[](https://github.com/pytorch/rl/actions/workflows/test-linux.yml)
[](https://pytorch.org/rl/)
[](https://pytorch.github.io/rl/dev/bench/)
[](https://codecov.io/gh/pytorch/rl)
[](https://twitter.com/torchrl1)
[](https://www.python.org/downloads/)
[](https://github.com/pytorch/rl/blob/main/LICENSE)
<a href="https://pypi.org/project/torchrl"><img src="https://img.shields.io/pypi/v/torchrl" alt="pypi version"></a>
<a href="https://pypi.org/project/torchrl-nightly"><img src="https://img.shields.io/pypi/v/torchrl-nightly?label=nightly" alt="pypi nightly version"></a>
[](https://pepy.tech/project/torchrl)
[)](https://pepy.tech/project/torchrl-nightly)
[](https://discord.gg/cZs26Qq3Dd)
# TorchRL
<p align="center">
<img src="docs/source/_static/img/icon.png" width="200" >
</p>
[**Documentation**](#documentation-and-knowledge-base) | [**TensorDict**](#writing-simplified-and-portable-rl-codebase-with-tensordict) |
[**Features**](#features) | [**Examples, tutorials and demos**](#examples-tutorials-and-demos) | [**Citation**](#citation) | [**Installation**](#installation) |
[**Asking a question**](#asking-a-question) | [**Contributing**](#contributing)
**TorchRL** is an open-source Reinforcement Learning (RL) library for PyTorch.
It provides pytorch and **python-first**, low and high level abstractions for RL that are intended to be **efficient**, **modular**, **documented** and properly **tested**.
The code is aimed at supporting research in RL. Most of it is written in python in a highly modular way, such that researchers can easily swap components, transform them or write new ones with little effort.
This repo attempts to align with the existing pytorch ecosystem libraries in that it has a dataset pillar ([torchrl/envs](torchrl/envs)), [transforms](torchrl/envs/transforms), [models](torchrl/modules), data utilities (e.g. collectors and containers), etc.
TorchRL aims at having as few dependencies as possible (python standard library, numpy and pytorch). Common environment libraries (e.g. OpenAI gym) are only optional.
On the low-level end, torchrl comes with a set of highly re-usable functionals for [cost functions](torchrl/objectives/costs), [returns](torchrl/objectives/returns) and data processing.
TorchRL aims at (1) a high modularity and (2) good runtime performance. Read the [full paper](https://arxiv.org/abs/2306.00577) for a more curated description of the library.
## Getting started
Check our [Getting Started tutorials](https://pytorch.org/rl/index.html#getting-started) for quickly ramp up with the basic
features of the library!
## Documentation and knowledge base
The TorchRL documentation can be found [here](https://pytorch.org/rl).
It contains tutorials and the API reference.
TorchRL also provides a RL knowledge base to help you debug your code, or simply
learn the basics of RL. Check it out [here](https://pytorch.org/rl/reference/knowledge_base.html).
We have some introductory videos for you to get to know the library better, check them out:
- [TorchRL intro at PyTorch day 2022](https://youtu.be/cIKMhZoykEE)
- [PyTorch 2.0 Q&A: TorchRL](https://www.youtube.com/live/myEfUoYrbts?feature=share)
## Writing simplified and portable RL codebase with `TensorDict`
RL algorithms are very heterogeneous, and it can be hard to recycle a codebase
across settings (e.g. from online to offline, from state-based to pixel-based
learning).
TorchRL solves this problem through [`TensorDict`](https://github.com/pytorch/tensordict/),
a convenient data structure<sup>(1)</sup> that can be used to streamline one's
RL codebase.
With this tool, one can write a *complete PPO training script in less than 100
lines of code*!
<details>
<summary>Code</summary>
```python
import torch
from tensordict.nn import TensorDictModule
from tensordict.nn.distributions import NormalParamExtractor
from torch import nn
from torchrl.collectors import SyncDataCollector
from torchrl.data.replay_buffers import TensorDictReplayBuffer, \
LazyTensorStorage, SamplerWithoutReplacement
from torchrl.envs.libs.gym import GymEnv
from torchrl.modules import ProbabilisticActor, ValueOperator, TanhNormal
from torchrl.objectives import ClipPPOLoss
from torchrl.objectives.value import GAE
env = GymEnv("Pendulum-v1")
model = TensorDictModule(
nn.Sequential(
nn.Linear(3, 128), nn.Tanh(),
nn.Linear(128, 128), nn.Tanh(),
nn.Linear(128, 128), nn.Tanh(),
nn.Linear(128, 2),
NormalParamExtractor()
),
in_keys=["observation"],
out_keys=["loc", "scale"]
)
critic = ValueOperator(
nn.Sequential(
nn.Linear(3, 128), nn.Tanh(),
nn.Linear(128, 128), nn.Tanh(),
nn.Linear(128, 128), nn.Tanh(),
nn.Linear(128, 1),
),
in_keys=["observation"],
)
actor = ProbabilisticActor(
model,
in_keys=["loc", "scale"],
distribution_class=TanhNormal,
distribution_kwargs={"min": -1.0, "max": 1.0},
return_log_prob=True
)
buffer = TensorDictReplayBuffer(
LazyTensorStorage(1000),
SamplerWithoutReplacement()
)
collector = SyncDataCollector(
env,
actor,
frames_per_batch=1000,
total_frames=1_000_000
)
loss_fn = ClipPPOLoss(actor, critic, gamma=0.99)
optim = torch.optim.Adam(loss_fn.parameters(), lr=2e-4)
adv_fn = GAE(value_network=critic, gamma=0.99, lmbda=0.95, average_gae=True)
for data in collector: # collect data
for epoch in range(10):
adv_fn(data) # compute advantage
buffer.extend(data.view(-1))
for i in range(20): # consume data
sample = buffer.sample(50) # mini-batch
loss_vals = loss_fn(sample)
loss_val = sum(
value for key, value in loss_vals.items() if
key.startswith("loss")
)
loss_val.backward()
optim.step()
optim.zero_grad()
print(f"avg reward: {data['next', 'reward'].mean().item(): 4.4f}")
```
</details>
Here is an example of how the [environment API](https://pytorch.org/rl/reference/envs.html)
relies on tensordict to carry data from one function to another during a rollout
execution:
`TensorDict` makes it easy to re-use pieces of code across environments, models and
algorithms.
<details>
<summary>Code</summary>
For instance, here's how to code a rollout in TorchRL:
```diff
- obs, done = env.reset()
+ tensordict = env.reset()
policy = SafeModule(
model,
in_keys=["observation_pixels", "observation_vector"],
out_keys=["action"],
)
out = []
for i in range(n_steps):
- action, log_prob = policy(obs)
- next_obs, reward, done, info = env.step(action)
- out.append((obs, next_obs, action, log_prob, reward, done))
- obs = next_obs
+ tensordict = policy(tensordict)
+ tensordict = env.step(tensordict)
+ out.append(tensordict)
+ tensordict = step_mdp(tensordict) # renames next_observation_* keys to observation_*
- obs, next_obs, action, log_prob, reward, done = [torch.stack(vals, 0) for vals in zip(*out)]
+ out = torch.stack(out, 0) # TensorDict supports multiple tensor operations
```
</details>
Using this, TorchRL abstracts away the input / output signatures of the modules, env,
collectors, replay buffers and losses of the library, allowing all primitives
to be easily recycled across settings.
<details>
<summary>Code</summary>
Here's another example of an off-policy training loop in TorchRL (assuming
that a data collector, a replay buffer, a loss and an optimizer have been instantiated):
```diff
- for i, (obs, next_obs, action, hidden_state, reward, done) in enumerate(collector):
+ for i, tensordict in enumerate(collector):
- replay_buffer.add((obs, next_obs, action, log_prob, reward, done))
+ replay_buffer.add(tensordict)
for j in range(num_optim_steps):
- obs, next_obs, action, hidden_state, reward, done = replay_buffer.sample(batch_size)
- loss = loss_fn(obs, next_obs, action, hidden_state, reward, done)
+ tensordict = replay_buffer.sample(batch_size)
+ loss = loss_fn(tensordict)
loss.backward()
optim.step()
optim.zero_grad()
```
This training loop can be re-used across algorithms as it makes a minimal number of assumptions about the structure of the data.
