Bug fixes; now everything works

Author: Riccardo231

example/cnn_mnist_1d.f90

@@ -1,7 +1,7 @@
 program cnn_mnist_1d
 
     use nf, only: network, sgd, &
-      input, conv2d, maxpool1d, maxpool2d, flatten, dense, reshape, reshape2d, locally_connected_1d, &
+      input, conv1d, conv2d, maxpool1d, maxpool2d, flatten, dense, reshape, reshape2d, locally_connected_1d, &
       load_mnist, label_digits, softmax, relu
 
     implicit none
@@ -12,7 +12,7 @@ program cnn_mnist_1d
     real, allocatable :: validation_images(:,:), validation_labels(:)
     real, allocatable :: testing_images(:,:), testing_labels(:)
     integer :: n
-    integer, parameter :: num_epochs = 10
+    integer, parameter :: num_epochs = 25
 
     call load_mnist(training_images, training_labels, &
                     validation_images, validation_labels, &
@@ -21,9 +21,9 @@ program cnn_mnist_1d
     net = network([ &
       input(784), &
       reshape2d([28,28]), &
-      locally_connected_1d(filters=8, kernel_size=3, activation=relu()), &
+      conv1d(filters=8, kernel_size=3, activation=relu()), &
       maxpool1d(pool_size=2), &
-      locally_connected_1d(filters=16, kernel_size=3, activation=relu()), &
+      conv1d(filters=16, kernel_size=3, activation=relu()), &
       maxpool1d(pool_size=2), &
       dense(10, activation=softmax()) &
     ])
@@ -37,7 +37,7 @@ program cnn_mnist_1d
         label_digits(training_labels), &
         batch_size=16, &
         epochs=1, &
-        optimizer=sgd(learning_rate=0.003) &
+        optimizer=sgd(learning_rate=0.005) &
       )
 
       print '(a,i2,a,f5.2,a)', 'Epoch ', n, ' done, Accuracy: ', accuracy( &

src/nf/nf_layer_submodule.f90

@@ -121,6 +121,8 @@ pure module subroutine backward_2d(self, previous, gradient)
           call this_layer % backward(prev_layer % output, gradient)
         type is(reshape2d_layer)
           call this_layer % backward(prev_layer % output, gradient)
+        type is(input2d_layer)
+          call this_layer % backward(prev_layer % output, gradient)
         type is(locally_connected_1d_layer)
           call this_layer % backward(prev_layer % output, gradient)
         type is(conv1d_layer)
@@ -134,6 +136,8 @@ pure module subroutine backward_2d(self, previous, gradient)
             call this_layer % backward(prev_layer % output, gradient)
           type is(reshape2d_layer)
             call this_layer % backward(prev_layer % output, gradient)
+          type is(input2d_layer)
+            call this_layer % backward(prev_layer % output, gradient)
           type is(locally_connected_1d_layer)
             call this_layer % backward(prev_layer % output, gradient)
           type is(conv1d_layer)
@@ -149,6 +153,8 @@ pure module subroutine backward_2d(self, previous, gradient)
             call this_layer % backward(prev_layer % output, gradient)
           type is(locally_connected_1d_layer)
             call this_layer % backward(prev_layer % output, gradient)
+          type is(input2d_layer)
+            call this_layer % backward(prev_layer % output, gradient)
           type is(conv1d_layer)
             call this_layer % backward(prev_layer % output, gradient)
         end select

