A repository of datascience / ai apis explored in far too much detail.

Quick Links to API Tests

You're probably most interested in the public api unittests in api_tests:

• torch_api
  • torch
    • I am currently working my way through the unary and binary cell-wise operations.

These exist, but they're mostly empty at the moment.

• numpy_api
• pandas_api

Generally, all test files have globally unique names. This makes IDE search functions much smoother, at the cost of some local smurfing. For instance, despite being nested in local contexts where their meaning is unambiguous, the following files exist:

• np_ones_test.py
• torch_ones_test.py

Packages

• Apt Packages apt.txt
• Snap Packages snap.txt
• Python virtualenv Packages (pip-compile / pip-sync)
  • requirements.in
  • (local) constraints.in
  • (local) requirements.txt

Note: You need a working pytorch install. I'm using a cuda1.11 preview build, see the --link in requirements.txt, you may need to change this. I am open to making the bootstrapping code conditional on local torch needs. Hmu.

• Run grind repo install_packages
• Run grind repo bootstrap
• Verify the install:
  • Run grind test --slow

If you encounter errors (I need a special version of torch for my cuda build) you may be able to satisfy them by adding constraints to the .gitignored file constraints.in, here is mine:

# pytorch with cu111 builds.
--find-links https://download.pytorch.org/whl/torch_stable.html
torch==1.10.1+cu111
torchvision==0.11.2+cu111
torchaudio==0.10.1+cu111

Pretty much everything gets run via:

grind presubmit

This will:

• check all code for style violations
• check all code for type violations
• remove unused imports
• reformat all code
• run all tests that haven't been run

See: grind README

grind provides a hook to run many repository commands.

• grind check
• grind format
• grind test
• grind fasttest
• grind presubmit

Grind supports tab completion. If you'd like to set it up, add the following hook to your bash environment:

source <$REPO_DIR>/tools/grind_completion

Tests are written using hamcrest, the most flexible pytest fluent testing framework I'm currently aware of is PyHamcrest

Hamcrest is great, but it doesn't really dig into structural comparison over tensor-like objects, so tests are also written in local hamcrest extension libraries, generally under the name eggs, torch_eggs, np_eggs, etc.

"Eggs and Ham", you see.

With some additional evolution, eggs may become a standalone package, but I'll co-develop it against these api tests, and extract it once I've got a solid usage corpus to iterate designs against.