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
andtorchtraining.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 allextras
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 duringpip install
[neptune]
[comet]
[tensorboard]