TCN for Time Series Image Classification

[Vishu26]

I want to classify images where the features (channels) represent each time step. So, the input to TCN would have input_shape=(None, timesteps, height, width). The idea here is to incorporate spatial and temporal information for classification at the same time.

Is there a possibility I can modify TCN to use it as a ConvLSTM2D layer?

[philipperemy]

@Vishu26 you can't do it with TCN unless if you reshape your input to (None, timesteps, height * width). Which might be perfectly fine in terms of performance and accuracy.

There's an example that can do that. It's called MNIST pixel: https://github.com/philipperemy/keras-tcn/tree/master/tasks/mnist_pixel.

You can refer to the original paper in the README to learn about the task.

[Vishu26]

Thank you for the reply! I found a nice paper which implements what I need - https://arxiv.org/abs/1811.10166

[philipperemy]

Good to hear!