Resnet 1D and Variable Length Pooling for time series data

Pooling from variable length of activations

This is useful for dealing features of various length in time dimension. E.g., phonemes in speech data.

Softmax             o1        o2        o3
                    |         |         |

VarLenAvgPooling    p1        p2        p3
                   / \     /     \     / \
                  |   |   |       |   |   |

Activation        a1  a2  a3  a4  a5  a6  a7
                  |   |   |   |   |   |   |

Conv1D            o   o   o   o   o   o   o
                 /|\ /|\ /|\ /|\ /|\ /|\ /|\

Activation        a1  a2  a3  a4  a5  a6  a7
                  |   |   |   |   |   |   |

Conv1D            o   o   o   o   o   o   o
                 /|\ /|\ /|\ /|\ /|\ /|\ /|\

Time             --------------------------->

ResNet for time series data

Vanilla ResNet uses Conv2D for image data. However this architecture may be useful for deep Conv1D networks as well.

I tried two approaches in my code:

• use rectangular filters (different H, W) directly in ResNet2D
• shift to Conv1D entirely

It depends on your specific problem to answer which approach is better.

Credits

Some code are copied from pytorch official examples and CMU classes. Credits to the original authors.

• https://github.com/pytorch/examples/tree/master/mnist
• ...