Skip to content

Install

pip install gatle-ignite

Extras

Nothing optional is installed by default, and nothing optional is imported unless you ask for it: training with logger_name = ["text"] never imports wandb.

Extra Pulls in For
(none) torch, pytorch-ignite, ml_collections, numpy, tqdm training
wandb wandb W&B logging
discord requests Discord webhook notifications
examples torchvision the MNIST example
dev pytest, pytest-cov, ruff running the test suite and the linter
docs mkdocs-material, mkdocstrings, mkdocs-gen-files building this site 
pip install "gatle-ignite[wandb,examples]"

Requirements

  • Python ≥ 3.9
  • torch ≥ 2.4
  • pytorch-ignite ≥ 0.4.13, < 0.6. Both 0.4 and 0.5 work, so this installs into an existing 0.4.13 environment with no upgrade.

A fresh install needs none of the care below: pip takes the newest torch and the newest numpy, and they agree. The pairings only matter when you hold a dependency back, which with torch is common.

Holding torch back? Hold numpy back too.

numpy 2.0 was an ABI break. A torch built against numpy 1.x cannot use numpy 2 arrays, and the failure arrives late: an import-time UserWarning, then a RuntimeError on the first .numpy() call, often deep inside a metric:

UserWarning: Failed to initialize NumPy: _ARRAY_API not found
RuntimeError: Numpy is not available

pip will not catch this. torch declares no numpy bound, so pip installs the broken pair without complaint, and this package's metadata cannot say "numpy < 2, but only with an older torch". torch 2.4.1 with numpy 2 has been checked and works. If you pin a torch built against numpy 1.x (a CUDA driver is the usual reason to pin), pin numpy with it:

pip install "torch==<the version you need>" "numpy<2"

torchvision decides which torch you get

torchvision pins the exact torch it was built against, so pip install ".[examples]" can move torch under you. This package cannot express that pairing without pinning torch itself, which would make it un-co-installable with anything wanting a different one. If you care which torch you have, install it first and let torchvision resolve against it, or pin both together.

A GPU is optional. The synthetic example trains on CPU in seconds. With 0 or 1 GPU a run is a single process with no distributed backend, so the command does not change.

From a checkout

git clone https://github.com/ryanwongsa/gatle-ignite.git
cd gatle-ignite
pip install -e ".[dev,docs]"
pytest -q
gatle-ignite train --config=examples/synthetic/configs/synthetic_v0.py