Added behavior cloning baselines #531
Conversation
examples/baselines/bc/bc.py
Outdated
| if args.video: | ||
| os.makedirs(os.path.join(log_path, args.video_path)) | ||
| gpu = args.sim_backend == "gpu" | ||
| envs = make_eval_envs(args.env, args.num_envs, stats, control_mode,gpu) |
There was a problem hiding this comment.
we should have a CLI argument to modify the max episode steps number to be higher from the command line. The code assumes 100 at the moment? You can replace max_traj_len with max_episode_steps as an argument. This should then also be passed to the environment creation tool.
examples/baselines/bc/bc_rgbd.py
Outdated
| os.makedirs(os.path.join(log_path, args.video_path)) | ||
|
|
||
|
|
||
| envs = make_eval_envs(args.env, args.num_envs, control_mode, gpu) |
There was a problem hiding this comment.
I recommend sticking to using the flatteng RGBD observation wrapper as used in the PPO rgb code instead of making all the extra process obs functions.
examples/baselines/bc/bc_rgbd.py
Outdated
| successes.append(fin_info["success"]) | ||
| returns.append(rewards[mask]) | ||
|
|
||
| log_dict["charts/eval_elapsed_steps"] = np.concatenate(el_steps).mean() |
There was a problem hiding this comment.
for logged data, can you try to be as consistent as possible with the diffusion policy baseline?
Namely there should be
eval/success_once
eval/success_at_end
eval/episode_len
eval/return
examples/baselines/bc/bc_rgbd.py
Outdated
| rewards = np.zeros((args.num_envs,)) | ||
| obs, _ = envs.reset() | ||
|
|
||
| for _ in range(100): |
There was a problem hiding this comment.
100 is hardcoded here. should be max episode steps
examples/baselines/bc/bc_rgbd.py
Outdated
| load_count=args.load_count, | ||
| ) | ||
|
|
||
| gpu = args.sim_backend == "gpu" |
There was a problem hiding this comment.
rename gpu to gpu_sim. gpu alone is a bit ambigious
examples/baselines/bc/bc_rgbd.py
Outdated
| group=config["group"], | ||
| name=log_name.replace(os.path.sep, "__"), | ||
| id=str(uuid.uuid4()), | ||
| tags=["BC_RGBD"] |
There was a problem hiding this comment.
for consistency lets name the tag "bc" for both bc rgbd and bc state, and also label group to be "BehaviorCloning". Default project name in the tyro args should be ManiSkill.
moreover why is there a ID generated here? We don't need to use uuid.
Added the behavior cloning baseline for both state and rgbd observations.