Providing an easy way for devoloper to include irregular convolution in there models.

irregularly shaped kernels can reduce over fitting significantly. Although this has been known for years, there is not an easy way for devolopers to include irregular convolution in their models. This package aims to change that.

while traditional kernels look like this:

[w00, w01, w02,

w10, w11, w12, w20, w21, w22]

Irregular kernels look like this for example:

[0 , w01, 0,

w10, w11, w12, 0 , 0 , 0]

Two differently shaped kernels cannot learn to identify the same features. This promotes the network to learn more features and generalize better.

NOTE: at the moment, the only supported kernel size is (3, 3)

for Keras:

from IrregConv.keras_tools import IrregConv2D

you can use "IrregConv2D" in your keras model in the same way as normal keras layer. It takes all of the same inputs as keras.layers.Conv2D so the two can be easily interchanged. The two layers behave in the same way except "IrregConv2D" irregularly shaped kernels.

for PyTorch:

from IrregConv.torch_tools import IrregConv2D