Apply optimizer to model weights without data copy

src/nf/nf_dense_layer.f90

@@ -34,8 +34,10 @@ module nf_dense_layer
     procedure :: backward
     procedure :: forward
     procedure :: get_gradients
+    procedure :: get_gradients_ptr
     procedure :: get_num_params
     procedure :: get_params
+    procedure :: get_params_ptr
     procedure :: init
     procedure :: set_params
 
@@ -96,6 +98,12 @@ module function get_params(self) result(params)
         !! Parameters of this layer
     end function get_params
 
+    module subroutine get_params_ptr(self, w_ptr, b_ptr)
+      class(dense_layer), intent(in), target :: self
+      real, pointer, intent(out) :: w_ptr(:)
+      real, pointer, intent(out) :: b_ptr(:)
+    end subroutine get_params_ptr
+
     module function get_gradients(self) result(gradients)
       !! Return the gradients of this layer.
       !! The gradients are ordered as weights first, biases second.
@@ -105,6 +113,12 @@ module function get_gradients(self) result(gradients)
         !! Gradients of this layer
     end function get_gradients
 
+    module subroutine get_gradients_ptr(self, dw_ptr, db_ptr)
+      class(dense_layer), intent(in), target :: self
+      real, pointer, intent(out) :: dw_ptr(:)
+      real, pointer, intent(out) :: db_ptr(:)
+    end subroutine get_gradients_ptr
+
     module subroutine set_params(self, params)
       !! Set the parameters of this layer.
       !! The parameters are ordered as weights first, biases second.

src/nf/nf_dense_layer_submodule.f90

@@ -77,6 +77,15 @@ module function get_params(self) result(params)
   end function get_params
 
 
+  module subroutine get_params_ptr(self, w_ptr, b_ptr)
+    class(dense_layer), intent(in), target :: self
+    real, pointer, intent(out) :: w_ptr(:)
+    real, pointer, intent(out) :: b_ptr(:)
+    w_ptr(1:size(self % weights)) => self % weights
+    b_ptr => self % biases
+  end subroutine get_params_ptr
+
+
   module function get_gradients(self) result(gradients)
     class(dense_layer), intent(in), target :: self
     real, allocatable :: gradients(:)
@@ -93,6 +102,15 @@ module function get_gradients(self) result(gradients)
   end function get_gradients
 
 
+  module subroutine get_gradients_ptr(self, dw_ptr, db_ptr)
+    class(dense_layer), intent(in), target :: self
+    real, pointer, intent(out) :: dw_ptr(:)
+    real, pointer, intent(out) :: db_ptr(:)
+    dw_ptr(1:size(self % dw)) => self % dw
+    db_ptr => self % db
+  end subroutine get_gradients_ptr
+
+
   module subroutine set_params(self, params)
     class(dense_layer), intent(in out) :: self
     real, intent(in), target :: params(:)

src/nf/nf_network_submodule.f90

@@ -649,6 +649,7 @@ module subroutine update(self, optimizer, batch_size)
     integer, intent(in), optional :: batch_size
     integer :: batch_size_
     real, allocatable :: params(:)
+    real, pointer :: weights(:), biases(:), dw(:), db(:)
     integer :: n
 
     ! Passing the optimizer instance is optional. If not provided, and if the
@@ -693,9 +694,21 @@ module subroutine update(self, optimizer, batch_size)
     end do
 #endif
 
-    params = self % get_params()
-    call self % optimizer % minimize(params, self % get_gradients() / batch_size_)
-    call self % set_params(params)
+    do n = 2, size(self % layers)
+      select type(this_layer => self % layers(n) % p)
+        type is(dense_layer)
+          call this_layer % get_params_ptr(weights, biases)
+          call this_layer % get_gradients_ptr(dw, db)
+          call self % optimizer % minimize(weights, dw / batch_size_)
+          call self % optimizer % minimize(biases, db / batch_size_)
+        type is(locally_connected1d_layer)
+          !TODO
+        type is(conv1d_layer)
+          !TODO
+        type is(conv2d_layer)
+          !TODO
+      end select
+    end do
 
     ! Flush network gradients to zero.
     do n = 2, size(self % layers)

src/nf/nf_optimizers.f90

@@ -248,4 +248,4 @@ pure subroutine minimize_adagrad(self, param, gradient)
 
   end subroutine minimize_adagrad
 
-end module nf_optimizers
+end module nf_optimizers