Welcome to Torchélie’s documentation!

Pytorch Utils

• torchelie.nn
• torchelie.nn.utils
• torchelie.optim
• torchelie.lr_scheduler
• torchelie.utils
• torchelie.data_learning
• torchelie.loss
• torchelie.models
• torchelie.distributions
• torchelie.datasets
• torchelie.transforms
• torchelie.recipes

Algorithms and training

• Understanding recipes
• Using the predefined recipes and callbacks
• torchelie.callbacks
• torchelie.hyper

Torchélie aims to provide new utilities, layers, tools to pytorch users, as an extension library. It aims for minimal codependence between components, so that you can use only what you need with a strong pytorch flavour, without learning a new tool.

It provides:

• layers

• models (untrained)

• new optimizers and schedulers

• utility functions

• datasets and dataset utilities

• losses

• and some more specific things

Torchélie also tries to meet the likes of Ignite and Pytorch-Lightning by providing training loops with automatic logging, averaging, checkpointing, and visualisation. Unlike those however, instead of providing a “one size fits all” training loop, Torchélie aims to make writing them easy.

• Module Index

• Search Page