The torch_rl package actually contains the PyTorch implementation of two Actor-Critic deep reinforcement learning algorithms:

• Synchronous A3C (A2C)
• Proximal Policy Optimizations (PPO)

The package comes out of the box with starter files:

• generic RL scripts (in scripts) to train, visualize and evaluate an agent
• a default agent's model (in model.py)
• CSV and Tensorboard logging for easy debugging

These files use the torch_rl package to allow you to immediatly train an agent on MiniGrid environments without having to write any line of code and they can be easily adapted to other environments.

• Support:
  • Recurrent policies
  • Reward shaping
  • Wide variety of observation spaces: tensors or dict of tensors
  • Wide variety of action spaces: discrete or continuous
  • Observation preprocessing
• Fast:
  • Multiprocess: parallelising experience collection
  • CUDA

You have to clone the repository and then install the package:

git clone https://github.com/lcswillems/torch-rl.git
cd torch_rl
pip3 install -e torch_rl

To update the package, just do git pull. No need to install it again.

The starter files presented just below are best suited for MiniGrid environments that you can install with:

git clone https://github.com/maximecb/gym-minigrid.git
cd gym-minigrid
pip3 install -e .

But the starter files can be easily adapted to other environments.

The torch_rl package comes out of the box with starter files:

• generic RL scripts:
  • scripts/train.py to train an agent
  • scripts/visualize.py to visualize how an agent behaves
  • scripts/evaluate.py to evaluate agent's performances
• a default agent's model: model.py (more details below)
• utilitarian classes and functions (in utils) used by the scripts

To adapt these files to your needs, you may want to modify:

• model.py
• utils/format.py

An example of use:

python3 -m scripts.train --algo ppo --env MiniGrid-DoorKey-5x5-v0 --model DoorKey --save-interval 10

In this use case, the script loads the model in storage/DoorKey or creates it if it doesn't exist, then trains it with the PPO algorithm on the MiniGrid DoorKey environment, and saves it every 10 updates in the storage/DoorKey directory.

Note: You can define a different storage location in the environment variable TORCH_RL_STORAGE.

More generally, the script has 2 required arguments:

• --algo ALGO: name of the RL algorithm used to train
• --env ENV: name of the environment to train on

and a bunch of optional arguments among which:

• --recurrence N: gradient will be backpropagated over N timesteps. By default, N = 1. If N > 1, a LSTM is added to the model to have memory.
• --text: a GRU is added to the model to handle text input.
• ... (see more using --help)

During training, logs are printed in your terminal (and saved in text and CSV format):

Note: U gives the update number, F the total number of frames, FPS the number of frames per second, D the total duration, rR:μσmM the mean, std, min and max reshaped return per episode, F:μσmM the mean, std, min and max number of frames per episode, H the entropy, V the value, pL the policy loss, vL the value loss and ∇ the gradient norm.

During training, logs might also be plotted in Tensorboard if --tb is added.

Note: tensorboardX package is required and can be installed with pip3 install tensorboardX.

An example of use:

python3 -m scripts.visualize --env MiniGrid-DoorKey-5x5-v0 --model DoorKey

In this use case, the script displays how the model in storage/DoorKey behaves on the MiniGrid DoorKey environment.

More generally, the script has 2 required arguments:

• --env ENV: name of the environment to act on.
• --model MODEL: name of the trained model.

and a bunch of optional arguments among which:

• --argmax: select the action with highest probability
• ... (see more using --help)

An example of use:

python3 -m scripts.evaluate --env MiniGrid-DoorKey-5x5-v0 --model DoorKey

In this use case, the script prints in the terminal the performance among 100 episodes of the model in storage/DoorKey.

More generally, the script has 2 required arguments:

• --env ENV: name of the environment to act on.
• --model MODEL: name of the trained model.

and a bunch of optional arguments among which:

• --episodes N: number of episodes of evaluation. By default, N = 100.
• ... (see more using --help)

The default model is discribed by the following schema:

By default, the memory part (in red) and the langage part (in blue) are disabled. They can be enabled by setting to True the use_memory and use_text parameters of the model constructor.

This model can be easily adapted to your needs.

The package consists of:

• torch_rl.A2CAlgo and torch_rl.PPOAlgo classes for A2C and PPO algorithms
• torch_rl.ACModel and torch_rl.RecurrentACModel abstract classes for non-recurrent and recurrent actor-critic models
• a torch_rl.DictList class for making dictionnaries of lists list-indexable and hence batch-friendly

Here are detailed the most important parts of the package.

torch_rl.A2CAlgo and torch_rl.PPOAlgo have 2 methods:

• __init__ that may take, among the other parameters:
  • an acmodel actor-critic model, i.e. an instance of a class inheriting from either torch_rl.ACModel or torch_rl.RecurrentACModel.
  • a preprocess_obss function that transforms a list of observations into a list-indexable object X (e.g. a PyTorch tensor). The default preprocess_obss function converts observations into a PyTorch tensor.
  • a reshape_reward function that takes into parameter an observation obs, the action action taken, the reward reward received and the terminal status done and returns a new reward. By default, the reward is not reshaped.
  • a recurrence number to specify over how many timesteps gradient is backpropagated. This number is only taken into account if a recurrent model is used and must divide the num_frames_per_agent parameter and, for PPO, the batch_size parameter.
• update_parameters that first collects experiences, then update the parameters and finally returns logs.

torch_rl.ACModel has 2 abstract methods:

• __init__ that takes into parameter an observation_space and an action_space.
• forward that takes into parameter N preprocessed observations obs and returns a PyTorch distribution dist and a tensor of values value. The tensor of values must be of size N, not N x 1.

torch_rl.RecurrentACModel has 3 abstract methods:

• __init__ that takes into parameter the same parameters than torch_rl.ACModel.
• forward that takes into parameter the same parameters than torch_rl.ACModel along with a tensor of N memories memory of size N x M where M is the size of a memory. It returns the same thing than torch_rl.ACModel plus a tensor of N memories memory.
• memory_size that returns the size M of a memory.

Note: The preprocess_obss function must return a list-indexable object (e.g. a PyTorch tensor). If your observations are dictionnaries, your preprocess_obss function may first convert a list of dictionnaries into a dictionnary of lists and then make it list-indexable using the torch_rl.DictList class as follow:

>>> d = DictList({"a": [[1, 2], [3, 4]], "b": [[5], [6]]})
>>> d.a
[[1, 2], [3, 4]]
>>> d[0]
DictList({"a": [1, 2], "b": [5]})

Note: if you use a RNN, you will need to set batch_first to True.

An example of use of torch_rl.A2CAlgo and torch_rl.PPOAlgo classes is given in scripts/train.py.

An example of use of torch_rl.DictList is given in the preprocess_obss functions of utils/format.py.

An example of implementation of torch_rl.RecurrentACModel abstract class is defined in model.py

Examples of preprocess_obss functions are given in utils/format.py.