</details>
TensorDict supports multiple tensor operations on its device and shape
(the shape of TensorDict, or its batch size, is the common arbitrary N first dimensions of all its contained tensors):
<details>
<summary>Code</summary>
```python
# stack and cat
tensordict = torch.stack(list_of_tensordicts, 0)
tensordict = torch.cat(list_of_tensordicts, 0)
# reshape
tensordict = tensordict.view(-1)
tensordict = tensordict.permute(0, 2, 1)
tensordict = tensordict.unsqueeze(-1)
tensordict = tensordict.squeeze(-1)
# indexing
tensordict = tensordict[:2]
tensordict[:, 2] = sub_tensordict
# device and memory location
tensordict.cuda()
tensordict.to("cuda:1")
tensordict.share_memory_()
```
</details>
TensorDict comes with a dedicated [`tensordict.nn`](https://pytorch.github.io/tensordict/reference/nn.html)
module that contains everything you might need to write your model with it.
And it is `functorch` and `torch.compile` compatible!
<details>
<summary>Code</summary>
```diff
transformer_model = nn.Transformer(nhead=16, num_encoder_layers=12)
+ td_module = SafeModule(transformer_model, in_keys=["src", "tgt"], out_keys=["out"])
src = torch.rand((10, 32, 512))
tgt = torch.rand((20, 32, 512))
+ tensordict = TensorDict({"src": src, "tgt": tgt}, batch_size=[20, 32])
- out = transformer_model(src, tgt)
+ td_module(tensordict)
+ out = tensordict["out"]
```
The `TensorDictSequential` class allows to branch sequences of `nn.Module` instances in a highly modular way.
For instance, here is an implementation of a transformer using the encoder and decoder blocks:
```python
encoder_module = TransformerEncoder(...)
encoder = TensorDictSequential(encoder_module, in_keys=["src", "src_mask"], out_keys=["memory"])
decoder_module = TransformerDecoder(...)
decoder = TensorDictModule(decoder_module, in_keys=["tgt", "memory"], out_keys=["output"])
transformer = TensorDictSequential(encoder, decoder)
assert transformer.in_keys == ["src", "src_mask", "tgt"]
assert transformer.out_keys == ["memory", "output"]
```
`TensorDictSequential` allows to isolate subgraphs by querying a set of desired input / output keys:
```python
transformer.select_subsequence(out_keys=["memory"]) # returns the encoder
transformer.select_subsequence(in_keys=["tgt", "memory"]) # returns the decoder
```
</details>
Check [TensorDict tutorials](https://pytorch.github.io/tensordict/) to
learn more!
## Features
- A common [interface for environments](torchrl/envs)
which supports common libraries (OpenAI gym, deepmind control lab, etc.)<sup>(1)</sup> and state-less execution
(e.g. Model-based environments).
The [batched environments](torchrl/envs/batched_envs.py) containers allow parallel execution<sup>(2)</sup>.
A common PyTorch-first class of [tensor-specification class](torchrl/data/tensor_specs.py) is also provided.
TorchRL's environments API is simple but stringent and specific. Check the
[documentation](https://pytorch.org/rl/reference/envs.html)
and [tutorial](https://pytorch.org/rl/tutorials/pendulum.html) to learn more!
<details>
<summary>Code</summary>
```python
env_make = lambda: GymEnv("Pendulum-v1", from_pixels=True)
env_parallel = ParallelEnv(4, env_make) # creates 4 envs in parallel
tensordict = env_parallel.rollout(max_steps=20, policy=None) # random rollout (no policy given)
assert tensordict.shape == [4, 20] # 4 envs, 20 steps rollout
env_parallel.action_spec.is_in(tensordict["action"]) # spec check returns True
```
</details>
- multiprocess and distributed [data collectors](torchrl/collectors/collectors.py)<sup>(2)</sup>
that work synchronously or asynchronously.
Through the use of TensorDict, TorchRL's training loops are made very similar
to regular training loops in supervised
learning (although the "dataloader" -- read data collector -- is modified on-the-fly):
<details>
<summary>Code</summary>
```python
env_make = lambda: GymEnv("Pendulum-v1", from_pixels=True)
collector = MultiaSyncDataCollector(
[env_make, env_make],
policy=policy,
devices=["cuda:0", "cuda:0"],
total_frames=10000,
frames_per_batch=50,
...
)
for i, tensordict_data in enumerate(collector):
loss = loss_module(tensordict_data)
loss.backward()
optim.step()
optim.zero_grad()
collector.update_policy_weights_()
```
</details>
Check our [distributed collector examples](sota-implementations/distributed/collectors) to
learn more about ultra-fast data collection with TorchRL.
- efficient<sup>(2)</sup> and generic<sup>(1)</sup> [replay buffers](torchrl/data/replay_buffers/replay_buffers.py) with modularized storage:
<details>
<summary>Code</summary>
```python
storage = LazyMemmapStorage( # memory-mapped (physical) storage
cfg.buffer_size,
scratch_dir="/tmp/"
)
buffer = TensorDictPrioritizedReplayBuffer(
alpha=0.7,
beta=0.5,
collate_fn=lambda x: x,
pin_memory=device != torch.device("cpu"),
prefetch=10, # multi-threaded sampling
storage=storage
)
```
</details>
Replay buffers are also offered as wrappers around common datasets for *offline RL*:
<details>
<summary>Code</summary>
```python
from torchrl.data.replay_buffers import SamplerWithoutReplacement
from torchrl.data.datasets.d4rl import D4RLExperienceReplay
data = D4RLExperienceReplay(
"maze2d-open-v0",
split_trajs=True,
batch_size=128,
sampler=SamplerWithoutReplacement(drop_last=True),
)
for sample in data: # or alternatively sample = data.sample()
fun(sample)
```
</details>
- cross-library [environment transforms](torchrl/envs/transforms/transforms.py)<sup>(1)</sup>,
executed on device and in a vectorized fashion<sup>(2)</sup>,
which process and prepare the data coming out of the environments to be used by the agent:
<details>
<summary>Code</summary>
```python
env_make = lambda: GymEnv("Pendulum-v1", from_pixels=True)
env_base = ParallelEnv(4, env_make, device="cuda:0") # creates 4 envs in parallel
env = TransformedEnv(
env_base,
Compose(
ToTensorImage(),
ObservationNorm(loc=0.5, scale=1.0)), # executes the transforms once and on device
)
tensordict = env.reset()
assert tensordict.device == torch.device("cuda:0")
```
Other transforms include: reward scaling (`RewardScaling`), shape operations (concatenation of tensors, unsqueezing etc.), concatenation of
successive operations (`CatFrames`), resizing (`Resize`) and many more.
Unlike other libraries, the transforms are stacked as a list (and not wrapped in each other), which makes it
easy to add and remove them at will:
```python
env.insert_transform(0, NoopResetEnv()) # inserts the NoopResetEnv transform at the index 0
```
Nevertheless, transforms can access and execute operations on the parent environment:
```python
transform = env.transform[1] # gathers the second transform of the list
parent_env = transform.parent # returns the base environment of the second transform, i.e. the base env + the first transform
```
</details>
- various tools for distributed learning (e.g. [memory mapped tensors](https://github.com/pytorch/tensordict/blob/main/tensordict/memmap.py))<sup>(2)</sup>;
- various [architectures](torchrl/modules/models/) and models (e.g. [actor-critic](torchrl/modules/tensordict_module/actors.py))<sup>(1)</sup>:
<details>
<summary>Code</summary>
```python
# create an nn.Module
common_module = ConvNet(
bias_last_layer=True,
depth=None,
num_cells=[32, 64, 64],
kernel_sizes=[8, 4, 3],
strides=[4, 2, 1],
)
# Wrap it in a SafeModule, indicating what key to read in and where to
# write out the output
common_module = SafeModule(
common_module,
in_keys=["pixels"],
out_keys=["hidden"],
)
# Wrap the policy module in NormalParamsWrapper, such that the output
# tensor is split in loc and scale, and scale is mapped onto a positive space
policy_module = SafeModule(
NormalParamsWrapper(
MLP(num_cells=[64, 64], out_features=32, activation=nn.ELU)
),
in_keys=["hidden"],
out_keys=["loc", "scale"],
)
# Use a SafeProbabilisticTensorDictSequential to combine the SafeModule with a
# SafeProbabilisticModule, indicating how to build the
# torch.distribution.Distribution object and what to do with it
policy_module = SafeProbabilisticTensorDictSequential( # stochastic policy
policy_module,
SafeProbabilisticModule(
in_keys=["loc", "scale"],
out_keys="action",
distribution_class=TanhNormal,
),
)
value_module = MLP(
num_cells=[64, 64],
out_features=1,
activation=nn.ELU,
)
# Wrap the policy and value funciton in a common module
actor_value = ActorValueOperator(common_module, policy_module, value_module)
# standalone policy from this
standalone_policy = actor_value.get_policy_operator()
```
</details>
- exploration [wrappers](torchrl/modules/tensordict_module/exploration.py) and
[modules](torchrl/modules/models/exploration.py) to easily swap between exploration and exploitation<sup>(1)</sup>:
<details>
<summary>Code</summary>
```python
policy_explore = EGreedyWrapper(policy)
with set_exploration_type(ExplorationType.RANDOM):
tensordict = policy_explore(tensordict) # will use eps-greedy
with set_exploration_type(ExplorationType.MODE):
tensordict = policy_explore(tensordict) # will not use eps-greedy
```
</details>
- A series of efficient [loss modules](https://github.com/pytorch/rl/tree/main/torchrl/objectives)
and highly vectorized
[functional return and advantage](https://github.com/pytorch/rl/blob/main/torchrl/objectives/value/functional.py)
computation.
