File size: 5,213 Bytes
42df274 3e87e2a 42df274 3e87e2a 42df274 3e87e2a 42df274 3e87e2a 42df274 3e87e2a 42df274 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 |
---
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_bowling
type: atari_bowling
metrics:
- type: mean_reward
value: 46.40 +/- 5.28
name: mean_reward
verified: false
---
A(n) **APPO** model trained on the **atari_bowling** 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_bowling
```
## 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_bowling --train_dir=./train_dir --experiment=APPO-atari_bowling
```
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_bowling --train_dir=./train_dir --experiment=APPO-atari_bowling --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.
|