keraflow.layers.convolution.ConvolutionBase Class Reference

Base layer for convolution layers. More...

Inheritance diagram for keraflow.layers.convolution.ConvolutionBase:

Public Member Functions
def	__init__

Detailed Description

Base layer for convolution layers.

Do not use this layer in your code.

Constructor & Destructor Documentation

def keraflow.layers.convolution.ConvolutionBase.__init__	(		self,
			kernel_shape,
			strides,
			pad = 'valid',
			bias = True,
			init = 'glorot_uniform',
			activation = 'linear',
			kwargs
	)

Parameters

kernel_shape	tuple of int. Shape of the kernel in the pattern (nb_kernel, k_rows, k_cols ...).
strides	tuple of int. Steps for sliding each kernel for convolution.
pad	str, 'valid' of 'same'. See descriptions below.
bias	boolean. Whether to include a bias (i.e. make the layer affine rather than linear).
init	str/function. Function to initialize trainable parameters. See Initializations.
activation	str/function. Activation function applied on the output. See Activations.
kwargs	see Layer.__init__.

Note

We follow tensorflow's padding strategy explained here.

1. When pad='same', the output length (for each dimension) is computed as:

   output_length = ceil( float(input_length) / stride )

   And the padding (for each dimension) is computed as:

   padding = (output_length - 1) * stride + kernel_length - input_length
   pad_front = floor(padding / 2)
   pad_back = (padding % 2==0)? floor(padding / 2): floor(padding / 2) + 1

   When stride=1, output_length will be equal to input_length, which is the reason for the name same.
2. When pad='valid', the output length (for each dimension) is computed as:

   output_length = ceil(float(input_length-kernel_length+1)/stride)

   And the padding is always 0. When stride=1, output_length will be equal to input_length-1.