itmo-nss-team / torchcnnbuilder

Framework for the automatic creation of CNN architectures

TorchCNNBuilder

---

TorchCNNBuilder is an open-source framework for the automatic creation of CNN architectures. This framework should first of all help researchers in the applicability of CNN models for a huge range of tasks, taking over most of the writing of the architecture code. This framework is distributed under the 3-Clause BSD license. All the functionality is written only using pytorch (no third-party dependencies)

Installation

---

The simplest way to install framework is using pip:

pip install torchcnnbuilder

Usage examples

---

The basic structure of the framework is presented below. Each subdirectory has its own example of using the appropriate available functionality. You can check <directory>_examples.ipynb files in order to see the ways to use the proposed toolkit. In short, there is the following functionality:

• the ability to calculate the size of tensors after (transposed) convolutional layers
• preprocessing an n-dimensional time series in TensorDataset
• automatic creation of (transposed) convolutional sequences
• automatic creation of (transposed) convolutional layers and (transposed) blocks from convolutional layers

The structure of the main part of the package:

├── examples
│ ├── builder_examples.ipynb
│ ├── preprocess_examples.ipynb
│ ├── models_examples.ipynb
│ └── tools                     # additional functions for the examples
└── torchcnnbuilder
    ├── preprocess
    │ └── time_series.py
    ├── builder.py
    └── models.py

Initially, the library was created to help predict n-dimensional time series (geodata), so there is a corresponding functionality and templates of predictive models (like ForecasterBase)

Sources

---

• Forecasting of Sea Ice Concentration using CNN, PDE discovery and Bayesian Networks
• Surrogate Modelling for Sea Ice Concentration using Lightweight Neural Ensemble

torchcnnbuilder's Issues

Design the repository

Добавить readme и requirements

Architecture refactoring

• Переписать классы и функции без дублирования кода с докстрингами и комментариями
• Избавиться от захардкоженных параметров, вынести их в переменные
• Подумать над более подходящим названием классов и функций

Add examples

Добавить примеры:

Задача предсказания:

• на тестовых данных о льде с разным числом каналов на входе и выходе
• на тестовых данных о льде с разной размерностью картинки на входе и выходе
• на тестовых данных изменяющихся структур (поискать видео, желательно с повторяющимися структурами пример)

Future plans

Here you can find some notes about plans for fixes or expansion of functionality in this framework:

• add support for pooling and sampling torch layers
• add support for 1- and 3-dimensional convolutions in the Builder class
• add several other templates of ready-made architectures (Segnet, Unet, etc.)
• add functionality for working with the internal vector representation of tensors