linear2d_layer_no_bias: add optional argument to disable biases

Author: OneAdder

src/nf/nf_layer_constructors.f90

@@ -213,12 +213,14 @@ module function reshape(output_shape) result(res)
         !! Resulting layer instance
     end function reshape
 
-    module function linear2d(out_features) result(res)
+    module function linear2d(out_features, biases) result(res)
       !! Rank-2 (sequence_length, out_features) linear layer constructor.
       !! sequence_length is determined at layer initialization, based on the
       !! output shape of the previous layer.
       integer, intent(in) :: out_features
         !! Number of output features
+      logical, optional :: biases
+        !! Whether to use biases or not
       type(layer) :: res
         !! Resulting layer instance
     end function linear2d

src/nf/nf_layer_constructors_submodule.f90

@@ -163,12 +163,13 @@ module function reshape(output_shape) result(res)
   end function reshape
 
 
-  module function linear2d(out_features) result(res)
+  module function linear2d(out_features, biases) result(res)
     integer, intent(in) :: out_features
+    logical, optional :: biases
     type(layer) :: res
 
     res % name = 'linear2d'
-    allocate(res % p, source=linear2d_layer(out_features))
+    allocate(res % p, source=linear2d_layer(out_features, biases))
 
   end function linear2d
 

src/nf/nf_linear2d_layer.f90

@@ -9,7 +9,8 @@ module nf_linear2d_layer
   public :: linear2d_layer
 
   type, extends(base_layer) :: linear2d_layer
-    integer :: sequence_length, in_features, out_features, batch_size
+    integer :: sequence_length, in_features, out_features
+    logical :: use_biases
 
     real, allocatable :: weights(:,:)
     real, allocatable :: biases(:)
@@ -31,8 +32,9 @@ module nf_linear2d_layer
   end type linear2d_layer
 
   interface linear2d_layer
-    module function linear2d_layer_cons(out_features) result(res)
+    module function linear2d_layer_cons(out_features, biases) result(res)
       integer, intent(in) :: out_features
+      logical, optional, intent(in) :: biases
       type(linear2d_layer) :: res
     end function linear2d_layer_cons
   end interface linear2d_layer

src/nf/nf_linear2d_layer_submodule.f90

@@ -5,11 +5,17 @@
 
 contains
 
-  module function linear2d_layer_cons(out_features) result(res)
+  module function linear2d_layer_cons(out_features, biases) result(res)
     integer, intent(in) :: out_features
+    logical, optional, intent(in) :: biases
     type(linear2d_layer) :: res
 
     res % out_features = out_features
+    if (present(biases)) then
+      res % use_biases = biases
+    else
+      res % use_biases = .true.
+    end if
 
   end function linear2d_layer_cons
 
@@ -36,8 +42,10 @@ module subroutine init(self, input_shape)
     allocate(self % dw(self % in_features, self % out_features))
     self % dw = 0
 
-    allocate(self % db(self % out_features))
-    self % db = 0
+    if (self % use_biases) then
+      allocate(self % db(self % out_features))
+      self % db = 0
+    end if
 
   end subroutine init
 
@@ -48,9 +56,11 @@ pure module subroutine forward(self, input)
     integer :: i
 
     self % output(:,:) = matmul(input(:,:), self % weights)
-    do concurrent(i = 1:self % sequence_length)
-      self % output(i,:) = self % output(i,:) + self % biases
-    end do
+    if (self % use_biases) then
+      do concurrent(i = 1:self % sequence_length)
+        self % output(i,:) = self % output(i,:) + self % biases
+      end do
+    end if
 
   end subroutine forward
 
@@ -64,7 +74,9 @@ pure module subroutine backward(self, input, gradient)
     integer :: i
 
     self % dw = self % dw + matmul(transpose(input(:,:)), gradient(:,:))
-    self % db = self % db + sum(gradient(:,:), 1)
+    if (self % use_biases) then
+      self % db = self % db + sum(gradient(:,:), 1)
+    end if
     self % gradient(:,:) = matmul(gradient(:,:), transpose(self % weights))
   end subroutine backward
 
@@ -74,7 +86,10 @@ pure module function get_num_params(self) result(num_params)
     integer :: num_params
 
     ! Number of weights times number of biases
-    num_params = self % in_features * self % out_features + self % out_features
+    num_params = self % in_features * self % out_features
+    if (self % use_biases) then
+      num_params = num_params + self % out_features
+    end if
 
   end function get_num_params
 
@@ -87,10 +102,14 @@ module function get_params(self) result(params)
 
     w_(1: product(shape(self % weights))) => self % weights
 
-    params = [ &
-      w_, &
-      self % biases &
-    ]
+    if (self % use_biases) then
+      params = [ &
+        w_, &
+        self % biases &
+      ]
+    else
+      params = w_
+    end if
 
   end function get_params
 
@@ -103,10 +122,14 @@ module function get_gradients(self) result(gradients)
 
     dw_(1: product(shape(self % dw))) => self % dw
 
-    gradients = [ &
-      dw_, &
-      self % db &
-    ]
+    if (self % use_biases) then
+      gradients = [ &
+        dw_, &
+        self % db &
+      ]
+    else
+      gradients = dw_
+    end if
 
   end function get_gradients
 
@@ -127,10 +150,12 @@ module subroutine set_params(self, params)
       p_(1:self % in_features, 1:self % out_features) => params(1 : n)
       self % weights = p_
 
-      ! reshape the biases
-      self % biases = params(n + 1 : n + self % out_features)
+      if (self % use_biases) then
+        ! reshape the biases
+        self % biases = params(n + 1 : n + self % out_features)
+      end if
     end associate
 
   end subroutine set_params
 
-end submodule nf_linear2d_layer_submodule
+end submodule nf_linear2d_layer_submodule