src/nf/nf_network_submodule.f90

@@ -1,5 +1,6 @@
 submodule(nf_network) nf_network_submodule
 
+  use nf_conv1d_layer, only: conv1d_layer
   use nf_conv2d_layer, only: conv2d_layer
   use nf_dense_layer, only: dense_layer
   use nf_dropout_layer, only: dropout_layer
@@ -76,9 +77,9 @@ module function network_from_layers(layers) result(res)
             type is(conv2d_layer)
               res % layers = [res % layers(:n-1), flatten(), res % layers(n:)]
               n = n + 1
-            !type is(locally_connected_1d_layer)
-              !res % layers = [res % layers(:n-1), flatten(), res % layers(n:)]
-              !n = n + 1
+            type is(locally_connected_1d_layer)
+              res % layers = [res % layers(:n-1), flatten(), res % layers(n:)]
+              n = n + 1
             type is(maxpool2d_layer)
               res % layers = [res % layers(:n-1), flatten(), res % layers(n:)]
               n = n + 1
@@ -88,6 +89,9 @@ module function network_from_layers(layers) result(res)
             type is(maxpool1d_layer)
               res % layers = [res % layers(:n-1), flatten(), res % layers(n:)]
               n = n + 1
+            type is(conv1d_layer)
+              res % layers = [res % layers(:n-1), flatten(), res % layers(n:)]
+              n = n + 1
             !type is(reshape2d_layer)
             !  res % layers = [res % layers(:n-1), flatten(), res % layers(n:)]
             !  n = n + 1
@@ -179,6 +183,10 @@ module subroutine backward(self, output, loss)
             call self % layers(n) % backward(self % layers(n - 1), next_layer % gradient)
           type is(reshape2d_layer)
             call self % layers(n) % backward(self % layers(n - 1), next_layer % gradient)
+          type is(conv1d_layer)
+            call self % layers(n) % backward(self % layers(n - 1), next_layer % gradient)
+          type is(locally_connected_1d_layer)
+            call self % layers(n) % backward(self % layers(n - 1), next_layer % gradient)
         end select
       end if
 

test/test_conv1d_layer.f90

@@ -51,10 +51,6 @@ program test_conv1d_layer
     allocate(sample_input(1, 3))
     sample_input = 0
 
-    ! Print the sample input array
-    print *, "Sample Input:"
-    print *, sample_input
-  
     input_layer = input(1, 3)
     conv1d_layer = conv1d(filters, kernel_size)
     call conv1d_layer % init(input_layer)
@@ -66,10 +62,6 @@ program test_conv1d_layer
     call conv1d_layer % forward(input_layer)
     call conv1d_layer % get_output(output)
 
-    ! Print the output array after the forward pass
-    print *, "Output:"
-    print *, output
-  
     if (.not. all(abs(output) < tolerance)) then
       ok = .false.
       write(stderr, '(a)') 'conv1d layer with zero input and sigmoid function must forward to all 0.5.. failed'

test/test_conv1d_network.f90

@@ -9,7 +9,7 @@ program test_conv1d_network
     real, allocatable :: sample_input(:,:), output(:,:)
     logical :: ok = .true.
 
-    ! 2-layer convolutional network
+    ! 3-layer convolutional network
     net = network([ &
       input(3, 32), &
       conv1d(filters=16, kernel_size=3), &
@@ -26,7 +26,7 @@ program test_conv1d_network
     sample_input = 0
 
     call net % forward(sample_input)
-    call net % layers(2) % get_output(output)
+    call net % layers(3) % get_output(output)
 
     if (.not. all(shape(output) == [32, 28])) then
       write(stderr, '(a)') 'conv1d network output should have correct shape.. failed'
@@ -64,6 +64,7 @@ program test_conv1d_network
       end do
 
       if (.not. n <= num_iterations) then
+        
         write(stderr, '(a)') &
           'convolutional network 1 should converge in simple training.. failed'
         ok = .false.

test/test_locally_connected_1d_layer.f90

@@ -51,10 +51,6 @@ program test_locally_connected_1d_layer
     allocate(sample_input(1, 3))
     sample_input = 0
 
-    ! Print the sample input array
-    print *, "Sample Input:"
-    print *, sample_input
-  
     input_layer = input(1, 3)
     locally_connected_1d_layer = locally_connected_1d(filters, kernel_size)
     call locally_connected_1d_layer % init(input_layer)
@@ -66,9 +62,6 @@ program test_locally_connected_1d_layer
     call locally_connected_1d_layer % forward(input_layer)
     call locally_connected_1d_layer % get_output(output)
 
-    ! Print the output array after the forward pass
-    print *, "Output:"
-    print *, output
 
     if (.not. all(abs(output) < tolerance)) then
       ok = .false.