search

rwth-i6 / i6_models

Collection of NN-Model parts
Mozilla Public License 2.0
1 stars 0 forks source link

Even kernel size for pooling layers in the GenericFrontend not permitted #50

Open kuacakuaca opened 4 months ago

kuacakuaca commented 4 months ago

https://github.com/rwth-i6/i6_models/blob/3c9173691521778b1e8b4070c172cbe929e4826b/i6_models/parts/frontend/generic_frontend.py#L88

albertz commented 4 months ago

What is this issue about? You think this is a problem that even kernel sizes are not permitted? So you want to change that?

kuacakuaca commented 4 months ago

What is this issue about? You think this is a problem that even kernel sizes are not permitted? So you want to change that?

I want to use even kernel size for the pooling layers, is there any reason we don’t want to allow this?

albertz commented 4 months ago

As far as I remember, there were technical reasons why it did not work. (I actually was against having the check here at that point, because this makes it unclear why it was not allowed. Just trying it out shows the problem much more clearly.) Maybe it's actually not a problem anymore in a more recent PyTorch version? I don't know. Just try it out.

JackTemaki commented 4 months ago

I think the limitation (from PyTorch side) was only for Convolution with "padding=same"...

kuacakuaca commented 4 months ago

thanks! I'll just try.