constantinpape / inferno

A utility library around PyTorch

inferno

Inferno is a little library providing utilities and convenience functions/classes around PyTorch. It's a work-in-progress, but the first release is underway!

Current 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!

Show me the Code!

import torch.nn as nn
from inferno.io.box.cifar10 import get_cifar10_loaders
from inferno.trainers.basic import Trainer
from inferno.trainers.callbacks.logging.tensorboard import TensorboardLogger
from inferno.extensions.layers.convolutional import ConvELU2D
from inferno.extensions.layers.reshape import Flatten

# Fill these in:
LOG_DIRECTORY = '...'
DATASET_DIRECTORY = '...'
DOWNLOAD_CIFAR = True
USE_CUDA = True

# Build torch model
model = 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.Softmax()
)

# Load loaders
train_loader, validate_loader = get_cifar10_loaders(DATASET_DIRECTORY,
                                                    download=DOWNLOAD_CIFAR)

# Build trainer
trainer = Trainer(model) \
  .build_criterion('CrossEntropyLoss') \
  .build_metric('CategoricalError') \
  .build_optimizer('Adam') \
  .validate_every((2, 'epochs')) \
  .save_every((5, 'epochs')) \
  .set_max_num_epochs(10) \
  .build_logger(TensorboardLogger(), log_directory=LOG_DIRECTORY)

# Bind loaders
trainer \
    .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.

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. :)

Contributing

Got an idea? Awesome! Start a discussion by opening an issue or contribute with a pull request.

Who's Who?

As of today, this library is maintained by Nasim Rahaman with sizeable contributions from Maurice Weiler and Steffen Wolf @ Image Analysis and Learning Lab, Heidelberg Collaboratory for Image Processing.