• namespace: Rindow\NeuralNetworks\Layer
• classname: Conv2D

2D convolution layer. This layer performs a two-dimensional convolution of the input data with kernel parameters. It can be used for convolution of 2D image data of 1 channel or more. Input data can be a 4D array of batch size, height, width and number of channels.

Conv2D implements the operation:

• output := convolution(input,kernel) + bias

Methods

constructor

$builer->Conv2D(
    int $filters,
    int|array $kernel_size,
    int|array $strides=1,
    string $padding='valid',
    string $data_format='channels_last',
    int|array $dilation_rate=1,
    int $groups=1,
    string|object $activation='linear',
    bool $use_bias=true,
    string|callable $kernel_initializer='glorot_uniform',
    string|callable $bias_initializer='zeros',
    string $kernel_regularizer=null,
    string $bias_regularizer=null,
    string $activity_regularizer=null,
    string $kernel_constraint=null,
    string $bias_constraint=null,
    array $input_shape=null,
    string $name=null,
)

You can create a Conv2D layer instances with the Layer Builder.

Arguments

• filters: Positive integer. Number of output filters in the convolution. Reflected in the number of dimensions in the output space.
• kernel_size: Positive integer or array of integer. For a single integer, use a kernel size that has the same height and width. Specify the window size when performing convolution. This size is reflected in the size of the kernel parameter.

Options

• strides: Positive integer or array of integer. For a single integer, use a kernel size that has the same height and width. Specify the stride length of the convolution. default is 1.
• padding: Either “valid” or “same”. If it is “valid”, there is no padding. Performs convolution with the valid range of input. In the case of “same”, the size of the input is expanded so that the output becomes the same size as the input, and that part is padded with zeros. default is “valid”.
• data_format: Either “channels_last” or “channels_first”. Specify which of the input shapes is the channel.
• dilation_rate: Positive integer or array of integer. Specify the dilation rate of the convolution. default is 1.
• input_shape: Specify the first layer the shape of the input data. In input_shape, the batch dimension is not included.
• activation: String of Activation function. Default null.
• use_bias: Boolean. the layer uses a bias vector. default true.
• kernel_initializer: name of kernel initializer
• bias_initializer: name of bias initializer

Input shape

[Batch size, input data height, input data width, number of channels] when data_format is “channels_last”, [Batch size, number of input channels, input data height, input data width] when “channels_first”. Four-dimensional NDArray.

Output shape

[Batch size, output data height, output data width, number of filters] regardless of input shape and data_format. Four-dimensional NDArray.

Examples

$model->add($nn->layers()->Conv2D(30,3,
    stride:1,
    input_shape:[28,28,3],
    kernel_initializer:'he_normal',
    bias_initializer:'zeros',
));