" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/utils/test_dataset_creation.py:80:34 - warning: "_data" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/utils/test_dataset_creation.py:173:34 - warning: "_data" is protected and used outside of the class in which it is declared (reportPrivateUsage)
xxxxx/Minari/tests/utils/test_dataset_creation.py:287:34 - warning: "_data" is protected and used outside of the class in which it is declared (reportPrivateUsage)
4 errors, 24 warnings, 0 informations
Completed in 1.5sec

System Info
Describe the characteristic of your environment:

• Describe how Minari was installed (pip, docker, source, ...): See code example
• Operating system: MacOS
• Python version: Python 3.11

Checklist

• I have checked that there is no similar issue in the repo (required)

Writing this without fixes for now, things I'm noticing going through the library:

1. The dataset listed in README doesn't seem to exist. This is bad, as that's the first thing a beginner will likely try when trying to interact with the library. We should either swap it, or create it.
2. Remote datasets are 'door-cloned-v0', 'door-expert-v0', 'door-human-v0', ..., at least reading it initially. One issue here is that the dataset names don't really tell you much unless you already know the context - we should probably mention that they're specifically for mujoco, probably specifically for some specific model - maybe we can list the name of the gymnasium env?
3. Trying to do minari.download_dataset("door-human-v0") results in a pretty ugly/uncaught import error for gymnasium_robotics. Since gymnasium-robotics isn't a necessary dependency, we should have a more elegant way of handling this.
4. (3a) The import issue seems to happen because when downloading the dataset, we're creating the environment to get the obs and action spaces. Ignoring even the dependency issue, this seems very redundant - can't we just save the obs/action spaces in the dataset? They should be easy to serialize.
5. Downloading the same dataset twice repeats the download and overwrites it. We explicitly acknowledge this, but imo there should be an overwrite or force argument to do that behavior, and by default it should just skip downloading it. It might be worthwhile to think if we want to do explicit downloads like this in the first place, but that's another thing.
6. EpisodeData is a namedtuple, which leads to the default representation being an enormous printout of all observations, rewards etc. This is not very useful, so we should probably make it a proper dataclass instead, and override the default representation.

Stopping here for now, @rodrigodelazcano let me know if there's anything I'm missing anything for any of those points (like some additional reasons for doing X thing in a certain way that I'm unaware of). Otherwise we can make this into a to-do list of improvements

MacOS creates .DS_Store if you access to a folder using Finder.
If this is done on the datasets folder, this will produce an error:

NotADirectoryError: [Errno 20] Not a directory: '/Users/username/.minari/datasets/.DS_Store'

because it tries to read .DS_Store as a dataset

Question

Is there a software citation for Minari available for citing it in a paper?

Question

Hi,

the discord link is invalid now. Would you link to update discord invitation link?

best