Add online training with TD-MPC as proof of concept (#338)

这个提交包含在:
Alexander Soare
2024-07-25 11:16:38 +01:00
提交者 GitHub
父节点 abbb1d2367
当前提交 f8a6574698
共有 25 个文件被更改,包括 1291 次插入 和 233 次删除

查看文件

@ -26,6 +26,7 @@ test-end-to-end:
${MAKE} DEVICE=$(DEVICE) test-diffusion-ete-train
${MAKE} DEVICE=$(DEVICE) test-diffusion-ete-eval
${MAKE} DEVICE=$(DEVICE) test-tdmpc-ete-train
${MAKE} DEVICE=$(DEVICE) test-tdmpc-ete-train-with-online
${MAKE} DEVICE=$(DEVICE) test-tdmpc-ete-eval
${MAKE} DEVICE=$(DEVICE) test-default-ete-eval
${MAKE} DEVICE=$(DEVICE) test-act-pusht-tutorial
@ -113,7 +114,6 @@ test-diffusion-ete-eval:
env.episode_length=8 \
device=$(DEVICE) \
# TODO(alexander-soare): Restore online_steps to 2 when it is reinstated.
test-tdmpc-ete-train:
python lerobot/scripts/train.py \
policy=tdmpc \
@ -133,6 +133,28 @@ test-tdmpc-ete-train:
training.image_transforms.enable=true \
hydra.run.dir=tests/outputs/tdmpc/
test-tdmpc-ete-train-with-online:
python lerobot/scripts/train.py \
env=pusht \
env.gym.obs_type=environment_state_agent_pos \
policy=tdmpc_pusht_keypoints \
eval.n_episodes=1 \
eval.batch_size=1 \
env.episode_length=10 \
device=$(DEVICE) \
training.offline_steps=2 \
training.online_steps=20 \
training.save_checkpoint=false \
training.save_freq=10 \
training.batch_size=2 \
training.online_rollout_n_episodes=2 \
training.online_rollout_batch_size=2 \
training.online_steps_between_rollouts=10 \
training.online_buffer_capacity=15 \
eval.use_async_envs=true \
hydra.run.dir=tests/outputs/tdmpc_online/
test-tdmpc-ete-eval:
python lerobot/scripts/eval.py \
-p tests/outputs/tdmpc/checkpoints/000002/pretrained_model \