tidy up contributions from @jvdp1

Author: milancurcic

CMakeLists.txt

@@ -78,7 +78,7 @@ foreach(execid mnist network_save network_sync set_activation_function)
   add_test(test_${execid} bin/test_${execid})
 endforeach()
 
-foreach(execid mnist montesinos_uni montesinos_multi save_and_load simple sine)
+foreach(execid mnist mnist_epochs save_and_load simple sine)
   add_executable(example_${execid} src/tests/example_${execid}.f90)
   target_link_libraries(example_${execid} neural ${LIBS})
   add_test(example_${execid} bin/example_${execid})

README.md

@@ -266,13 +266,11 @@ program example_mnist
   batch_size = 100
   num_epochs = 10
 
-  if (this_image() == 1) then
-    write(*, '(a,f5.2,a)') 'Initial accuracy: ',&
-      net % accuracy(te_images, label_digits(te_labels)) * 100, ' %'
-  end if
+  if (this_image() == 1) print '(a,f5.2,a)', 'Initial accuracy: ', &
+    net % accuracy(te_images, label_digits(te_labels)) * 100, ' %'
 
   epochs: do n = 1, num_epochs
-    mini_batches: do i = 1, size(tr_labels) / batch_size
+    batches: do i = 1, size(tr_labels) / batch_size
 
       ! pull a random mini-batch from the dataset
       call random_number(pos)
@@ -286,12 +284,10 @@ program example_mnist
       ! train the network on the mini-batch
       call net % train(input, output, eta=3._rk)
 
-    end do mini_batches
+    end do batches
 
-    if (this_image() == 1) then
-      write(*, '(a,i2,a,f5.2,a)') 'Epoch ', n, ' done, Accuracy: ',&
-        net % accuracy(te_images, label_digits(te_labels)) * 100, ' %'
-    end if
+    if (this_image() == 1) print '(a,i2,a,f5.2,a)', 'Epoch ', n, ' done, Accuracy: ', &
+      net % accuracy(te_images, label_digits(te_labels)) * 100, ' %'
 
   end do epochs
 

src/lib/mod_network.f90

@@ -40,7 +40,7 @@ module mod_network
 
   interface network_type
     module procedure :: net_constructor
-  endinterface network_type
+  end interface network_type
 
 contains
 
@@ -286,19 +286,19 @@ subroutine train_epochs(self, x, y, eta, num_epochs, batch_size)
     nsamples = size(y, dim=2)
     nbatch = nsamples / batch_size
 
-    epoch: do n = 1, num_epochs
-     mini_batches: do i = 1, nbatch
+    epochs: do n = 1, num_epochs
+      batches: do i = 1, nbatch
 
-      !pull a random mini-batch from the dataset  
-      call random_number(pos)
-      batch_start = int(pos * (nsamples - batch_size + 1))
-      if (batch_start == 0) batch_start = 1
-      batch_end = batch_start + batch_size - 1
+        !pull a random mini-batch from the dataset  
+        call random_number(pos)
+        batch_start = int(pos * (nsamples - batch_size + 1))
+        if (batch_start == 0) batch_start = 1
+        batch_end = batch_start + batch_size - 1
 
-      call self % train(x(:,batch_start:batch_end), y(:,batch_start:batch_end), eta)
+        call self % train(x(:,batch_start:batch_end), y(:,batch_start:batch_end), eta)
 
-     enddo mini_batches
-    enddo epoch
+      end do batches
+    end do epochs
 
   end subroutine train_epochs
 

src/tests/example_mnist.f90

@@ -12,29 +12,44 @@ program example_mnist
 
   real(rk), allocatable :: tr_images(:,:), tr_labels(:)
   real(rk), allocatable :: te_images(:,:), te_labels(:)
+  real(rk), allocatable :: input(:,:), output(:,:)
 
   type(network_type) :: net
 
   integer(ik) :: i, n, num_epochs
-  integer(ik) :: batch_size
+  integer(ik) :: batch_size, batch_start, batch_end
+  real(rk) :: pos
 
   call load_mnist(tr_images, tr_labels, te_images, te_labels)
 
-  net = network_type([size(tr_images,dim=1), 10, size(label_digits(te_labels),dim=1)])
+  net = network_type([784, 30, 10])
 
-  batch_size = 1000
+  batch_size = 100
   num_epochs = 10
 
-  if (this_image() == 1) then
-    write(*, '(a,f5.2,a)') 'Initial accuracy: ',&
+  if (this_image() == 1) print '(a,f5.2,a)', 'Initial accuracy: ', &
+    net % accuracy(te_images, label_digits(te_labels)) * 100, ' %'
+
+  epochs: do n = 1, num_epochs
+    batches: do i = 1, size(tr_labels) / batch_size
+
+      ! pull a random mini-batch from the dataset
+      call random_number(pos)
+      batch_start = int(pos * (size(tr_labels) - batch_size + 1))
+      batch_end = batch_start + batch_size - 1
+
+      ! prepare mini-batch
+      input = tr_images(:,batch_start:batch_end)
+      output = label_digits(tr_labels(batch_start:batch_end))
+
+      ! train the network on the mini-batch
+      call net % train(input, output, eta=3._rk)
+
+    end do batches
+
+    if (this_image() == 1) print '(a,i2,a,f5.2,a)', 'Epoch ', n, ' done, Accuracy: ', &
       net % accuracy(te_images, label_digits(te_labels)) * 100, ' %'
-  end if
-
-  call net % train(tr_images, label_digits(tr_labels), 3._rk, num_epochs, batch_size)
-   
-  if (this_image() == 1) then
-    write(*, '(a,f5.2,a)') 'Epochs done, Accuracy: ',&
-     net % accuracy(te_images, label_digits(te_labels)) * 100, ' %'
-  endif
+
+  end do epochs
 
 end program example_mnist

src/tests/example_mnist_epochs.f90

@@ -0,0 +1,36 @@
+program example_mnist
+
+  ! A training example with the MNIST dataset.
+  ! Uses stochastic gradient descent and mini-batch size of 100.
+  ! Can be run in serial or parallel mode without modifications.
+
+  use mod_kinds, only: ik, rk
+  use mod_mnist, only: label_digits, load_mnist
+  use mod_network, only: network_type
+
+  implicit none
+
+  real(rk), allocatable :: tr_images(:,:), tr_labels(:)
+  real(rk), allocatable :: te_images(:,:), te_labels(:)
+
+  type(network_type) :: net
+
+  integer(ik) :: i, n, num_epochs
+  integer(ik) :: batch_size
+
+  call load_mnist(tr_images, tr_labels, te_images, te_labels)
+
+  net = network_type([size(tr_images, dim=1), 10, size(label_digits(tr_labels), dim=1)])
+
+  batch_size = 100
+  num_epochs = 10
+
+  if (this_image() == 1) print '(a,f5.2,a)', 'Initial accuracy: ', &
+    net % accuracy(te_images, label_digits(te_labels)) * 100, ' %'
+
+  call net % train(tr_images, label_digits(tr_labels), 3._rk, num_epochs, batch_size)
+   
+  if (this_image() == 1) print '(a,f5.2,a)', 'Epochs done, Accuracy: ', &
+    net % accuracy(te_images, label_digits(te_labels)) * 100, ' %'
+
+end program example_mnist