Temporal Convolutional Networks (Deep-TCN) in PyTorch

This repository provides an implementation of Temporal Convolutional Networks (TCN) [1] in PyTorch, with focus on flexibility and fine-grained control over architecture parameters.

Additionally, it incorporates separable convolutions and pooling layers, contributing to the creation of more streamlined and computationally efficient networks.

Installation

Using pip:

pip install deep-tcn

To install the dependencies for examples, run

pip install deep-tcn[examples]

Alternatively, you can clone the repository and install the package using poetry:

poetry install

This time, to install the dependencies needed to run the examples, run

poetry install --all extras

Features

• Causal Convolutions: Causal convolutions are employed, making the architecture suitable for sequential data.

• Separable Convolutions: The implementation includes support for separable convolutions, aiming to reduce the overall number of network parameters.

• (Channel) Pooling Layers: Channel pooling layers are integrated to further enhance the efficiency of the network by reducing dimensionality.

• Flexible Depth Configuration: Optionally, network depth can be increased by adding nondilated convolutions after dilated convolutional layers.

• Residual Blocks with Full Preactivation: Residual blocks are designed following the "full preactivation" design, according to [2]

• Supported Normalization Layers:

  • Group Normalization
  • Weight Normalization
  • Batch Normalization

Usage

Please refer to the scripts under examples/ as a starting point.

References

[1] He et al.: Identity Mappings in Deep Residual Networks. ArXiv, 2016. Link

[2] Lea et al.: Temporal Convolutional Networks: A Unified Approach to Action Segmentation. ArXiv, 2016. Link