<details>
<summary>Code</summary>
### Loss modules
```python
from torchrl.objectives import DQNLoss
loss_module = DQNLoss(value_network=value_network, gamma=0.99)
tensordict = replay_buffer.sample(batch_size)
loss = loss_module(tensordict)
```
### Advantage computation
```python
from torchrl.objectives.value.functional import vec_td_lambda_return_estimate
advantage = vec_td_lambda_return_estimate(gamma, lmbda, next_state_value, reward, done, terminated)
```
</details>
- a generic [trainer class](torchrl/trainers/trainers.py)<sup>(1)</sup> that
executes the aforementioned training loop. Through a hooking mechanism,
it also supports any logging or data transformation operation at any given
time.
- various [recipes](torchrl/trainers/helpers/models.py) to build models that
correspond to the environment being deployed.
If you feel a feature is missing from the library, please submit an issue!
If you would like to contribute to new features, check our [call for contributions](https://github.com/pytorch/rl/issues/509) and our [contribution](CONTRIBUTING.md) page.
## Examples, tutorials and demos
A series of [examples](examples/) are provided with an illustrative purpose:
- [DQN](sota-implementations/dqn)
- [DDPG](sota-implementations/ddpg/ddpg.py)
- [IQL](sota-implementations/iql/iql.py)
- [CQL](sota-implementations/iql/cql.py)
- [TD3](sota-implementations/td3/td3.py)
- [A2C](examples/a2c_old/a2c.py)
- [PPO](sota-implementations/ppo/ppo.py)
- [SAC](sota-implementations/sac/sac.py)
- [REDQ](sota-implementations/redq/redq.py)
- [Dreamer](sota-implementations/dreamer/dreamer.py)
- [Decision Transformers](sota-implementations/decision_transformer)
- [RLHF](examples/rlhf)
and many more to come!
Check the [examples markdown](sota-implementations/SOTA-IMPLEMENTATIONS.md) directory for more details
about handling the various configuration settings.
We also provide [tutorials and demos](https://pytorch.org/rl/#tutorials) that give a sense of
what the library can do.
## Citation
If you're using TorchRL, please refer to this BibTeX entry to cite this work:
```
@misc{bou2023torchrl,
title={TorchRL: A data-driven decision-making library for PyTorch},
author={Albert Bou and Matteo Bettini and Sebastian Dittert and Vikash Kumar and Shagun Sodhani and Xiaomeng Yang and Gianni De Fabritiis and Vincent Moens},
year={2023},
eprint={2306.00577},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
```
## Installation
Create a conda environment where the packages will be installed.
```
conda create --name torch_rl python=3.9
conda activate torch_rl
```
**PyTorch**
Depending on the use of functorch that you want to make, you may want to
install the latest (nightly) PyTorch release or the latest stable version of PyTorch.
See [here](https://pytorch.org/get-started/locally/) for a detailed list of commands,
including `pip3` or other special installation instructions.
**Torchrl**
You can install the **latest stable release** by using
```
pip3 install torchrl
```
This should work on linux, Windows 10 and OsX (Intel or Silicon chips).
On certain Windows machines (Windows 11), one should install the library locally (see below).
The **nightly build** can be installed via
```
pip install torchrl-nightly
```
which we currently only ship for Linux and OsX (Intel) machines.
Importantly, the nightly builds require the nightly builds of PyTorch too.
To install extra dependencies, call
```
pip3 install "torchrl[atari,dm_control,gym_continuous,rendering,tests,utils,marl,checkpointing]"
```
or a subset of these.
One may also desire to install the library locally. Three main reasons can motivate this:
- the nightly/stable release isn't available for one's platform (eg, Windows 11, nightlies for Apple Silicon etc.);
- contributing to the code;
- install torchrl with a previous version of PyTorch (note that this should also be doable via a regular install followed
by a downgrade to a previous pytorch version -- but the C++ binaries will not be available.)
To install the library locally, start by cloning the repo:
```
git clone https://github.com/pytorch/rl
```
Go to the directory where you have cloned the torchrl repo and install it (after
installing `ninja`)
```
cd /path/to/torchrl/
pip install ninja -U
python setup.py develop
```
(unfortunately, `pip install -e .` will not work).
On M1 machines, this should work out-of-the-box with the nightly build of PyTorch.
If the generation of this artifact in MacOs M1 doesn't work correctly or in the execution the message
`(mach-o file, but is an incompatible architecture (have 'x86_64', need 'arm64e'))` appears, then try
```
ARCHFLAGS="-arch arm64" python setup.py develop
```
To run a quick sanity check, leave that directory (e.g. by executing `cd ~/`)
and try to import the library.
```
python -c "import torchrl"
```
This should not return any warning or error.
**Optional dependencies**
The following libraries can be installed depending on the usage one wants to
make of torchrl:
```
# diverse
pip3 install tqdm tensorboard "hydra-core>=1.1" hydra-submitit-launcher
# rendering
pip3 install moviepy
# deepmind control suite
pip3 install dm_control
# gym, atari games
pip3 install "gym[atari]" "gym[accept-rom-license]" pygame
# tests
pip3 install pytest pyyaml pytest-instafail
# tensorboard
pip3 install tensorboard
# wandb
pip3 install wandb
```
**Troubleshooting**
If a `ModuleNotFoundError: No module named ‘torchrl._torchrl` errors occurs (or
a warning indicating that the C++ binaries could not be loaded),
it means that the C++ extensions were not installed or not found.
- One common reason might be that you are trying to import torchrl from within the
git repo location. The following code snippet should return an error if
torchrl has not been installed in `develop` mode:
```
cd ~/path/to/rl/repo
python -c 'from torchrl.envs.libs.gym import GymEnv'
```
If this is the case, consider executing torchrl from another location.
- If you're not importing torchrl from within its repo location, it could be
caused by a problem during the local installation. Check the log after the
`python setup.py develop`. One common cause is a g++/C++ version discrepancy
and/or a problem with the `ninja` library.
- If the problem persists, feel free to open an issue on the topic in the repo,
we'll make our best to help!
- On **MacOs**, we recommend installing XCode first.
With Apple Silicon M1 chips, make sure you are using the arm64-built python
(e.g. [here](https://betterprogramming.pub/how-to-install-pytorch-on-apple-m1-series-512b3ad9bc6)).
Running the following lines of code
```
wget https://raw.githubusercontent.com/pytorch/pytorch/master/torch/utils/collect_env.py
python collect_env.py
```
should display
```
OS: macOS *** (arm64)
```
and not
```
OS: macOS **** (x86_64)
```
Versioning issues can cause error message of the type ```undefined symbol```
and such. For these, refer to the [versioning issues document](knowledge_base/VERSIONING_ISSUES.md)
for a complete explanation and proposed workarounds.
## Asking a question
If you spot a bug in the library, please raise an issue in this repo.
If you have a more generic question regarding RL in PyTorch, post it on
the [PyTorch forum](https://discuss.pytorch.org/c/reinforcement-learning/6).
## Contributing
Internal collaborations to torchrl are welcome! Feel free to fork, submit issues and PRs.
You can checkout the detailed contribution guide [here](CONTRIBUTING.md).
As mentioned above, a list of open contributions can be found in [here](https://github.com/pytorch/rl/issues/509).
