TorchRL

Disclaimer

This library is not officially released yet and is subject to change.

The features are available before an official release so that users and collaborators can get early access and provide feedback. No guarantee of stability, robustness or backward compatibility is provided.

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), transforms, models, 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, returns and data processing.

TorchRL aims at (1) a high modularity and (2) good runtime performance.

Features

On the high-level end, TorchRL provides:

TensorDict, a convenient data structure⁽¹⁾ to pass data from one object to another without friction. TensorDict makes it easy to re-use pieces of code across environments, models and algorithms. For instance, here's how to code a rollout in TorchRL:

Code

- obs, done = env.reset()
+ tensordict = env.reset()
policy = TensorDictModule(
    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_tensordict(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

TensorDict abstracts away the input / output signatures of the modules, env, collectors, replay buffers and losses of the library, allowing its primitives to be easily recycled across settings. 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):

- 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()

Again, this training loop can be re-used across algorithms as it makes a minimal number of assumptions about the structure of the data.

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):

# 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_()

Check our TensorDict tutorial for more information.

An associated TensorDictModule class which is functorch-compatible!

Code

transformer_model = nn.Transformer(nhead=16, num_encoder_layers=12)
+ td_module = TensorDictModule(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 TensorDictSequence 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:

encoder_module = TransformerEncoder(...)
encoder = TensorDictModule(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 = TensorDictSequence(encoder, decoder)
assert transformer.in_keys == ["src", "src_mask", "tgt"]
assert transformer.out_keys == ["memory", "output"]

TensorDictSequence allows to isolate subgraphs by querying a set of desired input / output keys:

transformer.select_subsequence(out_keys=["memory"])  # returns the encoder
transformer.select_subsequence(in_keys=["tgt", "memory"])  # returns the decoder

The corresponding tutorial provides more context about its features.

a generic trainer class⁽¹⁾ that executes the aforementioned training loop. Through a hooking mechanism, it also supports any logging or data transformation operation at any given time.

A common interface for environments which supports common libraries (OpenAI gym, deepmind control lab, etc.)⁽¹⁾ and state-less execution (e.g. Model-based environments). The batched environments containers allow parallel execution⁽²⁾. A common pytorch-first class of tensor-specification class is also provided.

Code

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

multiprocess data collectors⁽²⁾ 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):

Code

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_()

efficient⁽²⁾ and generic⁽¹⁾ replay buffers with modularized storage:

Code

storage = LazyMemmapStorage(  # memory-mapped (physical) storage
    cfg.buffer_size,
    scratch_dir="/tmp/"
)
buffer = TensorDictPrioritizedReplayBuffer(
    buffer_size=10000,
    alpha=0.7,
    beta=0.5,
    collate_fn=lambda x: x,
    pin_memory=device != torch.device("cpu"),
    prefetch=10,  # multi-threaded sampling
    storage=storage
)

cross-library environment transforms⁽¹⁾, executed on device and in a vectorized fashion⁽²⁾, which process and prepare the data coming out of the environments to be used by the agent:

Code

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.), contatenation 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:

env.insert_transform(0, NoopResetEnv())  # inserts the NoopResetEnv transform at the index 0

Nevertheless, transforms can access and execute operations on the parent environment:

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

various tools for distributed learning (e.g. memory mapped tensors)⁽²⁾;

various architectures and models (e.g. actor-critic)⁽¹⁾:

Code

# 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 TensorDictModule, indicating what key to read in and where to
# write out the output
common_module = TensorDictModule(
    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 = NormalParamsWrapper(
    MLP(
        num_cells=[64, 64],
        out_features=32,
        activation=nn.ELU,
    )
)
# Wrap the nn.Module in a ProbabilisticTensorDictModule, indicating how
# to build the torch.distribution.Distribution object and what to do with it
policy_module = ProbabilisticTensorDictModule(  # stochastic policy
    TensorDictModule(
        policy_module,
        in_keys=["hidden"],
        out_keys=["loc", "scale"],
    ),
    dist_param_keys=["loc", "scale"],
    out_key_sample="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()

exploration wrappers and modules to easily swap between exploration and exploitation⁽¹⁾:

Code

policy_explore = EGreedyWrapper(policy)
with set_exploration_mode("random"):
    tensordict = policy_explore(tensordict)  # will use eps-greedy
with set_exploration_mode("mode"):
    tensordict = policy_explore(tensordict)  # will not use eps-greedy

A series of efficient loss modules and highly vectorized functional return and advantage computation.

