gatle-ignite

A config-driven pytorch-ignite training framework.

You write your custom functions: a config, a model, a dataset, a prep_batch. The framework owns the rest: engines, metrics, checkpointing, resume, logging, mixed precision, and distributed training.

Why this exists

gatle-ignite grew out of my own work over the years, from Kaggle competitions to my PhD research and the research that followed. I like working with pytorch-ignite and built my experiments on it, and they all shared the same template on top of it: the same trainer, config layout, checkpointing and logging, which I copied into each new project and then adapted. This package gathers that template into one place, to streamline that work and make future projects easier to start: a new one begins with gatle-ignite init rather than another round of copying.

It is shaped the way I already wrote my experiments: a config file, a model, a dataset and a prep_batch, with each piece named by a plain dotted path and kept in its own folder.

gatle-ignite is an independent project, not affiliated with or endorsed by the PyTorch-Ignite team.

The one rule

A config module is the single source of truth, and every component is selected by a dotted module path that gets imported at runtime. There is no registry and no decorator: you write a module that exposes a fixed entrypoint name, and you point a config at it.

cfg.model_name     = "models.mlp"                       # a module exposing Model(**model_params)
cfg.criterion_name = "gatle_ignite.losses.composite"
cfg.optimizer_name = "gatle_ignite.optimizers.adamw"    # builtins are just modules too

Builtins have no special status: they are named in full, exactly like yours. (The one exception is cfg.logger_name, a list of enabled sinks, which takes short names like "text".)

The whole task-specific trainer

from gatle_ignite import BaseTrainer, to_device

class Trainer(BaseTrainer):
    def prep_batch(self, batch, split="train", **kwargs):
        x, y = batch
        return to_device({"model_input": {"x": x}, "targets": {"labels": y}})

That is not an excerpt.

Install

pip install gatle-ignite

Requires Python ≥3.9, torch ≥2.4, pytorch-ignite ≥0.4.13 (0.4 and 0.5 both work). Extras are in the install guide.

Start a project

gatle-ignite init my_project && cd my_project
gatle-ignite train --config=configs/my_project_v0.py   # trains as-is, on synthetic data

One directory per concern (configs/, dataloaders/, losses/, metrics/, models/, trainer/), and a concern you do not have is a directory you do not create. The generated README says where a loss, a metric or a sampler goes, and which builtins mean you need not write one.

Or read the examples

git clone https://github.com/ryanwongsa/gatle-ignite.git && cd gatle-ignite
pip install -e ".[dev]"
gatle-ignite train --config=examples/synthetic/configs/synthetic_v0.py   # CPU, seconds, nothing to download

Nine worked tasks, each the same shape as what init writes: a GAN's two optimizers, a diffusion sampler at eval, seq2seq, distillation with a second model, EMA, contrastive.

Documentation

• Your first task: six small files
• Entrypoint contracts: what each config field expects
• Config fields: every field, generated from the code
• Status: what is verified, and what is not

License

BSD-3-Clause. See LICENSE.