Contributors are recommended to install [pre-commit hooks](https://pre-commit.com/) (using `pre-commit install`). pre-commit will check for linting related issues when the code is committed locally. You can disable th check by appending `-n` to your commit command: `git commit -m <commit message> -n`
## Disclaimer
This library is released as a PyTorch beta feature.
BC-breaking changes are likely to happen but they will be introduced with a deprecation
warranty after a few release cycles.
# License
TorchRL is licensed under the MIT License. See [LICENSE](LICENSE) for details.
Raw data
{
"_id": null,
"home_page": "https://github.com/pytorch/rl",
"name": "torchrl-nightly",
"maintainer": null,
"docs_url": null,
"requires_python": null,
"maintainer_email": null,
"keywords": null,
"author": "torchrl contributors",
"author_email": "vmoens@fb.com",
"download_url": null,
"platform": null,
"description": "[](https://github.com/pytorch/rl/actions/workflows/test-linux.yml)\n[](https://pytorch.org/rl/)\n[](https://pytorch.github.io/rl/dev/bench/)\n[](https://codecov.io/gh/pytorch/rl)\n[](https://twitter.com/torchrl1)\n[](https://www.python.org/downloads/)\n[](https://github.com/pytorch/rl/blob/main/LICENSE)\n<a href=\"https://pypi.org/project/torchrl\"><img src=\"https://img.shields.io/pypi/v/torchrl\" alt=\"pypi version\"></a>\n<a href=\"https://pypi.org/project/torchrl-nightly\"><img src=\"https://img.shields.io/pypi/v/torchrl-nightly?label=nightly\" alt=\"pypi nightly version\"></a>\n[](https://pepy.tech/project/torchrl)\n[)](https://pepy.tech/project/torchrl-nightly)\n[](https://discord.gg/cZs26Qq3Dd)\n\n# TorchRL\n\n<p align=\"center\">\n <img src=\"docs/source/_static/img/icon.png\" width=\"200\" >\n</p>\n\n[**Documentation**](#documentation-and-knowledge-base) | [**TensorDict**](#writing-simplified-and-portable-rl-codebase-with-tensordict) |\n[**Features**](#features) | [**Examples, tutorials and demos**](#examples-tutorials-and-demos) | [**Citation**](#citation) | [**Installation**](#installation) |\n[**Asking a question**](#asking-a-question) | [**Contributing**](#contributing)\n\n**TorchRL** is an open-source Reinforcement Learning (RL) library for PyTorch.\n\nIt provides pytorch and **python-first**, low and high level abstractions for RL that are intended to be **efficient**, **modular**, **documented** and properly **tested**.\nThe code is aimed at supporting research in RL. Most of it is written in python in a highly modular way, such that researchers can easily swap components, transform them or write new ones with little effort.\n\nThis repo attempts to align with the existing pytorch ecosystem libraries in that it has a dataset pillar ([torchrl/envs](torchrl/envs)), [transforms](torchrl/envs/transforms), [models](torchrl/modules), data utilities (e.g. collectors and containers), etc.\nTorchRL aims at having as few dependencies as possible (python standard library, numpy and pytorch). Common environment libraries (e.g. OpenAI gym) are only optional.\n\nOn the low-level end, torchrl comes with a set of highly re-usable functionals for [cost functions](torchrl/objectives/costs), [returns](torchrl/objectives/returns) and data processing.\n\nTorchRL aims at (1) a high modularity and (2) good runtime performance. Read the [full paper](https://arxiv.org/abs/2306.00577) for a more curated description of the library.\n\n## Getting started\n\nCheck our [Getting Started tutorials](https://pytorch.org/rl/index.html#getting-started) for quickly ramp up with the basic \nfeatures of the library!\n\n## Documentation and knowledge base\n\nThe TorchRL documentation can be found [here](https://pytorch.org/rl).\nIt contains tutorials and the API reference.\n\nTorchRL also provides a RL knowledge base to help you debug your code, or simply\nlearn the basics of RL. Check it out [here](https://pytorch.org/rl/reference/knowledge_base.html).\n\nWe have some introductory videos for you to get to know the library better, check them out:\n\n- [TorchRL intro at PyTorch day 2022](https://youtu.be/cIKMhZoykEE)\n- [PyTorch 2.0 Q&A: TorchRL](https://www.youtube.com/live/myEfUoYrbts?feature=share)\n\n## Writing simplified and portable RL codebase with `TensorDict`\n\nRL algorithms are very heterogeneous, and it can be hard to recycle a codebase\nacross settings (e.g. from online to offline, from state-based to pixel-based \nlearning).\nTorchRL solves this problem through [`TensorDict`](https://github.com/pytorch/tensordict/),\na convenient data structure<sup>(1)</sup> that can be used to streamline one's\nRL codebase.\nWith this tool, one can write a *complete PPO training script in less than 100\nlines of code*!\n\n <details>\n <summary>Code</summary>\n\n ```python\n import torch\n from tensordict.nn import TensorDictModule\n from tensordict.nn.distributions import NormalParamExtractor\n from torch import nn\n \n from torchrl.collectors import SyncDataCollector\n from torchrl.data.replay_buffers import TensorDictReplayBuffer, \\\n LazyTensorStorage, SamplerWithoutReplacement\n from torchrl.envs.libs.gym import GymEnv\n from torchrl.modules import ProbabilisticActor, ValueOperator, TanhNormal\n from torchrl.objectives import ClipPPOLoss\n from torchrl.objectives.value import GAE\n \n env = GymEnv(\"Pendulum-v1\")\n model = TensorDictModule(\n nn.Sequential(\n nn.Linear(3, 128), nn.Tanh(),\n nn.Linear(128, 128), nn.Tanh(),\n nn.Linear(128, 128), nn.Tanh(),\n nn.Linear(128, 2),\n NormalParamExtractor()\n ),\n in_keys=[\"observation\"],\n out_keys=[\"loc\", \"scale\"]\n )\n critic = ValueOperator(\n nn.Sequential(\n nn.Linear(3, 128), nn.Tanh(),\n nn.Linear(128, 128), nn.Tanh(),\n nn.Linear(128, 128), nn.Tanh(),\n nn.Linear(128, 1),\n ),\n in_keys=[\"observation\"],\n )\n actor = ProbabilisticActor(\n model,\n in_keys=[\"loc\", \"scale\"],\n distribution_class=TanhNormal,\n distribution_kwargs={\"min\": -1.0, \"max\": 1.0},\n return_log_prob=True\n )\n buffer = TensorDictReplayBuffer(\n LazyTensorStorage(1000),\n SamplerWithoutReplacement()\n )\n collector = SyncDataCollector(\n env,\n actor,\n frames_per_batch=1000,\n total_frames=1_000_000\n )\n loss_fn = ClipPPOLoss(actor, critic, gamma=0.99)\n optim = torch.optim.Adam(loss_fn.parameters(), lr=2e-4)\n adv_fn = GAE(value_network=critic, gamma=0.99, lmbda=0.95, average_gae=True)\n for data in collector: # collect data\n for epoch in range(10):\n adv_fn(data) # compute advantage\n buffer.extend(data.view(-1))\n for i in range(20): # consume data\n sample = buffer.sample(50) # mini-batch\n loss_vals = loss_fn(sample)\n loss_val = sum(\n value for key, value in loss_vals.items() if\n key.startswith(\"loss\")\n )\n loss_val.backward()\n optim.step()\n optim.zero_grad()\n print(f\"avg reward: {data['next', 'reward'].mean().item(): 4.4f}\")\n ```\n </details>\n\nHere is an example of how the [environment API](https://pytorch.org/rl/reference/envs.html)\nrelies on tensordict to carry data from one function to another during a rollout\nexecution:\n\n\n`TensorDict` makes it easy to re-use pieces of code across environments, models and\nalgorithms.