nnlayer_conv: Construct a convolution layer.

In andrie/mxNeuralNetExtra: Helper Functions to Make it Easier to Use neuralNet in the MicrosoftML Package

Description

Construct a convolution layer.

Usage

nnlayer_conv(layer, kernelshape, inputshape, name, inputname, stride, lowerpad,
  sharing, mapcount = 1, activation = c("rlinear"), ...)

Arguments

layer	A layer object, e.g. using nnlayer_input, or NULL
kernelshape	Numeric vector describing the number of inputs in each dimension, e.g. c(3, 15, 15)
inputshape	Numeric vector describing the number of inputs in each dimension, e.g. c(3, 15, 15). If layer is specified, this can be NULL.
name	Name of the layer
inputname	Name of the preceding layer. If layer is specified, this can be NULL.
stride	Numeric vector describing the number of inputs in each dimension, e.g. c(3, 15, 15).
lowerpad	Numeric vector describing the number of inputs in each dimension, e.g. c(3, 15, 15).
sharing	Logical vector.
mapcount	Number of maps to create.
activation	Activation function, e.g. rlinear
...	Other arguments passed from nnlayer_pool and nnlayer_norm

See Also

Other layer.definition.functions: nnlayer_full, nnlayer_input, nnlayer_norm, nnlayer_output, nnlayer_pool

Examples

# Use the layer functions to generate individual layer specifications

nnlayer_input(c(13, 13))
nnlayer_input(c(3, 7, 7), name = "pixels")

# Convolution layers automatically compute the output size and padding

nnlayer_conv(NULL, c(2, 2), 
           inputshape = c(13, 13), 
           name = "conv1", 
           inputname = "pixels"
)

nnlayer_conv(NULL, 
           c(2, 2), 
           inputshape = c(13, 13), 
           name = "conv1", 
           inputname = "pixels", 
           stride = c(2, 2)
)

nnlayer_conv(NULL, 
           c(1, 2, 2), 
           inputshape = c(3, 13, 13), 
           name = "conv1", 
           inputname = "pixels", 
           stride = c(1, 2, 2)
)

nnlayer_pool(NULL, 
           c(1, 2, 2), 
           inputshape = c(3, 13, 13), 
           name = "conv1", 
           inputname = "pixels", 
           stride = c(1, 2, 2)
)



# Specify the number of nodes in a fully connected layer

nnlayer_full(NULL, nodes = 100, name = "h3", inputname = "conv")

# Output layer

nnlayer_output(NULL, 6, name = "class", inputname = "h3")


# using magrittr pipes to connect layers ----------------------------------

require(magrittr)

nnlayer_input(c(3, 50, 50), name = "pixels") %>% 
  nnlayer_conv(
    kernelshape = c(1, 5, 5),
    name = "conv1", 
    stride = c(1, 2, 3)
  )

nnlayer_input(c(3, 50, 50), name = "pixels") %>% 
  nnlayer_conv(
    kernelshape = c(1, 5, 5),
    name = "conv1", 
    stride = c(1, 2, 3)
  ) %>% 
  nnlayer_pool(
    kernelshape = c(1, 5, 5),
    name = "conv1", 
    stride = c(1, 2, 3)
  )

nnlayer_norm(NULL, inputshape = c(3, 11, 5), kernelshape = c(1,5,5), name = "rnorm1", inputname = "conv")

nnlayer_input(c(3, 50, 50), name = "pixels") %>% 
  nnlayer_conv(
    kernelshape = c(1, 5, 5),
    name = "conv1", 
    stride = c(1, 2, 3)
  ) %>% 
  nnlayer_norm(
    kernelshape = c(1, 5, 5),
    name = "norm1", 
    stride = c(1, 2, 3),
    alpha = 0.0001,
    beta = 0.75
  )


nnlayer_input(c(3, 50, 50), name = "pixels") %>% 
  nnlayer_conv(
    kernelshape = c(3, 5, 5), 
    name = "conv1", 
    stride = c(1, 2, 2),
    mapcount = 48
  ) %>% 
  nnlayer_conv(
    kernelshape = c(1, 4, 4), 
    stride = c(1, 2, 2),
    name = "conv2"
  ) %>% 
  nnlayer_full(nodes = 100, name = "hid1") %>% 
  nnlayer_full(nodes = 30, name = "hid2") %>% 
  nnlayer_output(nodes = 6, name = "class")

andrie/mxNeuralNetExtra documentation built on May 10, 2019, 11:19 a.m.