conv: Create a convolutional layer

In ruta: Implementation of Unsupervised Neural Architectures

Create a convolutional layer

Description

Wrapper for a convolutional layer. The dimensions of the convolution operation are inferred from the shape of the input data. This shape must follow the pattern (batch_shape, x, [y, [z, ]], channel) where dimensions y and z are optional, and channel will be either 1 for grayscale images or generally 3 for colored ones.

Usage

conv(
  filters,
  kernel_size,
  padding = "same",
  max_pooling = NULL,
  average_pooling = NULL,
  upsampling = NULL,
  activation = "linear"
)

Arguments

filters	Number of filters learned by the layer
kernel_size	Integer or list of integers indicating the size of the weight matrices to be convolved with the image
padding	One of "valid" or "same" (case-insensitive). See layer_conv_2d for more details
max_pooling	NULL or an integer indicating the reduction ratio for a max pooling operation after the convolution
average_pooling	NULL or an integer indicating the reduction ratio for an average pooling operation after the convolution
upsampling	NULL or an integer indicating the augmentation ratio for an upsampling operation after the convolution
activation	Optional, string indicating activation function (linear by default)

Value

A construct with class "ruta_network"

See Also

Other neural layers: dense(), dropout(), input(), layer_keras(), output(), variational_block()

Examples

# Sample convolutional autoencoder
net <- input() +
 conv(16, 3, max_pooling = 2, activation = "relu") +
 conv(8, 3, max_pooling = 2, activation = "relu") +
 conv(8, 3, upsampling = 2, activation = "relu") +
 conv(16, 3, upsampling = 2, activation = "relu") +
 conv(1, 3, activation = "sigmoid")

ruta documentation built on Jan. 9, 2023, 1:20 a.m.