Skip to content

RobbenRibery/ued_llm_jax

 
 

Repository files navigation

Efficient baselines for autocurricula in JAX

Contents

🐢 Why minimax?

Unsupervised Environment Design (UED) is a promising approach to generating autocurricula for training robust deep reinforcement learning (RL) agents. However, existing implementations of common baselines require excessive amounts of compute. In some cases, experiments can require more than a week to complete using V100 GPUs. This long turn-around slows the rate of research progress in autocuriculum methods. minimax provides fast, JAX-based implementations of key UED baselines, which are based on the concept of minimax regret. By making use of fully-tensorized environment implementations, minimax baselines are fully-jittable and thus take full advantage of the hardware acceleration offered by JAX. In timing studies done on V100 GPUs and Xeon E5-2698 v4 CPUs, we find minimax baselines can run over 100x faster than previous reference implementations, like those in facebookresearch/dcd.

All autocurriculum algorithms implemented in minimax support multi-device training, which can be activated through a single command line flag. Using multiple devices for training can lead to further speed ups and allows scaling these autocurriculum methods to much larger batch sizes.

Shows Anuraghazra's GitHub Stats.

🐇 Hardware-accelerated baselines

minimax includes JAX-based implementations of

Additionally, minimax includes two new variants of PLR and ACCEL that further reduce wall time by better leveraging the massive degree of environment parallelism enabled by JAX:

  • Parallel PLR (PLR$^{||}$)

  • Parallel ACCEL (ACCEL$^{||}$)

In brief, these two new algorithms collect rollouts for new level evaluation, level replay, and, in the case of Parallel ACCEL, mutation evaluation, all in parallel (i.e. rather than sequentially, as done by Robust PLR and ACCEL). As a simple example for why this parallelization improves wall time, consider how Robust PLR with replay probability of 0.5 would require approximately 2x as many rollouts in order to reach the same number of RL updates as a method like DR, because updates are only performed on rollouts based on level replay. Parallelizing level replay rollouts alongside new level evaluation rollouts by using 2x the environment parallelism reduces the total number of parallel rollouts to equal the total number of updates desired, thereby matching the 1:1 rollout to update ratio of DR. The diagram below summarizes this difference.

Parallel DCD overview

minimax includes a fully-tensorized implementation of a maze environment that we call AMaze. This environment exactly reproduces the MiniGrid-based mazes used in previous UED studies in terms of dynamics, reward function, observation space, and action space, while running many orders of magnitude faster in wall time, with increasing environment parallelism.

🛠️ Install

  1. Use a virtual environment manager like conda or mamba to create a new environment for your project:
conda create -n minimax
conda activate minimax
  1. Install minimax via either pip install minimax-lib or pip install ued.

  2. That's it!

⚠️ Note that to enable hardware acceleration on GPU, you will need to make sure to install the latest version of jax>=0.4.19 and jaxlib>=0.4.19 that is compatible with your CUDA driver (requires minimum CUDA version of 11.8). See the official JAX installation guide for detailed instructions.

🏁 Quick start

The easiest way to get started is to play with the Python notebooks in the examples folder of this repository. We also host Colab versions of these notebooks:

*Depending on how the top-level flags are set, this notebook runs PLR, Robust PLR, Parallel PLR, ACCEL, or Parallel ACCEL.

minimax comes with high-performing hyperparameter configurations for several algorithms, including domain randomization (DR), PAIRED, PLR, and ACCEL for 60-block mazes. You can train using these settings by first creating the training command for executing minimax.train using the convenience script minimax.config.make_cmd:

python -m minimax.config.make_cmd --config maze/[dr,paired,plr,accel] | pbcopy,

followed by pasting and executing the resulting command into the command line.

See the docs for minimax.config.make_cmd to learn more about how to use this script to generate training commands from JSON configurations. You can browse the available JSON configurations for various autocurriculum methods in the configs folder.

Note that when logging and checkpointing are enabled, the main minimax.train script outputs this data as logs.csv and checkpoint.pkl respectively in an experiment directory located at <log_dir>/<xpid>, where log_dir and xpid are arguments specified in the command. You can then evaluate the checkpoint by using minimax.evaluate:

python -m minimax.evaluate \
--seed 1 \
--log_dir <absolute path log directory> \
--xpid_prefix <select checkpoints with xpids matching this prefix> \
--env_names <csv string of test environment names> \
--n_episodes <number of trials per test environment> \
--results_path <path to results folder> \
--results_fname <filename of output results csv>

🪸 Dive deeper

minimax system diagram

Training

The main entry for training is minimax.train. This script configures the training run based on command line arguments. It constructs an instance of ExperimentRunner to manage the training process on an update-cycle basis: These duties include constructing and delegating updates to an appropriate training runner for the specified autocurriculum algorithm and conducting logging and checkpointing. The training runner used by ExperimentRunner executes all autocurriculum-related logic. The system diagram above describes how these pieces fit together, as well as how minimax manages various, hierarchical batch dimensions.

