hope multivariate version

[angus924]

hope multivariate version

[angus924]

Thanks @Presburger.

I've uploaded a makeshift multivariate version of Hydra in hydra_multivariate.py. (Note that this is still experimental, I've only checked that it runs, I don't really know how well it works. It's probably also not particularly efficient at this stage. I will keep the univariate and multivariate code separate for the time being.)

The multivariate version works (at least, for the moment) by assigning a random subset of channels to each of the g groups (using groupwise/depthwise convolution with groups = g). You have to pass num_channels (i.e., the number of channels in the multivariate input) as an additional argument to the constructor. Otherwise its usage is the same as for the univariate version.

from hydra_multivariate import HydraMultivariate

[...] # load data

transform = HydraMultivariate(X_training.shape[-1], num_channels = X_training.shape[1])

[...]

A slight variation on this idea is to use multivariate/multichannel kernels. You just need to change torch.randn(divisor, k * _g, 1, 9) to torch.randn(divisor, k * _g, num_channels_per, 9) in __init__(...), and then change .sum(2) to .reshape(num_examples, -1, length) in forward(...).

[Presburger]

Thank You, @angus924 I will try to use this. minirocket is a excellent features extract tools. Thanks.