Library structure#

Introduction#

Eztorch as a library has a specific organization for how and where the files are created. Names of folders and files should be informative of their content.

General structure#

The General tree structure of Eztorch is the following:

eztorch
├── docs
├── eztorch
├── run
└── tests

Doc#

In docs/, you will find source documentation files about Eztorch such as in this file. We aim to make Eztorch accessible to anyone.

Eztorch#

The folder eztorch/ contains the actual library that we will detail below.

Run#

The folder run/ contains various scripts to make training or evaluation. These scripts thanks a Hydra configuration should perform a task agnostically to specific classes or architectures and be able to generalize to any relevant Hydra configuration.

Tests#

The folder tests/ contains various tests to ensure that components in Eztorch properly work.

As a general rule of thumb, the best would be to test individual components every time they are added, which is not the case for now.

Library structure#

The tree structure of the library is the following:

eztorch/eztorch
├── callbacks
├── configs
├── datamodules
├── datasets
├── evaluation
├── losses
├── models
├── optimizers
├── schedulers
├── transforms
├── utils
└── visualization

Callbacks#

The callbacks/ folder contains callbacks to feed the trainer based on Pytorch-Lightning.

Datamodules#

The datamodules/ folder contains datamodules for various image and video datasets.

Datasets#

The datasets/ folder contains datasets tools for various image and video datasets.

Among these tools, there are:

  • Datasets for videos, for example datasets/hmdb51.py.

  • Wrapper around datasets, for example datasets/dict_dataset.py.

  • Other tools such as clip samplers for video clip generation in datasets/clip_samplers/.

Evaluation#

The evaluation/ folder contains the tools for evaluation networks. For now, there are:

  • Linear classifier evaluation for SSL protocol.

  • Testing time augmentations functions to average predictions of several augmentations.

Losses#

The losses/ folder contains individual component losses.

Models#

The models/ folder contains models and tools to configure the various models supported by Eztorch for images and videos.

It follows the following structure:

eztorch/eztorch/models
├── heads
├── modules
├── siamese
└── trunks

Heads#

The heads/ folder contains various heads for your models.

Heads are the part of your models that should specialize to a task:

  • Linear heads for classification task.

  • MLP heads for self-supervised learning projections and predictions.

  • …

Modules#

The modules/ folder contains layers or modules for your models that accomplish a specific action.

It could be gathering tensors, or splitting batch normalization layers in parts, …

Siamese#

The siamese/ folder contains siamese models for self-supervised learning.

Currently implemented methods are:

  • MoCo (v1, v2, v3)

  • ReSSL

  • SCE

  • SimCLR

Trunks#

The trunks/ folder contains the trunks, backbone, or encoders, used in Eztorch for Image and Video modalities.

Timm and Pytorchvideo models are also available.

Optimizers#

The optimizers/ folder contains the custom optimizers (such as LARS) but also factories to easily instantiate them.

Schedulers#

The schedulers/ folder contains the custom schedulers (such as Cosine Annealing) but also factories to easily instantiate them.

Transforms#

The transforms/ folder contains all custom transforms defined by Eztorch for images and videos.

Transforms from Torchvision or Kornia are also available. The latter permits GPU transforms.

Utils#

The utils/ folder contains all utils functions to help Eztorch functionalities and allow various training tricks.

Visualization#

The visualization/ contains tools to visualize data.