Skip to content

Latest commit

 

History

History
101 lines (59 loc) · 3.48 KB

ENV.md

File metadata and controls

101 lines (59 loc) · 3.48 KB
Raw

Environment Setup

Step-1: Conda Environment

Recreate the Python 3.8 conda environment:

conda env create -f environment.yaml

Use a Different Python Environment (Optional)

If you want to use a different python environment, for example a newer version of python. First downgrade pip, setuptools and wheel (for compatibility with gym==0.21.0):

pip install pip==21 setuptools==65.5.0 wheel==0.38.0

Then use pip to install packages from the environment.yaml, according to pip-installing-environment-yml-as-if-its-a-requirements-txt.

Step-2: Install Cuda 12

Optional. Skip if you already have cuda 12 installed and CUDA_HOME environment variable set.

Download cuda_12.3.2_545.23.08_linux.run from NVIDIA CUDA Toolkit Archive and install it:

bash ./cuda_12.3.2_545.23.08_linux.run --silent   --toolkit --toolkitpath=$HOME/cuda-12.3
export CUDA_HOME=$HOME/cuda-12.3

Step-3: Install Dependencies

Some packages cannot be installed from the above conda env create directly, we need to install them separately.

ALOHA GYM

pip install gym-aloha==0.1.1 --ignore-requires-python

Pytorch3D

export NVCC_FLAGS="--generate-code arch=compute_80,code=sm_80 --generate-code arch=compute_86,code=sm_86 --generate-code arch=compute_87,code=sm_87 --generate-code arch=compute_89,code=sm_89" # adjust this according to your GPU

pip install git+https://github.com/facebookresearch/pytorch3d.git@stable

xformers

export TORCH_CUDA_ARCH_LIST="8.0;8.6;8.7;8.9"  # adjust this according to your GPU
pip install -v -U git+https://github.com/facebookresearch/xformers.git@main#egg=xformers

To speed up the installation, you can install ninja and set MAX_JOBS:

pip install ninja
export MAX_JOBS=1 # make it larger will speed up installation, buy may cause failure on xformers installation.

RLBench

CoppeliaSim & PyRep & RLBench

Please follow the instruction in https://github.com/stepjam/RLBench https://github.com/mlzxy/RLBench.arp.

If your server does not provide GUI and sudo permission, like mine, please follow the headless rendering tips for rlbench in Eval.md.

Note 📑: The only difference between RLBench.arp and the RLBench from PerAct (with rich language variants) is that I comment out an unnecessary length check at https://github.com/mlzxy/RLBench.arp/blob/3f57db0801b783a6c55fe1bf5bfea051882a17e9/rlbench/utils.py#L119. This enables me to simplify the RLBench dataset. See issue #1.

Faster Point-Renderer (optional but recommended)

Install by downloading point-renderer from RVT. I did not include it in the repo for LICENSE reason.

cd rlb
git clone https://github.com/NVlabs/RVT
cp -rf RVT/rvt/libs/point-renderer ./
rm -rf ./RVT/
cd ./point-renderer
# remove line "from mvt.utils import ForkedPdb" from `point_renderer/rvt_renderer.py`
pip install -e .
cd ../../

This is optional if you use pytorch3d renderer. To skip this, just set render_with_cpp=False in rlbench configuration files.

But the pretrained models are trained with this renderer, setting render_with_cpp=False will cause a little performance drop (they render slightly different pictures). But I believe it won't affect anything if you train from scratch.