\n <details>\n <summary>Code</summary>\n \n For instance, here's how to code a rollout in TorchRL:\n\n ```diff\n - obs, done = env.reset()\n + tensordict = env.reset()\n policy = SafeModule(\n model,\n in_keys=[\"observation_pixels\", \"observation_vector\"],\n out_keys=[\"action\"],\n )\n out = []\n for i in range(n_steps):\n - action, log_prob = policy(obs)\n - next_obs, reward, done, info = env.step(action)\n - out.append((obs, next_obs, action, log_prob, reward, done))\n - obs = next_obs\n + tensordict = policy(tensordict)\n + tensordict = env.step(tensordict)\n + out.append(tensordict)\n + tensordict = step_mdp(tensordict) # renames next_observation_* keys to observation_*\n - obs, next_obs, action, log_prob, reward, done = [torch.stack(vals, 0) for vals in zip(*out)]\n + out = torch.stack(out, 0) # TensorDict supports multiple tensor operations\n ```\n </details>\n\nUsing this, TorchRL abstracts away the input / output signatures of the modules, env, \ncollectors, replay buffers and losses of the library, allowing all primitives\nto be easily recycled across settings.\n\n <details>\n <summary>Code</summary>\n\n Here's another example of an off-policy training loop in TorchRL (assuming \n that a data collector, a replay buffer, a loss and an optimizer have been instantiated):\n \n ```diff\n - for i, (obs, next_obs, action, hidden_state, reward, done) in enumerate(collector):\n + for i, tensordict in enumerate(collector):\n - replay_buffer.add((obs, next_obs, action, log_prob, reward, done))\n + replay_buffer.add(tensordict)\n for j in range(num_optim_steps):\n - obs, next_obs, action, hidden_state, reward, done = replay_buffer.sample(batch_size)\n - loss = loss_fn(obs, next_obs, action, hidden_state, reward, done)\n + tensordict = replay_buffer.sample(batch_size)\n + loss = loss_fn(tensordict)\n loss.backward()\n optim.step()\n optim.zero_grad()\n ```\n This training loop can be re-used across algorithms as it makes a minimal number of assumptions about the structure of the data.\n </details>\n\n TensorDict supports multiple tensor operations on its device and shape\n (the shape of TensorDict, or its batch size, is the common arbitrary N first dimensions of all its contained tensors):\n\n <details>\n <summary>Code</summary>\n\n ```python\n # stack and cat\n tensordict = torch.stack(list_of_tensordicts, 0)\n tensordict = torch.cat(list_of_tensordicts, 0)\n # reshape\n tensordict = tensordict.view(-1)\n tensordict = tensordict.permute(0, 2, 1)\n tensordict = tensordict.unsqueeze(-1)\n tensordict = tensordict.squeeze(-1)\n # indexing\n tensordict = tensordict[:2]\n tensordict[:, 2] = sub_tensordict\n # device and memory location\n tensordict.cuda()\n tensordict.to(\"cuda:1\")\n tensordict.share_memory_()\n ```\n </details>\n\nTensorDict comes with a dedicated [`tensordict.nn`](https://pytorch.github.io/tensordict/reference/nn.html)\nmodule that contains everything you might need to write your model with it.\nAnd it is `functorch` and `torch.compile` compatible!\n\n <details>\n <summary>Code</summary>\n\n ```diff\n transformer_model = nn.Transformer(nhead=16, num_encoder_layers=12)\n + td_module = SafeModule(transformer_model, in_keys=[\"src\", \"tgt\"], out_keys=[\"out\"])\n src = torch.rand((10, 32, 512))\n tgt = torch.rand((20, 32, 512))\n + tensordict = TensorDict({\"src\": src, \"tgt\": tgt}, batch_size=[20, 32])\n - out = transformer_model(src, tgt)\n + td_module(tensordict)\n + out = tensordict[\"out\"]\n ```\n\n The `TensorDictSequential` class allows to branch sequences of `nn.Module` instances in a highly modular way.\n For instance, here is an implementation of a transformer using the encoder and decoder blocks:\n ```python\n encoder_module = TransformerEncoder(...)\n encoder = TensorDictSequential(encoder_module, in_keys=[\"src\", \"src_mask\"], out_keys=[\"memory\"])\n decoder_module = TransformerDecoder(...)\n decoder = TensorDictModule(decoder_module, in_keys=[\"tgt\", \"memory\"], out_keys=[\"output\"])\n transformer = TensorDictSequential(encoder, decoder)\n assert transformer.in_keys == [\"src\", \"src_mask\", \"tgt\"]\n assert transformer.out_keys == [\"memory\", \"output\"]\n ```\n\n `TensorDictSequential` allows to isolate subgraphs by querying a set of desired input / output keys:\n ```python\n transformer.select_subsequence(out_keys=[\"memory\"]) # returns the encoder\n transformer.select_subsequence(in_keys=[\"tgt\", \"memory\"]) # returns the decoder\n ```\n </details>\n\n Check [TensorDict tutorials](https://pytorch.github.io/tensordict/) to\n learn more!\n\n\n## Features\n\n- A common [interface for environments](torchrl/envs)\n which supports common libraries (OpenAI gym, deepmind control lab, etc.)<sup>(1)</sup> and state-less execution \n (e.g. Model-based environments).\n The [batched environments](torchrl/envs/batched_envs.py) containers allow parallel execution<sup>(2)</sup>.\n A common PyTorch-first class of [tensor-specification class](torchrl/data/tensor_specs.py) is also provided.\n TorchRL's environments API is simple but stringent and specific. Check the \n [documentation](https://pytorch.org/rl/reference/envs.html)\n and [tutorial](https://pytorch.org/rl/tutorials/pendulum.html) to learn more!\n <details>\n <summary>Code</summary>\n\n ```python\n env_make = lambda: GymEnv(\"Pendulum-v1\", from_pixels=True)\n env_parallel = ParallelEnv(4, env_make) # creates 4 envs in parallel\n tensordict = env_parallel.rollout(max_steps=20, policy=None) # random rollout (no policy given)\n assert tensordict.shape == [4, 20] # 4 envs, 20 steps rollout\n env_parallel.action_spec.is_in(tensordict[\"action\"]) # spec check returns True\n ```\n </details>\n\n- multiprocess and distributed [data collectors](torchrl/collectors/collectors.py)<sup>(2)</sup>\n that work synchronously or asynchronously.\n Through the use of TensorDict, TorchRL's training loops are made very similar\n to regular training loops in supervised\n learning (although the \"dataloader\" -- read data collector -- is modified on-the-fly):\n <details>\n <summary>Code</summary>\n\n ```python\n env_make = lambda: GymEnv(\"Pendulum-v1\", from_pixels=True)\n collector = MultiaSyncDataCollector(\n [env_make, env_make],\n policy=policy,\n devices=[\"cuda:0\", \"cuda:0\"],\n total_frames=10000,\n frames_per_batch=50,\n ...\n )\n for i, tensordict_data in enumerate(collector):\n loss = loss_module(tensordict_data)\n loss.backward()\n optim.step()\n optim.zero_grad()\n collector.update_policy_weights_()\n ```\n </details>\n\n Check our [distributed collector examples](sota-implementations/distributed/collectors) to\n learn more about ultra-fast data collection with TorchRL.\n\n- efficient<sup>(2)</sup> and generic<sup>(1)</sup> [replay buffers](torchrl/data/replay_buffers/replay_buffers.py) with modularized storage:\n <details>\n <summary>Code</summary>\n\n ```python\n storage = LazyMemmapStorage( # memory-mapped (physical) storage\n cfg.buffer_size,\n scratch_dir=\"/tmp/\"\n )\n buffer = TensorDictPrioritizedReplayBuffer(\n alpha=0.7,\n beta=0.5,\n collate_fn=lambda x: x,\n pin_memory=device != torch.device(\"cpu\"),\n prefetch=10, # multi-threaded sampling\n storage=storage\n )\n ```\n </details>\n\n Replay buffers are also offered as wrappers around common datasets for *offline RL*:\n <details>\n <summary>Code</summary>\n\n ```python\n from torchrl.data.replay_buffers import SamplerWithoutReplacement\n from torchrl.data.datasets.d4rl import D4RLExperienceReplay\n data = D4RLExperienceReplay(\n \"maze2d-open-v0\",\n split_trajs=True,\n batch_size=128,\n sampler=SamplerWithoutReplacement(drop_last=True),\n )\n for sample in data: # or alternatively sample = data.sample()\n fun(sample)\n ```\n </details>\n\n\n- cross-library [environment transforms](torchrl/envs/transforms/transforms.py)<sup>(1)</sup>,\n executed on device and in a vectorized fashion<sup>(2)</sup>,\n which process and prepare the data coming out of the environments to be used by the agent:\n <details>\n <summary>Code</summary>\n\n ```python\n env_make = lambda: GymEnv(\"Pendulum-v1\", from_pixels=True)\n env_base = ParallelEnv(4, env_make, device=\"cuda:0\") # creates 4 envs in parallel\n env = TransformedEnv(\n env_base,\n Compose(\n ToTensorImage(),\n ObservationNorm(loc=0.5, scale=1.0)), # executes the transforms once and on device\n )\n tensordict = env.reset()\n assert tensordict.device == torch.device(\"cuda:0\")\n ```\n Other transforms include: reward scaling (`RewardScaling`), shape operations (concatenation of tensors, unsqueezing etc.), concatenation of\n successive operations (`CatFrames`), resizing (`Resize`) and many more.