Code

Loss modules

from torchrl.objectives.costs import DQNLoss
loss_module = DQNLoss(value_network=value_network, gamma=0.99)
tensordict = replay_buffer.sample(batch_size)
loss = loss_module(tensordict)

Advantage computation

from torchrl.objectives.returns.functional import vec_td_lambda_return_estimate
advantage = vec_td_lambda_return_estimate(gamma, lmbda, next_state_value, reward, done)

various recipes to build models that correspond to the environment being deployed.

Examples, tutorials and demos

A series of examples are provided with an illustrative purpose:

and many more to come!

We also provide tutorials and demos that give a sense of what the library can do.

Installation

Create a conda environment where the packages will be installed.

conda create --name torch_rl python=3.9
conda activate torch_rl

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:

Stable

# For CUDA 10.2
conda install pytorch torchvision cudatoolkit=10.2 -c pytorch
# For CUDA 11.3
conda install pytorch torchvision cudatoolkit=11.3 -c pytorch
# For CPU-only build
conda install pytorch torchvision cpuonly -c pytorch

pip3 install functorch

Nightly

# For CUDA 10.2
pip3 install --pre torch torchvision --extra-index-url https://download.pytorch.org/whl/nightly/cu102
# For CUDA 11.3
pip3 install --pre torch torchvision --extra-index-url https://download.pytorch.org/whl/nightly/cu113
# For CPU-only build
pip3 install --pre torch torchvision --extra-index-url https://download.pytorch.org/whl/nightly/cpu

and functorch

pip3 install ninja  # Makes the build go faster
pip3 install "git+https://github.com/pytorch/functorch.git"

If this fails, you can get the latest version of functorch that was marked to be compatible with the current torch version:

pip3 install ninja  # Makes the build go faster
PYTORCH_VERSION=`python -c "import torch.version; print(torch.version.git_version)"`
pip3 install "git+https://github.com/pytorch/pytorch.git@$PYTORCH_VERSION#subdirectory=functorch"

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, try erasing the previously created build artifacts (torchrl.egg-info/, build/, torchrl/_torchsl.so) or re-clone the library from GitHub, then try

PYTORCH_VERSION=`python -c "import torch.version; print(torch.version.git_version)"`
ARCHFLAGS="-arch arm64" pip3 install "git+https://github.com/pytorch/pytorch.git@$PYTORCH_VERSION#subdirectory=functorch"

Torchrl

You can install the latest release by using

pip3 install torchrl

This should work on linux and MacOs (not M1). For Windows and M1/M2 machines, one should install the library locally (see below).

To install extra dependencies, call

pip3 install "torchrl[atari,dm_control,gym_continuous,rendering,tests,utils]"

or a subset of these.

Alternatively, as the library is at an early stage, it may be wise to install it in develop mode as this will make it possible to pull the latest changes and benefit from them immediately. Start by cloning the repo:

git clone https://github.com/facebookresearch/rl

Go to the directory where you have cloned the torchrl repo and install it

cd /path/to/torchrl/
python setup.py develop

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 gym_retro

# 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, 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. Indeed 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.

On MacOs, we recommend installing XCode first. With Apple Silicon M1 chips, make sure you are using the arm64-built python (e.g. here). 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)

Running examples

Examples are coded in a very similar way but the configuration may change from one algorithm to another (e.g. async/sync data collection, hyperparameters, ratio of model updates / frame etc.) To train an algorithm it is therefore advised to use the predefined configurations that are found in the configs sub-folder in each algorithm directory:

python examples/ppo/ppo.py --config=examples/ppo/configs/humanoid.txt

Note that using the config files requires the configargparse library.

One can also overwrite the config parameters using flags, e.g.

python examples/ppo/ppo.py --config=examples/ppo/configs/humanoid.txt --frame_skip=2 --collection_devices=cuda:1

Each example will write a tensorboard log in a dedicated folder, e.g. ppo_logging/....

Contributing

Internal collaborations to torchrl are welcome! Feel free to fork, submit issues and PRs. You can checkout the detailed contribution guide here.

Contributors are recommended to install pre-commit hooks (using pre-commit install). pre-commit will check for linting related issues when the code is commited locally. You can disable th check by appending -n to your commit command: git commit -m <commit message> -n

Upcoming features

In the near future, we plan to:

provide tutorials on how to design new actors or environment wrappers;
implement IMPALA (as a distributed RL example) and Meta-RL algorithms;
improve the tests, documentation and nomenclature.

We welcome any contribution, should you want to contribute to these new features or any other, lister or not, in the issues section of this repository.

License

TorchRL is licensed under the MIT License. See LICENSE for details.

ronert/rl