Currently, minimax includes training runners for the following classes of autocurricula:

Runner Algorithm class --train_runner
DRRunner Domain randomization dr
PLRRunner Replay-based curricula, including ACCEL plr
PAIREDRunner Curricula via a co-adapting teacher environment design policy paired

The below table summarizes how various autocurriculum methods map to these runners and the key arguments that must be set differently from the default settings in order to switch the runner's behavior to each method.

Algorithm Reference Runner Key args
DR Tobin et al, 2019 DRRunner
Minimax adversary Dennis et al, 2020 PAIREDRunner ued_score='neg_return'
PAIRED Dennis et al, 2020 PAIREDRunner
Population PAIRED Dennis et al, 2020 PAIREDRunner n_students >= 2, ued_score='population_regret'
PLR Jiang et al, 2021 PLRRunner plr_use_robust_plr=False
Robust PLR Jiang et al, 2021a PLRRunner
ACCEL Parker-Holder et al, 2022 PLRRunner plr_mutation_fn != None, plr_n_mutations > 0
Parallel PLR Jiang et al, 2023 PLRRunner plr_use_parallel_eval=True
Parallel ACCEL Jiang et al, 2023 PLRRunner plr_use_parallel_eval=True, plr_mutation_fn != None, plr_n_mutations > 0

See the docs on minimax.train for a comprehensive guide on how to configure command-line arguments for running various autocurricula methods via minimax.train.

Logging

By default, minimax.train generates a folder in the directory specified by the --log_dir argument, named according to --xpid. This folder contains the main training logs, logs.csv, which are updated with a new row every --log_interval rollout cycles.

Checkpointing

Latest checkpoint: The latest model checkpoint is saved as checkpoint.pkl. The model is checkpointed every --checkpoint_interval number of updates, where each update corresponds to a full rollout and update cycle for each participating agent. For the same number of environment interaction steps, methods may differ in the number of gradient updates performed by participating agents, so checkpointing based on number of update cycles controls for this potential discrepency. For example, methods based on Robust PLR, like ACCEL, do not perform student gradient updates every rollout cycle.

Archived checkpoints: Separate archived model checkpoints can be saved at specific intervals by specifying a positive value for the argument --archive_interval. For example, setting --archive_interval=1000 will result in saving model checkpoints every 1000 updates, named checkpoint_1000.tar, checkpoint_2000.tar, and so on. These archived models are saved in addition to checkpoint.pkl, which always stores the latest checkpoint, based on --checkpoint_interval.

Evaluating

Once training completes, you can evaluate the resulting checkpoint.pkl on test environments using minimax.evaluate. This script can evaluate an individual checkpoint or group of checkpoints created via training runs with a shared experiment ID prefix (--xpid value), e.g. each corresponding to different training seeds of the same experiment configuration. Each checkpoint is evaluated over --n_episodes episodes for each of the test environments, specified via a csv string of test environment names passed in via --env_names. The evaluation results can be optionally written to a csv file in --results_path, if a --results_fname is provided.

See the docs on minimax.evaluation for a comprehensive guide on how to configure command line arguments for minimax.evaluate.

Multi-device training

All autocurriculum algorithms in minimax support multi-device training via shmap across the environment batch dimension (see the system diagram above). In order to shard rollouts and gradient updates along the environment batch dimension across N devices, simply pass minimax.train the additional argument --n_devices=N. By default, n_devices=1.

🏝️ Environments

Supported environments

Maze Overview

minimax currently includes AMaze, a fully-tensorized implementation of the partially-observable maze navigation environments featured in previous UED studies (see example AMaze environments in the figure above). The minimax implementation of the maze environment fully replicates the original MiniGrid-based dynamics, reward functions, observation space, action space. See the environment docs fo more details.

We look forward to working with the greater RL community in continually expanding the set of environments integrated with minimax.

Adding environments

In order to integrate into minimax's fully-jittable training logic, environments should be implemented in a tensorized fashion via JAX. All environments must implement the Environment interface. At a high level, Environment subclasses should implement reset and step logic assuming a single environment instance (no environment parallelism). Parallelism is automatically achieved via the training runner logic included withminimax (See the paper and system diagram above for a quick overview of how this is performed).

