Additional notes and errata

Working with compiled Tensorflow binaries on GCP

Compute engine instances built with their deeplearning family of images includes pre-compiled tensorflow binaries. Hooray! How the hell do you use 'em?

These binaries are compiled for install with conda or mamba, and located in the following path.

/opt/deeplearning/binaries/tensorflow/{somename}-tf-{version}-{cpu|gpu}-{hash}.tar.bz2

You can install these binaries directly, but you'll first need to resolve some dependencies. My initial install list was:

mamba install -c conda-forge jupyter jupyterlab nodejs
pip install --upgrade pyarrow

Installing pyarrow from pip was required for me because going through mamba would have required downgrading tensorflow from 2.4.3 to 2.3.0. Dumb. Once you've got your dependencies square, identify the binary you want to install and pass it directly to mamba.

mamba install /opt/deeplearning/binaries/tensorflow/dlenv-tf-2-4-gpu-1.0.20210512-py37hfeb105c_0.tar.bz2

Fixing NVIDIA driver issues on GCP deep learning images

Though GCP has a series of deeplearning images, they are mostly built with broken NVIDIA drivers. The issue is that the driver configuration breaks once the instance is shut down. Why the fuck would they know this and not patch the image? I don't know, but this StackOverflow thread has instructions for fixing it (also posted on the google-dl-platform groups page).

Applying scalers to feature and response data

Here's a nifty article on different scalers to apply to model training data. Several of these methods are implemented in the myco.scalers module.