Skip to content

Latest commit

 

History

History
22 lines (18 loc) · 2.58 KB

README.md

File metadata and controls

22 lines (18 loc) · 2.58 KB
Raw

👓 Exporting habitat-lab policies/networks in torchscript to load into spotsim2real environment without libraries and environment conflicts

You might want to train new policies or networks in habitat-lab. However, habitat-lab conda environment packages and spot-ros (used for spot-sim2real) environment packages might create version incompatabilities. Thus, we export a model in intermediate representation (IR) using torchscript module provided in Pytorch. Disentangling the deployment and development environment of a model, provides freedom to the model developer. We made a conversion script to convert mobile-gaze policy that was trained in new version of habitat-lab to torchscript model.

To use pytorch_to_torchscript.py , use python pytorch_to_torchscipt.py -c conversionparams.py add the relevant conversion parameters in conversionparams.yaml

In general, these are the steps you can follow for conversion.

  1. Load the pytorch model with class files, transfer the model on cuda
  2. Pass some random input tensor to the model and trace it's forward pass using torch.jit.trace, usage example can be found here
  3. Save the traced model as modelXX.torchscript, replace modelXX with desired name
  4. To load the model in spotsim2real use torch.jit.load(path/to/saved/torchscript/model, map_location="cuda:0")
  5. To download existing torchscript checkpoints run these commands from spot-sim2real folder
cd spot_rl_experiments/weights
git clone git clone https://huggingface.co/spaces/jimmytyyang/spot-sim2real-data
unzip spot-sim2real-data/weight/torchscript.zip && rm -rf spot-sim2real-torchscript-data && cd ../..

Solving Pytorch & CUDA error in recent Habitat-lab version setup

We encountered cuda error when setting up the recent habitat-lab version, it installs a recent version (2.2.1) of pytorch and CUDA 11.8. However, the hardware driver was older than 11.8 thus torch.cuda.is_available() was False and showing driver old error. To fix that first uninstall pytorch using pip uninstall pytorch torchvision torchaudio in your habitat-lab conda env then run the following conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch in same habitat-lab env (this is the pytorch and cuda version we use for spot-sim2real/spot-ros env)