A key design decision of minimax is to separate environment parameters into two groups:

  • Static parameters are fixed throughout training. These parameters are frozen hyperparameters defining some unchanging aspect of the underlying environment distribution, e.g. the width, height, or maximum number of walls of maze environments considered during training. These static parameters are encapsulated in an EnvParams dataclass.

  • Free parameters can change per environment instance (e.g. across each instance in a parallel rollout batch). These parameters might correspond to aspects like the specific wall map defining the maze layout or the starting position of the agent. Free parameters are simply treated as part of the fully-traceable EnvState, taking the form of an arbitrary pytree.

All environments supporting the Environment interface will interoperate with DRRunner and PLRRunner (though for ACCEL mode, where mutation_fn != None, a mutation operator must additionally be defined).

Environment design with a co-adapting teacher policy

In PAIRED-style autocurricula, a teacher policy generates environment instances in order to maximize some curriculum objective, e.g. relative regret. The teacher's decision-making process corresponds to its own MDP.

To support such autocurricula, minimax follows the pattern of implementing separate Environment subclasses for each of student and teacher MDPs. A convenience class called UEDEnvironment is then initialized with instances of the student and teacher MDPs. The UEDEnvironment instance exposes a unified interface for resetting and stepping the teacher and student, which is then used in the training runner. For example, stepping the UEDEnvironment instance for the teacher (via the step_teacher method) produces an environment instance, which can then be used with the reset_student method to reset the state of the UEDEnvironment object to that particular environment instance. Subsequent calls of the step_student method then operate within this environment instance. Following this pattern, integration of a new environment with PAIREDRunner requires implementing the corresponding Environment subclass for the teacher MDP (the decision process ). See minimax/envs/maze/maze_ued.py for an example based on the maze environment.

Environment operators

Custom environment operators can also be defined in minimax.

  • Comparators take two environment instances, as represented by their EnvState pytrees and return True iff the two instances are deemed equal. If a comparator is registered for an environment, training runners can use the comparator to enforce uniqueness of environment instances for many purposes, e.g. making sure the members of the PLR buffer are all unique.

  • Mutators take an environment instance, as represented by an EnvState pytree, and apply some modification to the instance, returning the modified (or "mutated") instance. Mutators are used by ACCEL to mutate environment instances in the PLR buffer. New environments seeking integration with the ACCEL mode of the PLRRunner should implement and register a default mutation operator.

Registration

Each new Environment subclass should be registered with the envs module:

  • Student environments should be registered using envs.registration.register. See src/minimax/maze/maze.py for an example.

  • Teacher environments should be registered using envs.registration.register_ued. See envs/maze/maze_ued.py for an example.

  • Mutators should be registered using envs.registration.register_mutator. See envs/maze/maze_mutators.py for an example.

  • Comparators should be registered using envs.registration.register_comparator. See envs/maze/maze_comparators.py for an example.

🤖 Agents

In minimax agents correspond to a particular data-seeking learning algorithm, e.g. PPO. A model corresponds to a module that implements the policy (or value function) used by the agent. Any agent that follows the Agent interface should be usable in any minimax compatible environment.

Model forward passes are assumed to return a tuple of (value_prediction, policy_logits, carry).

Registration

Custom model classes should be registered for a particular environment for which they are designed. See models/maze/gridworld_models.py for an example. After registration, the model can be easily retrieved and via models.make(env_name, model_name, **model_kwargs).

🚀 Roadmap

Many exciting features are planned for future releases of minimax. Features planned for near-term release include:

  • Add support for JaxMARL (multi-agent RL environments) via an IPPO mode in DRRunner and PLRRunner.
  • Extend Parsnip with methods for composing argument specs across multiple files to reduce the size of arguments.py currently used for train.py.
  • Add support for Jumanji (combinatorial optimization environments), via an appropriate decorator class.

You can suggest new features or ways to improve current functionality by creating an issue in this repository.

🪪 License

minimax is licensed under Apache 2.0.

📜 Citation

For attribution in academic contexts, please cite this work as

@article{jiang2023minimax,
    title={minimax: Efficient Baselines for Autocurricula in JAX},
    author={Jiang, Minqi and Dennis, Michael and Grefenstette, Edward and Rocktäschel, Tim},
    booktitle={Agent Learning in Open-Endedness Workshop at NeurIPS},
    year={2023}
}

About

Efficient baselines for autocurricula in JAX.

Resources

License

Code of conduct

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%