godshatter

Being a thousand independent shards of ideas isn't always fun, but at least it's not boring.

Development

Install and use pre-commit git hooks. pre-commit rules should be really generic and apply to the entire repo. If we want more specific per-project rules, we'll need a better git hooks project.

Dependency management

The projects in godshatter aren't meant to be built or packaged, so the project metadata is pretty simple!

We use pip and requirements files to specify dependencies. Requirements files are great because they allow us to reference other local projects and their dependencies. Plus, they're simple and well-maintained.
We use pyproject.toml to track basic project metadata.

Why not conda?

conda is great because it can manage the many non-Python dependencies modern machine learning depends on. But I couldn't get it to work well with editable installs or combine conda environments easily. Ideally, I'd specify the dependencies of each project with conda-build, but that was a lot of work. Manually setting up a separate environment.yml file per project was fine, but I couldn't import other local projects while getting their dependencies.

PaperSpace

Installing NVIDIA drivers was annoying. I used the GCP script to install it, but I ran into problems because the Nouveau driver wasn't getting removed automatically. So I created a blacklist for it by following these instructions.

GCP setup

GCP sucks. When I started up my VM instance a couple days later, I got resource exhaustion. There's no transparency on which regions have available resources.

I'm trying PaperSpace which is more transparent. Let's see if I run into issues here too.

Create instance

Choose some compute zone. You can see all available ones with:

gcloud compute zones list

Find which machine types are available in the desired zone:

gcloud compute machine-types list --filter="zone:( europe-west1-b )"

I don't like using the Linux for Deep Learning boot images because they have some dependencies installed via conda. That caused some problems with pyenv, and it was simpler to just not have to worry about conda. I installed the base Buster (Debian 10) image. Then:

Install CUDA drivers
Install git: sudo apt install git-all
Install pyenv after installing dependencies

Setup SSH and remote desktop with VSCode

It's nicer to add SSH keys to your GCP project metadata so that you can SSH into any instance without extra setup. Here's a guide.

You might need to go to "Metadata" on GCP and re-add the SSH key if you add a new instance.

Now just SSH. For example, the username and key I use is:

ssh -i ~/.ssh/gcp wba@EXTERNAL_IP

It's also nice to add this config to your ~/.ssh/config with some alias so you can just ssh t4.