Refactor eval.py (#127)

这个提交包含在:
Alexander Soare
2024-05-03 17:33:16 +01:00
提交者 GitHub
父节点 b7b69fcc3d
当前提交 bccee745c3
共有 12 个文件被更改,包括 457 次插入 和 298 次删除

查看文件

@ -28,7 +28,8 @@ training:
action: "[i / ${fps} for i in range(${policy.chunk_size})]"
eval:
n_episodes:: 50
n_episodes: 50
batch_size: 50
# See `configuration_act.py` for more details.
policy:

查看文件

@ -28,6 +28,7 @@ training:
eval:
n_episodes: 50
batch_size: 50
override_dataset_stats:
# TODO(rcadene, alexander-soare): should we remove image stats as well? do we use a pretrained vision model?

查看文件

@ -8,6 +8,7 @@ training:
eval_freq: 5000
online_steps_between_rollouts: 1
online_sampling_ratio: 0.5
online_env_seed: 10000
batch_size: 256
grad_clip_norm: 10.0