keras-tcn

Why we included this project

Most people reach for an LSTM or GRU when a sequence problem shows up, but this library makes it easy to try the other path. It wraps the Temporal Convolutional Network architecture as a Keras layer you can drop into an existing model with one import and a few keyword arguments. The dilated causal convolutions give it a longer effective memory than recurrent cells of similar size, and training parallelizes across time steps instead of stepping through them one at a time. It is maintained against recent TensorFlow releases, works from both Python and R, and includes working examples. Just note it is a building block, not a full platform: bring your own data pipeline and training loop.