This repository contains source code for the initialization of convolutional layers in Keras. The initialization is based on Fuzzy modeling theory - Fuzzy transformation.
The initialization samples weights from the pre-defined 2D F-transform kernels and stack them up to create a filter tensor. Initialization of the first convolutional layer is unfortunately limited to 8 filters, second to 512 filters and any subsequent layer is basically without restriction. The initialization has reached better val/test accuracy on several datasets/architectures:
FTInit requires cv2, numpy, pillow, scipy.
Example of usage:
import FtInit
init = ftinit.FtInit()
# This creates a filter tensor with shape [5,5,1,1] and print its values.
print(init([5,5,1,1], dtype=float))
# Creates keras Conv2D layer with FtInit initializer
Conv2D(num_filters, kernel_size=kernel_size, strides=strides, padding="same" kernel_initializer=FtInit(), kernel_regularizer=l2(1e-4))