Skip to content

Latest commit

 

History

History
244 lines (188 loc) · 7.86 KB

training.md

File metadata and controls

244 lines (188 loc) · 7.86 KB
Raw

Benchmarking Training

Prerequisites

Before you begin, ensure you have done the following:

Install A2Perf

For detailed instructions, please refer to our Installation Guide.

Proper Submission folder structure

Your submission should be placed in the A2Perf/a2perf/a2perf_benchmark_submission/ directory. The structure should look like this:

A2Perf/
├── a2perf/
│   ├── a2perf_benchmark_submission/
│   │   ├── init.py
│   │   ├── train.py
│   │   ├── inference.py
│   │   ├── requirements.txt
│   │   └── [your supporting files and directories]

To get started quickly, you can use our template repository as a starting point. You can find the template at:

https://github.com/Farama-Foundation/a2perf-benchmark-submission/tree/template

This template provides the basic structure and files needed for your submission. You can clone this repository and modify it to fit your specific implementation.

Explanation of files:

  • __init__.py: The __init__.py file can be an empty file.

  • train.py: The train.py file includes the function train(), which A2Perf calls for the training of your algorithm.

  • inference.py Next, the inference.py file is subsequently used for benchmarking the trained agent. This file includes several key functions.

    load_model(env): This function loads and returns the trained agent. A2Perf passes the environment that is being tested via the env parameter. This allows the agent loading logic to use any context needed, such as the environment object.
    preprocess_observation(observation): Preprocesses the observation before feeding it to the agent. If no preprocessing is required, simply return the initial observation.
    infer_once(model, observation): Passes a single observation to the loaded agent and returns the predicted action. This function performs a single inference step.

  • requirements.txt: Specifies any package dependencies required to run the submission code. This file may include version constraints for the dependencies. Having an explicit requirements file ensures a consistent environment for evaluation.

Update the a2perf_benchmark_submission Submodule

For the purposes of this tutorial, we will use the repository of our local baselines as the "submission." To use this, please update the a2perf_benchmark_submission submodule to the following branch: baselines-local

Navigate to the Submodule Directory

cd a2perf/a2perf_benchmark_submission

Checkout the branch with code for baselines

git fetch origin
git checkout baselines-local

Pull Latest Changes

git pull origin baselines-local

Back to the Main Directory

Return to the root directory of the A2Perf repository:

cd ../..

Running the Training Benchmark

Running locally with XManager (Docker)

Running the Benchmark

xmanager launch xm_launch.py -- \
  --experiment-name=test \
  --root-dir=~/gcs/a2perf/experiments/ \
  --domain=QuadrupedLocomotion-DogPace-v0  \
  --submission-gin-config-path=a2perf/submission/configs/quadruped_locomotion/train.gin \
  --user=$USER \
  --participant-module-path=a2perf/a2perf_benchmark_submission \
  --participant-args="gin_config_path=configs/quadruped_locomotion/dog_pace/ppo.gin"

Command line arguments

  • root-dir: Specifies the directory where experiment logs and artifacts will be saved.
  • gin-config: Points to the Gin configuration file for the Dog Pace environment in the Quadruped Locomotion domain.
  • participant-module-path: Indicates the path to the directory containing the submission code. Adjust this path to point to your a2perf_benchmark_submission directory.
  • participant-args: Provides additional arguments for the participant's code, including the path to the algorithm-specific Gin configuration file.

XManager will automatically launch a Docker container with the necessary dependencies installed. It will also create a new experiment directory at ~/gcs/a2perf/experiments/<experiment-id>/test/1/. The number 1 is included because we are running a single work unit in the experiment. For more details on work units, refer to XManager's documentation.

Important: Make note of the <experiment-id> in your experiment directory path. You will need this ID when running the inference benchmark later. The <experiment-id> is a unique identifier for your training run and is typically a long string of numbers.

The experiment directory will contain all logs and artifacts generated during the benchmark. Here is how the directory structure will look at the end of training:

~/gcs/a2perf/experiments/1724700456099
└── test
    └── 1
        ├── collect
        ├── metrics
        ├── policies
        ├── submission_config.gin
        ├── training_complete
        └── train

  • collect/: Contains TensorBoard summaries for each of the separate collection policies. Each subdirectory (e.g., actor_0, actor_1, ...) corresponds to a different actor's summary data.

    Example structure:

    collect/
├── actor_0
│   └── summaries
│       └── 0
├── actor_1
│   └── summaries
│       └── 1
└── ...

  • metrics/: Stores system metrics collected during training, such as train_emissions.csv, which logs energy consumption and emissions data.

  • policies/: Contains the saved policies generated during training. This includes different policies such as collect_policy, greedy_policy, and the main policy. Each subdirectory represents a different policy and contains the necessary files for TensorFlow models, including saved models and variables.

    Example structure:

    policies/
├── collect_policy
├── greedy_policy
└── policy

  • train/: Contains additional checkpoint information and TensorBoard logs from the training process, which are useful for monitoring training progress and debugging.

Running Locally Without Docker

If you prefer to run the benchmark locally without using Docker, follow these steps:

Installing Dependencies

To run the benchmark locally, you need to manually install the required Python dependencies. Run the following command:

pip install -r A2Perf/a2perf/a2perf_benchmark_submission/requirements.txt

Running the Benchmark

Once the dependencies are installed, you can run the benchmark for the Quadruped Locomotion - Dog Pace environment with the following command:

cd A2perf
export A2PERF_ROOT=$(pwd)
a2perf $A2PERF_ROOT/a2perf/a2perf_benchmark_submission \
  --root-dir=~/gcs/a2perf/experiments/test_without_docker \
  --submission-gin-config-path=$A2PERF_ROOT/a2perf/submission/configs/quadruped_locomotion/train.gin \
  --participant-args="gin_config_path=configs/quadruped_locomotion/dog_pace/ppo.gin"

Command line arguments

  • root-dir: Specifies the directory where experiment logs and artifacts will be saved.
  • gin-config: Points to the Gin configuration file for the Dog Pace environment in the Quadruped Locomotion domain.
  • participant-module-path: Indicates the path to the directory containing the submission code. Adjust this path to point to your a2perf_benchmark_submission directory.
  • participant-args: Provides additional arguments for the participant's code, including the path to the algorithm-specific Gin configuration file.

Make sure to adjust the paths according to your setup if they differ from the example provided.