Skip to content

Latest commit

 

History

History
160 lines (106 loc) · 4 KB

mnist.md

File metadata and controls

160 lines (106 loc) · 4 KB

Load MNIST data

The MNIST dataset containts 70.000 hand-written digits. Each digit is coded as 785 variables, that is, 28x28 pixel grey values and 1 label encoding the digit.

library(snedata)

# fetch the MNIST data set from the MNIST website
if (!exists("mnist"))
  mnist <- download_mnist()

# view the fifth digit
show_mnist_digit(mnist, 2)

Convert data

Convert the first 1000 digits to bnnlib sequence format.

library(bnnlib)

## Loading required package: ggplot2

## Loading required package: gridExtra

seqset = SequenceSet()

n_seq <- 1000

for (i in 1:n_seq) {
  seq = Sequence()
  input_vals <- simplify2array(mnist[i,1:784]/256)
  target_vals <- bnnlib:::onehot(mnist[i,785], alphabet = 0:9)
#  Sequence_add_from_array(seq, input_vals, target_vals, 784, 10)
  Sequence_add_from_array(seq, input_vals, input_vals, 784, 784)
  SequenceSet_add_copy_of_sequence(seqset, seq)
}

labels <-  mnist[1:n_seq, 785]

Train a network

One thousand digits with 200 epochs take about 90s on a 2020 laptop.

dims<-784

net_sparse <- NetworkFactory_createSparseAutoencoder(dims,10,0.1);

bp <- ImprovedRPropTrainer(net_sparse)
Trainer_sparsity_beta_set(bp, 0.1)

## NULL

library(tictoc)
tic()
Trainer_train__SWIG_0(bp, seqset, 50)

## NULL

toc()

## 25.998 sec elapsed

library(ggplot2)
plotTrainingerror(bp) + scale_y_log10()

## `geom_smooth()` using method = 'loess' and formula 'y ~ x'

Plot predictions

plotPreds <- function(sequence, net, dim=28) {
  x = getOutputs(net, sequence )
  #x2 = sapply(1:(84*84), function(x){ Sequence_get_input_at(sequence,0,x-1)})
  x2 = getInputs(sequence)
  #getInputs()
  predimg <- matrix(nrow=dim, ncol=dim)
  inpimg <- matrix(nrow=dim, ncol=dim)
  for (i in 1:dim) {
    predimg[i,] <- x[1:dim+(i-1)*dim]
    inpimg[i,] <- x2[1:dim+(i-1)*dim]
  }
  
  par(mfrow=c(2,2))
  image(inpimg, main="Input")
  image(predimg, main="Prediction")
  image(predimg-inpimg, main="Difference")
  image(abs(predimg-inpimg),main="Absolute Difference")
}


plotPreds(SequenceSet_get(seqset,4), net_sparse)


plotPreds(SequenceSet_get(seqset,102), net_sparse)


plotPreds(SequenceSet_get(seqset,8), net_sparse)


plotPreds(SequenceSet_get(seqset,40), net_sparse)


plotPreds(SequenceSet_get(seqset,41), net_sparse)

Plot latent space projection

pos <- which(sapply(getNodeNames(net_sparse) , function(x){startsWith(x,"Hidden")}))

result <- matrix(nrow=0,ncol=length(pos))
for (i in 1:SequenceSet_size(seqset)) {
  sequence <- SequenceSet_get(seqset,i-1)
  x = getActivations(net_sparse, sequence)[1,pos]
  result <- rbind(result, x)
}

result <- data.frame(result)

cols <- rainbow(10)
par(mfrow=c(1,1))
if (sum(dist(result))!=0) {
  coords <- cmdscale(dist(result),2)
  plot(coords,type="n")
  text(x=coords[,1],y=coords[,2], labels=labels,col = cols[labels])
}

Modify network

Remove input layer

ens <- Network_get_ensemble_by_name(net_sparse, "Input")  
Network_remove_ensemble(net_sparse, ens)
Network_sort_nodes(net_sparse)
Network_in_size_set(net_sparse, 10)

Create artificial inputs

  seq = Sequence()
  input_vals <- rep(0,10)
  Sequence_add_from_array(seq, input_vals, input_vals, 784, 784)


getActivations(net_sparse, seq)