pyml.neural_network.layer.transformation.convolutional.Convolutional

class Convolutional(in_channels, out_channels, kernel_shape, conv_operation='cross-correlation', padding=0, stride=1, weight_regularizer_l1=0, weight_regularizer_l2=0, bias_regularizer_l1=0, bias_regularizer_l2=0)

Bases: _Transformation, _TrainableTransformation

Convolutional layer for a neural network.

This class implements a convolutional layer with options for various parameters like kernel shape, padding, and stride. It is used for performing convolution or cross-correlation operations.

Parameters:

• in_channels (int) – The number of input channels or depth.

• out_channels (int) – The number of output channels.

• kernel_shape (int or tuple[int,int]) – The shape of the convolutional kernel. If it’s an integer, it’s treated as a square kernel. If it’s a tuple, it should specify the height and width of the kernel.

• conv_operation (str, optional) – The type of convolution operation, either ‘convolution’ or ‘cross-correlation’. By default ‘cross-correlation’. Not implemented yet.

• padding (int or tuple[int,int] or str, optional) – Padding to apply to the input. It can be an integer, a tuple of two integers for height and width padding, or one of the strings ‘valid’, ‘full’, or ‘same’. By default 0.

• stride (int or tuple[int,int], optional) – The stride to use during convolution. It can be an integer or a tuple of two integers specifying the vertical and horizontal strides. By default 1.

• weight_regularizer_l1 (float, optional) – L1 regularization strength for kernel weights, by default 0.

• weight_regularizer_l2 (float, optional) – L2 regularization strength for kernel weights, by default 0.

• bias_regularizer_l1 (float, optional) – L1 regularization strength for biases, by default 0.

• bias_regularizer_l2 (float, optional) – L2 regularization strength for biases, by default 0.

Variables:

• kernel_shape (tuple[int, int, int, int]) – The shape of the convolutional kernel in the format (output channels, input channels, height, width).

• flip_kernel (bool) – True if using cross-correlation, False if using convolution.

Methods

__init__
apply_kernel	Apply the convolution operation to a slice of the input using a given kernel and bias.
apply_padding	Apply padding to a given input matrix.
apply_striding	Performs striding on a matrix.
backward	Perform the backward pass through the convolutional layer.
forward	Perform the forward pass through the convolutional layer.
get_parameters	Return parameters
set_adjacent_layers	Set adjacent layers which are needed for the model to iterate through the layers.
set_parameters	Sets the parameters for this layer

apply_kernel(X_slice, kernel, bias)

Apply the convolution operation to a slice of the input using a given kernel and bias.

Return type:

int

Parameters:

• X_slice (numpy.ndarray) – A slice of the input matrix that matches the shape of the convolutional kernel.

• kernel (numpy.ndarray) – The convolutional kernel.

• bias (numpy.ndarray) – The bias associated with the kernel.

Returns:

The result of applying the convolution operation to the input slice using the given kernel and bias.

Return type:

int

apply_padding(X)

Apply padding to a given input matrix.

Return type:

ndarray

Parameters:

X (numpy.ndarray) – Input matrix of the shape (Number of Batches, Channels/Depth, Height, Width).

Returns:

Input matrix with applied padding.

Return type:

numpy.ndarray

static apply_striding(X, stride=1)

Performs striding on a matrix.

Return type:

ndarray

Parameters:

• X (numpy.ndarray) – Input matrix of the dimension (Number of Batches, Channels/Depth, Height, Width)

• stride (int or tuple[int, int], optional) – Specifies the iteration step size of rows and columns to keep. If set to 1, all columns and rows will be kept. Stride can either be a single int or a tuple, where the first parameter sets the stride for rows (height), and the second parameter sets the stride for columns (width). By default 1.

Returns:

Input matrix with applied stride.

Return type:

numpy.ndarray

backward(dvalues)

Perform the backward pass through the convolutional layer.

Return type:

None

Parameters:

dvalues (numpy.ndarray) – The gradients of the loss with respect to the layer’s output.

forward(inputs)

Perform the forward pass through the convolutional layer.

Return type:

None

Parameters:

inputs (numpy.ndarray) – Input data of the shape (Number of Batches, Channels/Depth, Height, Width).

get_parameters()

Return parameters

Return type:

tuple[ndarray, ndarray]

Returns:

Returns on first index the weights (kernel) and on second index the biases

Return type:

tuple[numpy.ndarray, numpy.ndarray]

set_adjacent_layers(previous_layer, next_layer)

Set adjacent layers which are needed for the model to iterate through the layers.

Parameters:

• previous_layer (_Layer) – Layer that is previous to this layer.

• next_layer (_Layer) – Layer that is subsequent to this layer.

set_parameters(weights, biases)

Sets the parameters for this layer

Return type:

None

Parameters:

• weights (numpy.ndarray) – Weights (kernel) of this layer.

• biases (numpy.ndarray) – Biases of this layer.