Where things go

The table gatle-ignite init writes into your project's README, generated from that same file so the two cannot drift.

One directory per concern. Create only the ones you need: this scaffold ships six.

You want	Put it in	Named by	Built-in alternative
a model	models/	cfg.model_name	none
a model's parts / method state	models/model_utils/	nothing: your model imports it	none
a dataset	dataloaders/	cfg.train_ds_name	none
a sampler	dataloaders/data_utils/	*_ds_params.sampler_params.cls_name	none
a collate fn	dataloaders/data_utils/	*_ds_params.collate_fn.cls_name	none
a vocab / raw data helper	dataloaders/data_utils/	nothing: your get_ds imports it	none
an augmentation	augmentation/	cfg.aug_name	none
a loss term	losses/loss_functions/	criterion_params.dict_of_loss_params[k].cls_name	none
a whole criterion	losses/	cfg.criterion_name	gatle_ignite.losses.composite
a metric	metrics/	{train,val,tester}_metrics[k].cls_name	none
an optimizer	optimizer/	cfg.optimizer_name	gatle_ignite.optimizers.*
an LR scheduler	scheduler/	cfg.lr_scheduler	gatle_ignite.schedulers.*
a trainer	trainer/	cfg.main_runner	none
a logger backend	callbacks/backends/	a dotted entry in cfg.logger_name	gatle_ignite.callbacks.backends.*
a script you run by hand	scripts/	nothing: you run it	none
an experiment	configs/	--config=	none

Copy-paste skeletons for each: Templates.