library_name: sample-factory
tags:
- deep-reinforcement-learning
- reinforcement-learning
- sample-factory
model-index:
- name: APPO
results:
- task:
type: reinforcement-learning
name: reinforcement-learning
dataset:
name: atari_tennis
type: atari_tennis
metrics:
- type: mean_reward
value: 23.50 +/- 0.50
name: mean_reward
verified: false
A(n) APPO model trained on the atari_tennis environment.
This model was trained using Sample-Factory 2.0: https://github.com/alex-petrenko/sample-factory. 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_tennis
About the Model
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.
The aim is to reach state-of-the-art (SOTA) performance on each atari environment. I will flag the models with SOTA when they reach at or near these levels.
The hyperparameters used in the model are the ones I have pushed to my fork of sample-factory: https://github.com/MattStammers/sample-factory. Given that https://huggingface.co/edbeeching has kindly shared his. I saved time and energy by using many of his tuned hyperparameters to maximise performance. However, he used 2 billion training steps. I have started as explained above at 10 million then moved to 100m to see how performance goes:
hyperparameters = {
"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,
}
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_tennis --train_dir=./train_dir --experiment=APPO-atari_tennis
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_tennis --train_dir=./train_dir --experiment=APPO-atari_tennis --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.