\n\n Unlike other libraries, the transforms are stacked as a list (and not wrapped in each other), which makes it\n easy to add and remove them at will:\n ```python\n env.insert_transform(0, NoopResetEnv()) # inserts the NoopResetEnv transform at the index 0\n ```\n Nevertheless, transforms can access and execute operations on the parent environment:\n ```python\n transform = env.transform[1] # gathers the second transform of the list\n parent_env = transform.parent # returns the base environment of the second transform, i.e. the base env + the first transform\n ```\n </details>\n\n- various tools for distributed learning (e.g. [memory mapped tensors](https://github.com/pytorch/tensordict/blob/main/tensordict/memmap.py))<sup>(2)</sup>;\n- various [architectures](torchrl/modules/models/) and models (e.g. [actor-critic](torchrl/modules/tensordict_module/actors.py))<sup>(1)</sup>:\n <details>\n <summary>Code</summary>\n\n ```python\n # create an nn.Module\n common_module = ConvNet(\n bias_last_layer=True,\n depth=None,\n num_cells=[32, 64, 64],\n kernel_sizes=[8, 4, 3],\n strides=[4, 2, 1],\n )\n # Wrap it in a SafeModule, indicating what key to read in and where to\n # write out the output\n common_module = SafeModule(\n common_module,\n in_keys=[\"pixels\"],\n out_keys=[\"hidden\"],\n )\n # Wrap the policy module in NormalParamsWrapper, such that the output\n # tensor is split in loc and scale, and scale is mapped onto a positive space\n policy_module = SafeModule(\n NormalParamsWrapper(\n MLP(num_cells=[64, 64], out_features=32, activation=nn.ELU)\n ),\n in_keys=[\"hidden\"],\n out_keys=[\"loc\", \"scale\"],\n )\n # Use a SafeProbabilisticTensorDictSequential to combine the SafeModule with a\n # SafeProbabilisticModule, indicating how to build the\n # torch.distribution.Distribution object and what to do with it\n policy_module = SafeProbabilisticTensorDictSequential( # stochastic policy\n policy_module,\n SafeProbabilisticModule(\n in_keys=[\"loc\", \"scale\"],\n out_keys=\"action\",\n distribution_class=TanhNormal,\n ),\n )\n value_module = MLP(\n num_cells=[64, 64],\n out_features=1,\n activation=nn.ELU,\n )\n # Wrap the policy and value funciton in a common module\n actor_value = ActorValueOperator(common_module, policy_module, value_module)\n # standalone policy from this\n standalone_policy = actor_value.get_policy_operator()\n ```\n </details>\n\n- exploration [wrappers](torchrl/modules/tensordict_module/exploration.py) and\n [modules](torchrl/modules/models/exploration.py) to easily swap between exploration and exploitation<sup>(1)</sup>:\n <details>\n <summary>Code</summary>\n\n ```python\n policy_explore = EGreedyWrapper(policy)\n with set_exploration_type(ExplorationType.RANDOM):\n tensordict = policy_explore(tensordict) # will use eps-greedy\n with set_exploration_type(ExplorationType.MODE):\n tensordict = policy_explore(tensordict) # will not use eps-greedy\n ```\n </details>\n\n- A series of efficient [loss modules](https://github.com/pytorch/rl/tree/main/torchrl/objectives)\n and highly vectorized\n [functional return and advantage](https://github.com/pytorch/rl/blob/main/torchrl/objectives/value/functional.py)\n computation.\n\n <details>\n <summary>Code</summary>\n\n ### Loss modules\n ```python\n from torchrl.objectives import DQNLoss\n loss_module = DQNLoss(value_network=value_network, gamma=0.99)\n tensordict = replay_buffer.sample(batch_size)\n loss = loss_module(tensordict)\n ```\n\n ### Advantage computation\n ```python\n from torchrl.objectives.value.functional import vec_td_lambda_return_estimate\n advantage = vec_td_lambda_return_estimate(gamma, lmbda, next_state_value, reward, done, terminated)\n ```\n\n </details>\n\n- a generic [trainer class](torchrl/trainers/trainers.py)<sup>(1)</sup> that\n executes the aforementioned training loop. Through a hooking mechanism,\n it also supports any logging or data transformation operation at any given\n time.\n\n- various [recipes](torchrl/trainers/helpers/models.py) to build models that\n correspond to the environment being deployed.\n\nIf you feel a feature is missing from the library, please submit an issue!\nIf you would like to contribute to new features, check our [call for contributions](https://github.com/pytorch/rl/issues/509) and our [contribution](CONTRIBUTING.md) page.\n\n\n## Examples, tutorials and demos\n\nA series of [examples](examples/) are provided with an illustrative purpose:\n- [DQN](sota-implementations/dqn)\n- [DDPG](sota-implementations/ddpg/ddpg.py)\n- [IQL](sota-implementations/iql/iql.py)\n- [CQL](sota-implementations/iql/cql.py)\n- [TD3](sota-implementations/td3/td3.py)\n- [A2C](examples/a2c_old/a2c.py)\n- [PPO](sota-implementations/ppo/ppo.py)\n- [SAC](sota-implementations/sac/sac.py)\n- [REDQ](sota-implementations/redq/redq.py)\n- [Dreamer](sota-implementations/dreamer/dreamer.py)\n- [Decision Transformers](sota-implementations/decision_transformer)\n- [RLHF](examples/rlhf)\n\nand many more to come!\n\nCheck the [examples markdown](sota-implementations/SOTA-IMPLEMENTATIONS.md) directory for more details \nabout handling the various configuration settings.\n\nWe also provide [tutorials and demos](https://pytorch.org/rl/#tutorials) that give a sense of\nwhat the library can do.\n\n## Citation\n\nIf you're using TorchRL, please refer to this BibTeX entry to cite this work:\n```\n@misc{bou2023torchrl,\n title={TorchRL: A data-driven decision-making library for PyTorch}, \n author={Albert Bou and Matteo Bettini and Sebastian Dittert and Vikash Kumar and Shagun Sodhani and Xiaomeng Yang and Gianni De Fabritiis and Vincent Moens},\n year={2023},\n eprint={2306.00577},\n archivePrefix={arXiv},\n primaryClass={cs.LG}\n}\n```\n\n## Installation\n\nCreate a conda environment where the packages will be installed.\n\n```\nconda create --name torch_rl python=3.9\nconda activate torch_rl\n```\n\n**PyTorch**\n\nDepending on the use of functorch that you want to make, you may want to \ninstall the latest (nightly) PyTorch release or the latest stable version of PyTorch.\nSee [here](https://pytorch.org/get-started/locally/) for a detailed list of commands, \nincluding `pip3` or other special installation instructions.\n\n**Torchrl**\n\nYou can install the **latest stable release** by using\n```\npip3 install torchrl\n```\nThis should work on linux, Windows 10 and OsX (Intel or Silicon chips).\nOn certain Windows machines (Windows 11), one should install the library locally (see below).