Installing Transformers on Jetson Nano

I’m finally back to playing around with my Jetson Nano. Transformers (in current and previous forms) has been my favorite library over the past 18 months or so. I decided to play around with it this evening, but found that it wasn’t trivial to install on this device. Here’s my guide to getting up and running:

Create Virtual Environment

pip install virtualenvwrapper
mkvirtualenv transformers
deactivate
workon transformers

Install PyTorch

A common theme in Jetson guides is that many libraries don’t have Python wheels for Jetson Nano’s ARM aarch64 architecture. Fortunately, in the case of PyTorch itself, it’s easy enough to find downloadable wheels on the Nvidia forums. Install PyTorch 1.4.0 and a few dependencies with:

wget https://nvidia.box.com/shared/static/ncgzus5o23uck9i5oth2n8n06k340l6k.whl \
    -O torch-1.4.0-cp36-cp36m-linux_aarch64.whl
sudo apt-get install libopenblas-base
pip install Cython
pip install numpy torch-1.4.0-cp36-cp36m-linux_aarch64.whl

Install SentencePiece

At the time of writing, if you try to pip install transformers at this point, you’ll first run into issues with SentencePiece. Google doesn’t have recent versions, besides this problematic latest release, available on PyPI. This is a bit annoying, but it’s easy enough to checkout an older tag (0.1.84) and install from source:

git clone https://github.com/google/sentencepiece
git checkout tags/v0.1.84
cd sentencepiece/python
python setup.py install

Install Tokenizers

Within the last few months, HuggingFace introduced Tokenizers, a dependency of and closely related sibling library to Transformers. Tokenizers accelerates things with Rust, so you’ll need to have it installed:

curl https://sh.rustup.rs -sSf | sh
rustc --version

The second line probably won’t work until you restart your session (this will add the Rust compiler to your .bashrc). If you’re connected via SSH, you probably just want to exit and then reconnect. After you do this, try that line again to make sure everything is setup correctly.

At this point, you shouldn’t have any issue installing with Pip:

pip install tokenizers

Install Transformers

At this point, you should be able to install Transformers (2.5.1, as of writing):

pip install transformers

I realized after the fact that this reinstalled / downgraded Tokenizers to the exact version listed by Transformers, so I could have skipped the Pip install of it before. It didn’t hurt anything, though.

Make sure to really prove to yourself that everything is working by jumping into a terminal, importing the library, and playing around with it. Maybe load a pre-trained model or use torch.cuda.is_available() to make sure CUDA is being used.

🚀