Output layer inconsistency

[dopplerchase]

When we forked this repo and made a few more flexible changes, there was a missed change. NOTE: This is not a breaking code issue. It is more of an inconsistency.

I am specifically referencing the kernel_size changes we made. The original package by Kyle didn't let you change the convolutional kernel size from the top level method (e.g., models.unet(...)). This made for a limited hyperparameter search for many using unets. Thus, if you go back into the history of the pulls here you will see how we passed the kernel_size paramater up to the top level method.

The issue opened here arises from the changes. Unfortunately, it looks like kernel_size was not passed into the output layer of the 3+ unets. The code issue is here, which is from https://github.com/dopplerchase/keras-unet-collection/blob/main/keras_unet_collection/_model_unet_3plus_2d.py

        OUT = CONV_output(X_decoder[0], n_labels, kernel_size=3, l1=l1, l2=l2,
                          activation=output_activation, name='{}_output_final'.format(name))

        model = Model([IN,], [OUT,])

It is possible that this hard coded kernel with size 3 is in other places, but I have not checked elsewhere yet.