\n\nThe **nightly build** can be installed via\n```\npip install torchrl-nightly\n```\nwhich we currently only ship for Linux and OsX (Intel) machines.\nImportantly, the nightly builds require the nightly builds of PyTorch too.\n\nTo install extra dependencies, call\n```\npip3 install \"torchrl[atari,dm_control,gym_continuous,rendering,tests,utils,marl,checkpointing]\"\n```\nor a subset of these.\n\nOne may also desire to install the library locally. Three main reasons can motivate this:\n- the nightly/stable release isn't available for one's platform (eg, Windows 11, nightlies for Apple Silicon etc.);\n- contributing to the code;\n- install torchrl with a previous version of PyTorch (note that this should also be doable via a regular install followed\n by a downgrade to a previous pytorch version -- but the C++ binaries will not be available.)\n\nTo install the library locally, start by cloning the repo:\n```\ngit clone https://github.com/pytorch/rl\n```\n\nGo to the directory where you have cloned the torchrl repo and install it (after\ninstalling `ninja`)\n```\ncd /path/to/torchrl/\npip install ninja -U\npython setup.py develop\n```\n\n(unfortunately, `pip install -e .` will not work).\n\nOn M1 machines, this should work out-of-the-box with the nightly build of PyTorch.\nIf the generation of this artifact in MacOs M1 doesn't work correctly or in the execution the message\n`(mach-o file, but is an incompatible architecture (have 'x86_64', need 'arm64e'))` appears, then try\n\n```\nARCHFLAGS=\"-arch arm64\" python setup.py develop\n```\n\nTo run a quick sanity check, leave that directory (e.g. by executing `cd ~/`)\nand try to import the library.\n```\npython -c \"import torchrl\"\n```\nThis should not return any warning or error.\n\n**Optional dependencies**\n\nThe following libraries can be installed depending on the usage one wants to\nmake of torchrl:\n```\n# diverse\npip3 install tqdm tensorboard \"hydra-core>=1.1\" hydra-submitit-launcher\n\n# rendering\npip3 install moviepy\n\n# deepmind control suite\npip3 install dm_control\n\n# gym, atari games\npip3 install \"gym[atari]\" \"gym[accept-rom-license]\" pygame\n\n# tests\npip3 install pytest pyyaml pytest-instafail\n\n# tensorboard\npip3 install tensorboard\n\n# wandb\npip3 install wandb\n```\n\n**Troubleshooting**\n\nIf a `ModuleNotFoundError: No module named \u2018torchrl._torchrl` errors occurs (or\na warning indicating that the C++ binaries could not be loaded),\nit means that the C++ extensions were not installed or not found.\n\n- One common reason might be that you are trying to import torchrl from within the\n git repo location. The following code snippet should return an error if\n torchrl has not been installed in `develop` mode:\n ```\n cd ~/path/to/rl/repo\n python -c 'from torchrl.envs.libs.gym import GymEnv'\n ```\n If this is the case, consider executing torchrl from another location.\n- If you're not importing torchrl from within its repo location, it could be\n caused by a problem during the local installation. Check the log after the\n `python setup.py develop`. One common cause is a g++/C++ version discrepancy\n and/or a problem with the `ninja` library.\n- If the problem persists, feel free to open an issue on the topic in the repo,\n we'll make our best to help!\n- On **MacOs**, we recommend installing XCode first. \n With Apple Silicon M1 chips, make sure you are using the arm64-built python\n (e.g. [here](https://betterprogramming.pub/how-to-install-pytorch-on-apple-m1-series-512b3ad9bc6)).\n Running the following lines of code\n ```\n wget https://raw.githubusercontent.com/pytorch/pytorch/master/torch/utils/collect_env.py\n python collect_env.py\n ```\n should display\n ```\n OS: macOS *** (arm64)\n ```\n and not\n ```\n OS: macOS **** (x86_64)\n ```\n\nVersioning issues can cause error message of the type ```undefined symbol```\nand such. For these, refer to the [versioning issues document](knowledge_base/VERSIONING_ISSUES.md)\nfor a complete explanation and proposed workarounds.\n\n## Asking a question\n\nIf you spot a bug in the library, please raise an issue in this repo.\n\nIf you have a more generic question regarding RL in PyTorch, post it on\nthe [PyTorch forum](https://discuss.pytorch.org/c/reinforcement-learning/6).\n\n## Contributing\n\nInternal collaborations to torchrl are welcome! Feel free to fork, submit issues and PRs.\nYou can checkout the detailed contribution guide [here](CONTRIBUTING.md).\nAs mentioned above, a list of open contributions can be found in [here](https://github.com/pytorch/rl/issues/509).\n\nContributors are recommended to install [pre-commit hooks](https://pre-commit.com/) (using `pre-commit install`). pre-commit will check for linting related issues when the code is committed locally. You can disable th check by appending `-n` to your commit command: `git commit -m <commit message> -n`\n\n\n## Disclaimer\n\nThis library is released as a PyTorch beta feature.\nBC-breaking changes are likely to happen but they will be introduced with a deprecation\nwarranty after a few release cycles.\n\n# License\nTorchRL is licensed under the MIT License. See [LICENSE](LICENSE) for details.\n\n\n",
"bugtrack_url": null,
"license": "MIT",
"summary": null,
"version": "2024.4.2",
"project_urls": {
"Homepage": "https://github.com/pytorch/rl"
},
"split_keywords": [],
"urls": [
{
"comment_text": "",
"digests": {
"blake2b_256": "29758ec250b69c27fd6cdc40fdf217cc165d89e6c167036510c0695c872ddae8",
"md5": "6c1e6cbe4e0eddcf430c040027291969",
"sha256": "7262b2edcd7a783036c3df8ac0bef52658474a628fd32a64eb06867f79109556"
},
"downloads": -1,
"filename": "torchrl_nightly-2024.4.2-cp310-cp310-macosx_10_15_x86_64.whl",
"has_sig": false,
"md5_digest": "6c1e6cbe4e0eddcf430c040027291969",
"packagetype": "bdist_wheel",
"python_version": "cp310",
"requires_python": null,
"size": 933144,
"upload_time": "2024-04-02T11:23:58",
"upload_time_iso_8601": "2024-04-02T11:23:58.980344Z",
"url": "https://files.pythonhosted.org/packages/29/75/8ec250b69c27fd6cdc40fdf217cc165d89e6c167036510c0695c872ddae8/torchrl_nightly-2024.4.2-cp310-cp310-macosx_10_15_x86_64.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "ea5983ad3296d3f855b7fc3b1d2d9ef9ce776983d92ffb70dceafaf64a37554e",
"md5": "46cecf1293c55394f4ae58d44c4b0627",
"sha256": "62888a57e6967349ccf03690b1556bcadb31f01b081b97b6a744036aa7bea4a0"
},
"downloads": -1,
"filename": "torchrl_nightly-2024.4.2-cp310-cp310-manylinux1_x86_64.whl",
"has_sig": false,
"md5_digest": "46cecf1293c55394f4ae58d44c4b0627",
"packagetype": "bdist_wheel",
"python_version": "cp310",
"requires_python": null,
"size": 6248849,
"upload_time": "2024-04-02T11:26:39",
"upload_time_iso_8601": "2024-04-02T11:26:39.846101Z",
"url": "https://files.pythonhosted.org/packages/ea/59/83ad3296d3f855b7fc3b1d2d9ef9ce776983d92ffb70dceafaf64a37554e/torchrl_nightly-2024.4.2-cp310-cp310-manylinux1_x86_64.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "9402a8a661038dde17897ca6853cb48798b0b800a0154b5301bdfda7403426a6",
"md5": "815fcc4194052d2f330439c693586361",
"sha256": "cf13d34dd813785af6a571f7f0601375462e96443395da97ec30f3c679342dea"
},
"downloads": -1,
"filename": "torchrl_nightly-2024.4.2-cp310-cp310-win_amd64.whl",
"has_sig": false,
"md5_digest": "815fcc4194052d2f330439c693586361",
"packagetype": "bdist_wheel",
"python_version": "cp310",
