appo-atari_timepilot-superhuman

About the Project

This project is an attempt to maximise performance of high sample throughput APPO RL models in Atari environments in as carbon efficient a manner as possible using a single, not particularly high performance single machine. It is about demonstrating the generalisability of on-policy algorithms to create good performance quickly (by sacrificing sample efficiency) while also proving that this route to RL production is accessible to even hobbyists like me (I am a gastroenterologist not a computer scientist).

In terms of throughput I am managing to reach throughputs of 2,500 - 3,000 across both policies using sample factory using two Quadro P2200's (not particularly powerful GPUs) each loaded up about 60% (3GB). Previously using the stable baselines 3 (sb3) implementation of PPO it would take about a week to train an atari agent to 100 million timesteps synchronously. By comparison the sample factory async implementation takes only just over 2 hours to achieve the same result. That is about 84 times faster with only typically a 21 watt burn per GPU. I am thus very grateful to Alex Petrenko and all the sample factory team for their work on this.

Project Aims

This model as with all the others in the benchmarks was trained initially asynchronously un-seeded to 10 million steps for the purposes of setting a sample factory async baseline for this model on this environment but only 3/57 made it anywhere near sota performance.

I then re-trained the models with 100 million timesteps- at this point 2 environments maxed out at sota performance (Pong and Freeway) with four approaching sota performance - (atlantis, boxing, tennis and fishingderby.) =6/57 near sota.

The aim now is to try and reach state-of-the-art (SOTA) performance on a further block of atari environments using up to 1 billion training timesteps initially with appo. I will flag the models with SOTA when they reach at or near these levels.

After this I will switch on V-Trace to see if the Impala variations perform any better with the same seed (I have seeded '1234')

About the Model

The hyperparameters used in the model are described in my shell script on my fork of sample-factory: https://github.com/MattStammers/sample-factory. Given that https://huggingface.co/edbeeching has kindly shared his parameters, I saved time and energy by using many of his tuned hyperparameters to reduce carbon inefficiency:

hyperparameters =  {
  "help": false,
  "algo": "APPO",
  "env": "atari_asteroid",
  "experiment": "atari_asteroid_APPO",
  "train_dir": "./train_atari",
  "restart_behavior": "restart",
  "device": "gpu",
  "seed": 1234,
  "num_policies": 2,
  "async_rl": true,
  "serial_mode": false,
  "batched_sampling": true,
  "num_batches_to_accumulate": 2,
  "worker_num_splits": 1,
  "policy_workers_per_policy": 1,
  "max_policy_lag": 1000,
  "num_workers": 16,
  "num_envs_per_worker": 2,
  "batch_size": 1024,
  "num_batches_per_epoch": 8,
  "num_epochs": 4,
  "rollout": 128,
  "recurrence": 1,
  "shuffle_minibatches": false,
  "gamma": 0.99,
  "reward_scale": 1.0,
  "reward_clip": 1000.0,
  "value_bootstrap": false,
  "normalize_returns": true,
  "exploration_loss_coeff": 0.0004677351413,
  "value_loss_coeff": 0.5,
  "kl_loss_coeff": 0.0,
  "exploration_loss": "entropy",
  "gae_lambda": 0.95,
  "ppo_clip_ratio": 0.1,
  "ppo_clip_value": 1.0,
  "with_vtrace": false,
  "vtrace_rho": 1.0,
  "vtrace_c": 1.0,
  "optimizer": "adam",
  "adam_eps": 1e-05,
  "adam_beta1": 0.9,
  "adam_beta2": 0.999,
  "max_grad_norm": 0.0,
  "learning_rate": 0.0003033891184,
  "lr_schedule": "linear_decay",
  "lr_schedule_kl_threshold": 0.008,
  "lr_adaptive_min": 1e-06,
  "lr_adaptive_max": 0.01,
  "obs_subtract_mean": 0.0,
  "obs_scale": 255.0,
  "normalize_input": true,
  "normalize_input_keys": [
    "obs"
  ],
  "decorrelate_experience_max_seconds": 0,
  "decorrelate_envs_on_one_worker": true,
  "actor_worker_gpus": [],
  "set_workers_cpu_affinity": true,
  "force_envs_single_thread": false,
  "default_niceness": 0,
  "log_to_file": true,
  "experiment_summaries_interval": 3,
  "flush_summaries_interval": 30,
  "stats_avg": 100,
  "summaries_use_frameskip": true,
  "heartbeat_interval": 10,
  "heartbeat_reporting_interval": 60,
  "train_for_env_steps": 100000000,
  "train_for_seconds": 10000000000,
  "save_every_sec": 120,
  "keep_checkpoints": 2,
  "load_checkpoint_kind": "latest",
  "save_milestones_sec": 1200,
  "save_best_every_sec": 5,
  "save_best_metric": "reward",
  "save_best_after": 100000,
  "benchmark": false,
  "encoder_mlp_layers": [
    512,
    512
  ],
  "encoder_conv_architecture": "convnet_atari",
  "encoder_conv_mlp_layers": [
    512
  ],
  "use_rnn": false,
  "rnn_size": 512,
  "rnn_type": "gru",
  "rnn_num_layers": 1,
  "decoder_mlp_layers": [],
  "nonlinearity": "relu",
  "policy_initialization": "orthogonal",
  "policy_init_gain": 1.0,
  "actor_critic_share_weights": true,
  "adaptive_stddev": false,
  "continuous_tanh_scale": 0.0,
  "initial_stddev": 1.0,
  "use_env_info_cache": false,
  "env_gpu_actions": false,
  "env_gpu_observations": true,
  "env_frameskip": 4,
  "env_framestack": 4,
  "pixel_format": "CHW"
}

A(n) APPO model trained on the atari_timepilot environment.

This model was trained using Sample-Factory 2.0: https://github.com/alex-petrenko/sample-factory. Sample factory is a high throughput on-policy RL framework. I have been using Documentation for how to use Sample-Factory can be found at https://www.samplefactory.dev/

Downloading the model

After installing Sample-Factory, download the model with:

python -m sample_factory.huggingface.load_from_hub -r MattStammers/APPO-atari_timepilot

Using the model

To run the model after download, use the enjoy script corresponding to this environment:

python -m sf_examples.atari.enjoy_atari --algo=APPO --env=atari_timepilot --train_dir=./train_dir --experiment=APPO-atari_timepilot

You can also upload models to the Hugging Face Hub using the same script with the --push_to_hub flag. See https://www.samplefactory.dev/10-huggingface/huggingface/ for more details

Training with this model

To continue training with this model, use the train script corresponding to this environment:

python -m sf_examples.atari.train_atari --algo=APPO --env=atari_timepilot --train_dir=./train_dir --experiment=APPO-atari_timepilot --restart_behavior=resume --train_for_env_steps=10000000000

Note, you may have to adjust --train_for_env_steps to a suitably high number as the experiment will resume at the number of steps it concluded at.