Inferno

Inferno is a little library providing utilities and convenience functions/classes around PyTorch. It's a work-in-progress, but the latest release (0.3.1) should be fairly stable!

• Free software: Apache Software License 2.0

• Documentation: http://inferno-pytorch.readthedocs.io (Work in Progress).

Features

• Current features include:

• 
  

• a basic Trainer class to encapsulate the training boilerplate (iteration/epoch loops, validation and checkpoint creation),

• a graph API for building models with complex architectures, powered by networkx.

• easy data-parallelism over multiple GPUs,

• a submodule for torch.nn.Module-level parameter initialization,

• a submodule for data preprocessing / transforms,

• support for Tensorboard (best with atleast tensorflow-cpu installed)

• a callback API to enable flexible interaction with the trainer,

• various utility layers with more underway,

• a submodule for volumetric datasets, and more!

import torch.nn as nnfrom inferno.io.box.cifar import get_cifar10_loadersfrom inferno.trainers.basic import Trainerfrom inferno.trainers.callbacks.logging.tensorboard import TensorboardLoggerfrom inferno.extensions.layers.convolutional import ConvELU2Dfrom inferno.extensions.layers.reshape import Flatten# Fill these in:LOG_DIRECTORY = '...'SAVE_DIRECTORY = '...'DATASET_DIRECTORY = '...'DOWNLOAD_CIFAR = TrueUSE_CUDA = True# Build torch modelmodel = nn.Sequential(
    ConvELU2D(in_channels=3, out_channels=256, kernel_size=3),
    nn.MaxPool2d(kernel_size=2, stride=2),
    ConvELU2D(in_channels=256, out_channels=256, kernel_size=3),
    nn.MaxPool2d(kernel_size=2, stride=2),
    ConvELU2D(in_channels=256, out_channels=256, kernel_size=3),
    nn.MaxPool2d(kernel_size=2, stride=2),
    Flatten(),
    nn.Linear(in_features=(256 * 4 * 4), out_features=10),
    nn.LogSoftmax(dim=1)
)# Load loaderstrain_loader, validate_loader = get_cifar10_loaders(DATASET_DIRECTORY,                                                    download=DOWNLOAD_CIFAR)# Build trainertrainer = Trainer(model) 
  .build_criterion('NLLLoss') 
  .build_metric('CategoricalError') 
  .build_optimizer('Adam') 
  .validate_every((2, 'epochs')) 
  .save_every((5, 'epochs')) 
  .save_to_directory(SAVE_DIRECTORY) 
  .set_max_num_epochs(10) 
  .build_logger(TensorboardLogger(log_scalars_every=(1, 'iteration'),                                  log_images_every='never'),                log_directory=LOG_DIRECTORY)# Bind loaderstrainer 
    .bind_loader('train', train_loader) 
    .bind_loader('validate', validate_loader)if USE_CUDA:
  trainer.cuda()# Go!trainer.fit()

To visualize the training progress, navigate to LOG_DIRECTORY and fire up tensorboard with

$ tensorboard --logdir=${PWD} --port=6007

and navigate to localhost:6007 with your browser.

Installation

Conda packages for python >= 3.6 for all distributions are availaible on conda-forge:

$ conda install -c pytorch -c conda-forge inferno

Future Features:

• Planned features include:

• 
  

• a class to encapsulate Hogwild! training over multiple GPUs,

• minimal shape inference with a dry-run,

• proper packaging and documentation,

• cutting-edge fresh-off-the-press implementations of what the future has in store. :)

Credits

All contributors are listed here_. .. _here: https://inferno-pytorch.github.io/inferno/html/authors.html

This package was partially generated with Cookiecutter and the audreyr/cookiecutter-pypackage project template + lots of work by Thorsten.