Open siaoliu opened 11 months ago
I use a 3080 or 2080 for all my experiments. Training time is about 1 day at max. (block stacking finetuning stage is another extra day but it converges far earlier)
It trains faster if you have more CPUs and can increase the number of training envs
Thanks for your reply! I have attempted this on my device. However, I encountered some issues.
env = SubprocVecEnv([make_env(i) for i in range(exp_cfg.n_envs)])
get errors
ConnectionResetError: [Errno 104] Connection reset by peer
I will reclone this repo in my laptop, really hope get your help~ Your work has opened up entirely new perspectives on generalization in robotics for my perspectives, and this line of research deserves greater attention from more researchers.
it would be appreciated if the author could provide more detailed installation instructions. I tried using the check_env function in stable-baselines3 but encountered issues creating the environment, likely due to version incompatibility between the various installed packages.
Thanks again for authors such an excellent work, I have solved the training issues and write a install.md as follows. When using a server without screen, there are some errors. Maybe the env does not support the headless mode?
Install torch
pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113
upgrade the setuptool & wheel for installing gym < 0.22
pip install --upgrade pip setuptools==57.5.0 wheel==0.38.3
install Mani-skill2 to install some dependencies, this package will remove later, but can install some package in one command.
pip install mani-skill2
install other denpendency
pip install stable-baselines3==1.8 transformers wandb omegaconf pyglet open3d tensorboard moviepy
install all envs
pip install minigrid dm_control x-magical
local install for most tasks
pip install -e ./paper_rl/
pip install -e .
pip install -e external/ManiSkill2
If want to try the Opendrawer task, we need to check the sapien version to v1:
conda create -n tr2_open --clone tr2
conda activate tr2_open
pip uninstall mani-skill2
pip install sapien==1.1
pip install external/ManiSkill
Sorry for the very delayed reply, It'll be a little hard to debug with just ConnectionResetError: [Errno 104] Connection reset by peer
. This bug happens usually if one of the parallel envs has an error. Could you show the full error log?
This was a fairly "old" project so I may need some time to go through my old environment and see if I can pull out a more reproducible environment.yml file for use with conda.
Thanks for such excellent work by the authors!!! Due to limited computational resources on my platform, I would like to inquire about the hardware requirements for reproducing this work. Is an Nvidia RTX-3090 sufficient for the task, and could you provide some information on the duration of each training session in your work?