sgoodfriend's picture
PPO playing BreakoutNoFrameskip-v4 from https://github.com/sgoodfriend/rl-algo-impls/tree/e47a44c4d891f48885af0b1605b30d19fc67b5af
81552e0
metadata
library_name: rl-algo-impls
tags:
  - BreakoutNoFrameskip-v4
  - ppo
  - deep-reinforcement-learning
  - reinforcement-learning
model-index:
  - name: ppo
    results:
      - metrics:
          - type: mean_reward
            value: 516.88 +/- 155.01
            name: mean_reward
        task:
          type: reinforcement-learning
          name: reinforcement-learning
        dataset:
          name: BreakoutNoFrameskip-v4
          type: BreakoutNoFrameskip-v4

PPO Agent playing BreakoutNoFrameskip-v4

This is a trained model of a PPO agent playing BreakoutNoFrameskip-v4 using the /sgoodfriend/rl-algo-impls repo.

All models trained at this commit can be found at https://api.wandb.ai/links/sgoodfriend/v4wd7cp5.

Training Results

This model was trained from 3 trainings of PPO agents using different initial seeds. These agents were trained by checking out e47a44c. The best and last models were kept from each training. This submission has loaded the best models from each training, reevaluates them, and selects the best model from these latest evaluations (mean - std).

algo env seed reward_mean reward_std eval_episodes best wandb_url
ppo BreakoutNoFrameskip-v4 1 502.562 161.406 16 wandb
ppo BreakoutNoFrameskip-v4 2 426.562 85.8509 16 wandb
ppo BreakoutNoFrameskip-v4 3 516.875 155.012 16 * wandb

Prerequisites: Weights & Biases (WandB)

Training and benchmarking assumes you have a Weights & Biases project to upload runs to. By default training goes to a rl-algo-impls project while benchmarks go to rl-algo-impls-benchmarks. During training and benchmarking runs, videos of the best models and the model weights are uploaded to WandB.

Before doing anything below, you'll need to create a wandb account and run wandb login.

Usage

/sgoodfriend/rl-algo-impls: https://github.com/sgoodfriend/rl-algo-impls

Note: While the model state dictionary and hyperaparameters are saved, the latest implementation could be sufficiently different to not be able to reproduce similar results. You might need to checkout the commit the agent was trained on: e47a44c.

# Downloads the model, sets hyperparameters, and runs agent for 3 episodes
python enjoy.py --wandb-run-path=sgoodfriend/rl-algo-impls-benchmarks/kxf84p5u

Setup hasn't been completely worked out yet, so you might be best served by using Google Colab starting from the colab_enjoy.ipynb notebook.

Training

If you want the highest chance to reproduce these results, you'll want to checkout the commit the agent was trained on: e47a44c. While training is deterministic, different hardware will give different results.

python train.py --algo ppo --env BreakoutNoFrameskip-v4 --seed 3

Setup hasn't been completely worked out yet, so you might be best served by using Google Colab starting from the colab_train.ipynb notebook.

Benchmarking (with Lambda Labs instance)

This and other models from https://api.wandb.ai/links/sgoodfriend/v4wd7cp5 were generated by running a script on a Lambda Labs instance. In a Lambda Labs instance terminal:

git clone git@github.com:sgoodfriend/rl-algo-impls.git
cd rl-algo-impls
bash ./lambda_labs/setup.sh
wandb login
bash ./lambda_labs/benchmark.sh

Alternative: Google Colab Pro+

As an alternative, colab_benchmark.ipynb, can be used. However, this requires a Google Colab Pro+ subscription and running across 4 separate instances because otherwise running all jobs will exceed the 24-hour limit.

Hyperparameters

This isn't exactly the format of hyperparams in hyperparams/ppo.yml, but instead the Wandb Run Config. However, it's very close and has some additional data:

algo: ppo
algo_hyperparams:
  batch_size: 256
  clip_range: 0.1
  clip_range_decay: linear
  ent_coef: 0.01
  learning_rate: 0.00025
  learning_rate_decay: linear
  n_epochs: 4
  n_steps: 128
  vf_coef: 0.5
env: impala-BreakoutNoFrameskip-v4
env_hyperparams:
  frame_stack: 4
  n_envs: 8
  no_reward_fire_steps: 500
  no_reward_timeout_steps: 1000
  vec_env_class: subproc
env_id: BreakoutNoFrameskip-v4
eval_params:
  deterministic: false
n_timesteps: 10000000
policy_hyperparams:
  activation_fn: relu
  cnn_feature_dim: 256
  cnn_layers_init_orthogonal: false
  cnn_style: impala
  init_layers_orthogonal: true
seed: 3
use_deterministic_algorithms: true
wandb_entity: null
wandb_project_name: rl-algo-impls-benchmarks
wandb_tags:
- benchmark_e47a44c
- host_129-146-2-230