search

philipperemy / keras-tcn

Keras Temporal Convolutional Network.
MIT License
1.87k stars 454 forks source link

Question: Skip connections #254

Closed ajoomye closed 7 months ago

ajoomye commented 8 months ago

Hi, Thank you for the amazing repo. What does the skip connection do? If I want a simple TCN as shown in the examples with a kernel size of 2 and dilations [1, 2, 4, 8]. Do I have to set the skip connection as false? And if it is set to True, is there a separate connection between the input and output layer?

philipperemy commented 7 months ago

@ajoomye hi sorry for late reply. Skip connections allow a better gradient flow. Each output layer is connected to each other. The outputs are sum-ed and pass to the last layer to calculate the prediction. I'd say, set it to True always unless you have a good reason not to.

ajoomye commented 7 months ago

Dear @philipperemy , Thank you for the response.