From c50332de6bd3a825ea89388c7a126426e0d85eef Mon Sep 17 00:00:00 2001 From: AdamJelley Date: Wed, 19 Jun 2024 10:28:25 +0100 Subject: [PATCH] Update installation instructions. --- README.md | 9 ++++++--- requirements.txt | 9 --------- requirements/requirements.txt | 20 +++++++++++--------- 3 files changed, 17 insertions(+), 21 deletions(-) delete mode 100644 requirements.txt diff --git a/README.md b/README.md index d887a32..b024e22 100644 --- a/README.md +++ b/README.md @@ -24,13 +24,16 @@ Please note also that [ReBRAC](https://arxiv.org/abs/2305.09836) (developed conc ## Installation +Linux installation with [miniconda](https://docs.anaconda.com/miniconda/): ```bash git clone git@github.com:AdamJelley/EfficientOfflineRL.git && cd EfficientOfflineRL conda create -y -n EORL python=3.10 conda activate EORL -conda install -y -c conda-forge glew=2.1.0 glfw=3.3.10 -pip install -r requirements.txt +conda install -y -c conda-forge glew=2.1.0 glfw=1.12.0 +pip install -r requirements/requirements.txt ``` +Note that this will install [MuJoCo](https://github.com/google-deepmind/mujoco), with all dependencies and binaries included (thanks to [free-mujoco-py](https://github.com/vwxyzjn/free-mujoco-py)). However, this installation will only work on Linux. + Alternatively, you could use docker using the provided Dockerfile: ```bash docker build -t . @@ -55,7 +58,7 @@ Note these assume access to a CUDA device to run (otherwise set `--device cpu`). Other environments can be run by changing the config. Hyperparameters can be changed either in the relevant config or overridden with arguments as above. -The pre-training argument can also be set to `--pretrain BC` to only pretrain the actor (corresponding to the ablation in Appendix D of the paper), or ignored to run the default algorithm without pre-training. LayerNorm can also be removed from the actor or critic with `--actor_LN False` and `--critic_LN False` respectively. For `pretrain_steps`, `td_component` (referred to as $\lambda$ in the paper) and other hyperparameters used please see the paper. +The pre-training argument can also be set to `--pretrain BC` to only pretrain the actor (corresponding to the ablation in Appendix D of the paper), or ignored to run the default algorithm without pre-training. LayerNorm can also be removed from the actor or critic with `--actor_LN False` and `--critic_LN False` respectively. For `pretrain_steps`, `td_component` (referred to as $\lambda$ in the paper) and other hyperparameters used please see the paper. Videos of the agent during final evaluation can be recorded using `--render True`. ## Adroit Environments (Section 6 of Paper) diff --git a/requirements.txt b/requirements.txt deleted file mode 100644 index 8bf0f62..0000000 --- a/requirements.txt +++ /dev/null @@ -1,9 +0,0 @@ -git+https://github.com/Farama-Foundation/d4rl@master#egg=d4rl -cython==0.29.33 -numpy==1.23.1 -pandas==2.2.2 -patchelf==0.17.2.1 -pyrallis==0.3.1 -torch==2.3.1 -tqdm==4.64.1 -wandb==0.13.10 \ No newline at end of file diff --git a/requirements/requirements.txt b/requirements/requirements.txt index 00e3db5..212be1a 100644 --- a/requirements/requirements.txt +++ b/requirements/requirements.txt @@ -1,11 +1,13 @@ -# Main dependencies -git+https://github.com/tinkoff-ai/d4rl@master#egg=d4rl -tqdm==4.64.0 -wandb==0.12.21 -mujoco-py==2.1.2.14 +git+https://github.com/Farama-Foundation/d4rl@master#egg=d4rl +cython==0.29.33 +free-mujoco-py==2.1.6 +imageio[ffmpeg]==2.34.1 numpy==1.23.1 -gym[mujoco_py,classic_control]==0.23.0 ---extra-index-url https://download.pytorch.org/whl/cu113 -torch==1.11.0+cu113 -sortedcontainers==2.4.0 +pandas==2.2.2 +patchelf==0.17.2.1 +pre-commit==3.3.3 pyrallis==0.3.1 +ruff==0.0.278 +torch==2.3.1 +tqdm==4.64.1 +wandb==0.13.10 \ No newline at end of file