Connect flatten, conv2d, and maxpool2d layers in backward pass (#142)

* Connect flatten, conv2d, and maxpool2d layers in backward pass

* Bump minor version

Author: milancurcic

fpm.toml

@@ -1,5 +1,5 @@
 name = "neural-fortran"
-version = "0.12.0"
+version = "0.13.0"
 license = "MIT"
 author = "Milan Curcic"
 maintainer = "milancurcic@hey.com"

src/nf/nf_network_submodule.f90

@@ -283,7 +283,6 @@ end function get_activation_by_name
   pure module subroutine backward(self, output)
     class(network), intent(in out) :: self
     real, intent(in) :: output(:)
-    real, allocatable :: gradient(:)
     integer :: n, num_layers
 
     num_layers = size(self % layers)
@@ -296,18 +295,25 @@ pure module subroutine backward(self, output)
         ! Output layer; apply the loss function
         select type(this_layer => self % layers(n) % p)
           type is(dense_layer)
-            gradient = quadratic_derivative(output, this_layer % output)
+            call self % layers(n) % backward( &
+              self % layers(n - 1), &
+              quadratic_derivative(output, this_layer % output) &
+            )
         end select
       else
         ! Hidden layer; take the gradient from the next layer
         select type(next_layer => self % layers(n + 1) % p)
           type is(dense_layer)
-            gradient = next_layer % gradient
+            call self % layers(n) % backward(self % layers(n - 1), next_layer % gradient)
+          type is(flatten_layer)
+            call self % layers(n) % backward(self % layers(n - 1), next_layer % gradient)
+          type is(conv2d_layer)
+            call self % layers(n) % backward(self % layers(n - 1), next_layer % gradient)
+          type is(maxpool2d_layer)
+            call self % layers(n) % backward(self % layers(n - 1), next_layer % gradient)
         end select
       end if
 
-      call self % layers(n) % backward(self % layers(n - 1), gradient)
-
     end do
 
   end subroutine backward