1. Run the following function to create a new convolution layer:

# Create a new convolution layer
create_conv_layer <- function(input,
num_input_channels,
filter_size,
num_filters,
use_pooling=True)
{
# Shape of the filter-weights for the convolution.
shape1 = shape(filter_size, filter_size, num_input_channels, num_filters)
# Create new weights
weights = weight_variable(shape=shape1)
# Create new biases
biases = bias_variable(shape=shape(num_filters))
# Create the TensorFlow operation for convolution.
layer = tf$nn$conv2d(input=input,
filter=weights,
strides=shape(1L, 1L, 1L ,1L),
padding="SAME")
# Add the biases to the results of the convolution.
layer = layer + biases
# Use pooling (binary flag) to reduce the image resolution
if(use_pooling){
layer = tf$nn$max_pool(value=layer,
ksize=shape(1L, 2L, 2L, 1L),
strides=shape(1L, 2L, 2L, 1L),
padding='SAME')
}
# Add non-linearity using Rectified Linear Unit (ReLU).
layer = tf$nn$relu(layer)
# Retrun resulting layer and updated weights
return(list("layer" = layer, "weights" = weights))
}

2. Run the following function to generate plots of convolution layers:

drawImage_conv <- function(index, images.bw, images.lab=NULL,par_imgs=8) {
require(imager)
img <- images.bw[index,,,]
n_images <- dim(img)[3]
par(mfrow=c(par_imgs,par_imgs), oma=c(0,0,0,0),
mai=c(0.05,0.05,0.05,0.05),ann=FALSE,ask=FALSE)
for(i in 1:n_images){
img.bwmat <- as.cimg(img[,,i])
# Extract the label
if(!is.null(images.lab)){
lab = labels[[1]][images.lab[[index]]]
}
# Plot and output label
plot(img.bwmat,axes=FALSE,ann=FALSE)
}
par(mfrow=c(1,1))
}

3. Run the following function to generate plots of convolution layer weights:

drawImage_conv_weights <- function(weights_conv, par_imgs=8) {
require(imager)
n_images <- dim(weights_conv)[4]
par(mfrow=c(par_imgs,par_imgs), oma=c(0,0,0,0),
mai=c(0.05,0.05,0.05,0.05),ann=FALSE,ask=FALSE)
for(i in 1:n_images){
img.r.mat <- as.cimg(weights_conv[,,1,i])
img.g.mat <- as.cimg(weights_conv[,,2,i])
img.b.mat <- as.cimg(weights_conv[,,3,i])
img.col.mat <- imappend(list(img.r.mat,img.g.mat,img.b.mat),"c") 
#Bind the three channels into one image
# Plot and output label
plot(img.col.mat,axes=FALSE,ann=FALSE)
}
par(mfrow=c(1,1))
}