"requires_python": null,
"size": 890942,
"upload_time": "2024-04-02T11:30:46",
"upload_time_iso_8601": "2024-04-02T11:30:46.736340Z",
"url": "https://files.pythonhosted.org/packages/94/02/a8a661038dde17897ca6853cb48798b0b800a0154b5301bdfda7403426a6/torchrl_nightly-2024.4.2-cp310-cp310-win_amd64.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "cca2174a4a1eeb24729ae6e7c21cf8784be13a508c6aeef49e1a4f9fab58b234",
"md5": "687feb6f156df59f00fbd67b8fe674b4",
"sha256": "800e25cd31820135b3eded0466a0a35b33b6f61f50d67fd4cb18276e322a4fb2"
},
"downloads": -1,
"filename": "torchrl_nightly-2024.4.2-cp311-cp311-macosx_10_9_universal2.whl",
"has_sig": false,
"md5_digest": "687feb6f156df59f00fbd67b8fe674b4",
"packagetype": "bdist_wheel",
"python_version": "cp311",
"requires_python": null,
"size": 1131542,
"upload_time": "2024-04-02T11:24:00",
"upload_time_iso_8601": "2024-04-02T11:24:00.046166Z",
"url": "https://files.pythonhosted.org/packages/cc/a2/174a4a1eeb24729ae6e7c21cf8784be13a508c6aeef49e1a4f9fab58b234/torchrl_nightly-2024.4.2-cp311-cp311-macosx_10_9_universal2.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "d8fb97f165120b636e82b96a1b64f6b10b4d9ebd54f58dec4cac63900f1908ad",
"md5": "b33cb53737b44f66b86ec14dec8ab6c4",
"sha256": "1a07c3dda22e2008d474497cbe5bc0f1815db12283e93350da1cdb5fa4ef2132"
},
"downloads": -1,
"filename": "torchrl_nightly-2024.4.2-cp311-cp311-manylinux1_x86_64.whl",
"has_sig": false,
"md5_digest": "b33cb53737b44f66b86ec14dec8ab6c4",
"packagetype": "bdist_wheel",
"python_version": "cp311",
"requires_python": null,
"size": 6269167,
"upload_time": "2024-04-02T11:26:43",
"upload_time_iso_8601": "2024-04-02T11:26:43.447743Z",
"url": "https://files.pythonhosted.org/packages/d8/fb/97f165120b636e82b96a1b64f6b10b4d9ebd54f58dec4cac63900f1908ad/torchrl_nightly-2024.4.2-cp311-cp311-manylinux1_x86_64.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "e4faafe0bd9b3142a77f57aec37c90390e79b54769d5c25617daf2f52b7702ba",
"md5": "322438b87852d1d32e9b25e9cbaca943",
"sha256": "97239f9cd9508b23bf161151150ad6fd2de776bd9dc40f65f2cedcc913ba0aaf"
},
"downloads": -1,
"filename": "torchrl_nightly-2024.4.2-cp311-cp311-win_amd64.whl",
"has_sig": false,
"md5_digest": "322438b87852d1d32e9b25e9cbaca943",
"packagetype": "bdist_wheel",
"python_version": "cp311",
"requires_python": null,
"size": 887491,
"upload_time": "2024-04-02T11:30:39",
"upload_time_iso_8601": "2024-04-02T11:30:39.657473Z",
"url": "https://files.pythonhosted.org/packages/e4/fa/afe0bd9b3142a77f57aec37c90390e79b54769d5c25617daf2f52b7702ba/torchrl_nightly-2024.4.2-cp311-cp311-win_amd64.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "9591562243c9ae308395ccff9b1fb46f990f6423c5a2c79fb52d45d295fe9acc",
"md5": "7b931314a9d73acfeedbd5317f3809d1",
"sha256": "5a8c703847a994dccc32286941c6ad0badc788fb4f3323f46e9f4845c4b4455e"
},
"downloads": -1,
"filename": "torchrl_nightly-2024.4.2-cp38-cp38-macosx_11_0_x86_64.whl",
"has_sig": false,
"md5_digest": "7b931314a9d73acfeedbd5317f3809d1",
"packagetype": "bdist_wheel",
"python_version": "cp38",
"requires_python": null,
"size": 932960,
"upload_time": "2024-04-02T11:23:58",
"upload_time_iso_8601": "2024-04-02T11:23:58.396499Z",
"url": "https://files.pythonhosted.org/packages/95/91/562243c9ae308395ccff9b1fb46f990f6423c5a2c79fb52d45d295fe9acc/torchrl_nightly-2024.4.2-cp38-cp38-macosx_11_0_x86_64.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "bbf0cfcd4f264a227f3007c4ff84fefc49f73bf330462fd39e2c075b03c3de0c",
"md5": "5195166a04cc66283359cd98cfe43574",
"sha256": "57e332e2f3c14d4f6d30b74a1f3f1f2975ca4a0676ad37f62c46016caa7c7391"
},
"downloads": -1,
"filename": "torchrl_nightly-2024.4.2-cp38-cp38-manylinux1_x86_64.whl",
"has_sig": false,
"md5_digest": "5195166a04cc66283359cd98cfe43574",
"packagetype": "bdist_wheel",
"python_version": "cp38",
"requires_python": null,
"size": 6260293,
"upload_time": "2024-04-02T11:26:40",
"upload_time_iso_8601": "2024-04-02T11:26:40.946495Z",
"url": "https://files.pythonhosted.org/packages/bb/f0/cfcd4f264a227f3007c4ff84fefc49f73bf330462fd39e2c075b03c3de0c/torchrl_nightly-2024.4.2-cp38-cp38-manylinux1_x86_64.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "bdd2c5e00fc0c5dcb636f4795ecb5f781f565de4e805d8ede2286bdcd15626cc",
"md5": "cc28dfc16de5d19fb36d50cc343688aa",
"sha256": "30fb84a7880c0f648b7f67b32b39447ec4e02450c18d5491573d6726a78a8ddd"
},
"downloads": -1,
"filename": "torchrl_nightly-2024.4.2-cp38-cp38-win_amd64.whl",
"has_sig": false,
"md5_digest": "cc28dfc16de5d19fb36d50cc343688aa",
"packagetype": "bdist_wheel",
"python_version": "cp38",
"requires_python": null,
"size": 890936,
"upload_time": "2024-04-02T11:30:47",
"upload_time_iso_8601": "2024-04-02T11:30:47.376351Z",
"url": "https://files.pythonhosted.org/packages/bd/d2/c5e00fc0c5dcb636f4795ecb5f781f565de4e805d8ede2286bdcd15626cc/torchrl_nightly-2024.4.2-cp38-cp38-win_amd64.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "9f069f5b31bffa43920f36e343664d5472c45f2921778d1214bdb048aa8e7244",
"md5": "7e7edfc14bbbae1d30b98c73ac67b60c",
"sha256": "d9fd826d51bd2c9e56b676be00bc84af7c2556d5d3faa1ecd16af7e31a40adcf"
},
"downloads": -1,
"filename": "torchrl_nightly-2024.4.2-cp39-cp39-macosx_11_0_x86_64.whl",
"has_sig": false,
"md5_digest": "7e7edfc14bbbae1d30b98c73ac67b60c",
"packagetype": "bdist_wheel",
"python_version": "cp39",
"requires_python": null,
"size": 933412,
"upload_time": "2024-04-02T11:23:56",
"upload_time_iso_8601": "2024-04-02T11:23:56.643778Z",
"url": "https://files.pythonhosted.org/packages/9f/06/9f5b31bffa43920f36e343664d5472c45f2921778d1214bdb048aa8e7244/torchrl_nightly-2024.4.2-cp39-cp39-macosx_11_0_x86_64.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "ec9c4685d1a48e54aa652a0b58559391c412d4cddba24bb40c77f47fdd6d406f",
"md5": "85b41e3ae1467aef41d1d59780f72b47",
"sha256": "d1f9cc99fd349cd3df774031dca5f09e9900a2eb4b74dda0651b29140b0c59a1"
},
"downloads": -1,
"filename": "torchrl_nightly-2024.4.2-cp39-cp39-manylinux1_x86_64.whl",
"has_sig": false,
"md5_digest": "85b41e3ae1467aef41d1d59780f72b47",
"packagetype": "bdist_wheel",
"python_version": "cp39",
"requires_python": null,
"size": 6244214,
"upload_time": "2024-04-02T11:26:45",
"upload_time_iso_8601": "2024-04-02T11:26:45.476734Z",
"url": "https://files.pythonhosted.org/packages/ec/9c/4685d1a48e54aa652a0b58559391c412d4cddba24bb40c77f47fdd6d406f/torchrl_nightly-2024.4.2-cp39-cp39-manylinux1_x86_64.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "df1f3898bc8ee8048b2feff1a4b209aa6692e0ee02f2e1585e6948b61f33a294",
"md5": "5b6ecb99cb547ed12a1a93ccbcf7e9ee",
"sha256": "8ac4909f8de9fce4a53a7740b9635b8ecf410c66466c36319f5f625f92bc0a47"
},
"downloads": -1,
"filename": "torchrl_nightly-2024.4.2-cp39-cp39-win_amd64.whl",
"has_sig": false,
"md5_digest": "5b6ecb99cb547ed12a1a93ccbcf7e9ee",
"packagetype": "bdist_wheel",
"python_version": "cp39",
"requires_python": null,
"size": 888213,
"upload_time": "2024-04-02T11:30:47",
"upload_time_iso_8601": "2024-04-02T11:30:47.667303Z",
"url": "https://files.pythonhosted.org/packages/df/1f/3898bc8ee8048b2feff1a4b209aa6692e0ee02f2e1585e6948b61f33a294/torchrl_nightly-2024.4.2-cp39-cp39-win_amd64.whl",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2024-04-02 11:23:58",
"github": true,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"github_user": "pytorch",
"github_project": "rl",
"travis_ci": false,
"coveralls": false,
"github_actions": true,
"lcname": "torchrl-nightly"
}
JSON
