torchtraining

torchtraining is a functional PyTorch neural network training library which provides high level building blocks and integrations instead of trying to do everything for you under the hood (most of current approaches, e.g. pytorch-lightning, skorch).

Using unified approach (single ** operator) across metrics, callbacks and others all users have to do is inherit from specific object (usually torchtraining.Operation) and implement desired forward or use provided building blocks by us.

Tutorials

See Google Colab tutorials to get a feel and what’s possible with this framework:

• Introduction - quick tour around functionalities with CIFAR100 classification and tensorboard

• GAN training - more advanced example and creating you own pipeline components

Modules

Below you can find available modules, so be sure to skim through those to see what’s currently possible.

• torchtraining
• torchtraining.accelerators
• torchtraining.accumulators
• torchtraining.callbacks
• torchtraining.cast
• torchtraining.device
• torchtraining.epochs
• torchtraining.exceptions
• torchtraining.functional
• torchtraining.iterations
• torchtraining.loss
• torchtraining.metrics
• torchtraining.pytorch
• torchtraining.quantization
• torchtraining.steps

Integrations

Integration are currently WIP, please file any issues you find along the way

Following are currently available:

• comet-ml - module torchtraining.callbacks.comet, ‘Self-hosted and cloud-based meta machine learning platform allowing data scientists and teams to track, compare, explain and optimize experiments and models.’

• neptune.ai - module torchtraining.callbacks.neptune, ‘The most lightweight experiment management tool that fits any workflow’

• tensorboard - module torchtraining.callbacks.tensorboard, ‘Visualization and tooling needed for machine learning experimentation’

• horovod - module torchtraining.accelerators.horovod and torchtraining.accelerators.Horovod, ‘Simple distributed training framework for TensorFlow, Keras, PyTorch, and Apache MXNet.’

All of those can be installed via extras, see below.

Installation

Following installation methods are available:

pip:

To install latest release:

pip install --user torchtraining

To install nightly version:

pip install --user torchtraining-nightly

torchtraining integrations come with extra. Simply run:

pip install --user torchtraining[neptune, tensorboard]

To install necessary packages for additional torchtraining.callbacks modules. Available extras:

• [all] - install all extras in one go

• [callbacks] - third party callbacks: [neptune, comet, tensorboard]

• [accelerators] - third party accelerators: [horovod]

• [horovod] - HOROVOD_CUDA_HOME=/path/to/cuda may be required during pip install

• [neptune]

• [comet]

• [tensorboard]