Add Input2d layer by redesigning the input parameters to input layer constructors

Author: milancurcic

src/nf/nf_input2d_layer.f90

@@ -0,0 +1,48 @@
+module nf_input2d_layer
+
+  !! This module provides the `input2d_layer` type.
+
+  use nf_base_layer, only: base_layer
+  implicit none
+
+  private
+  public :: input2d_layer
+
+  type, extends(base_layer) :: input2d_layer
+    real, allocatable :: output(:,:)
+  contains
+    procedure :: init
+    procedure :: set
+  end type input2d_layer
+
+  interface input2d_layer
+    pure module function input2d_layer_cons(output_shape) result(res)
+      !! Create a new instance of the 2-d input layer.
+      !! Only used internally by the `layer % init` method.
+      integer, intent(in) :: output_shape(2)
+        !! Shape of the input layer
+      type(input2d_layer) :: res
+        !! 2-d input layer instance
+    end function input2d_layer_cons
+  end interface input2d_layer
+
+  interface
+
+    module subroutine init(self, input_shape)
+      !! Only here to satisfy the language rules
+      !! about deferred methods of abstract types.
+      !! This method does nothing for this type and should not be called.
+      class(input2d_layer), intent(in out) :: self
+      integer, intent(in) :: input_shape(:)
+    end subroutine init
+
+    pure module subroutine set(self, values)
+      class(input2d_layer), intent(in out) :: self
+        !! Layer instance
+      real, intent(in) :: values(:,:)
+        !! Values to set
+    end subroutine set
+
+  end interface
+
+end module nf_input2d_layer 

src/nf/nf_input2d_layer_submodule.f90

@@ -0,0 +1,23 @@
+submodule(nf_input2d_layer) nf_input2d_layer_submodule
+  implicit none
+contains
+
+  pure module function input2d_layer_cons(output_shape) result(res)
+    integer, intent(in) :: output_shape(2)
+    type(input2d_layer) :: res
+    allocate(res % output(output_shape(1), output_shape(2)))
+    res % output = 0
+  end function input2d_layer_cons
+
+  module subroutine init(self, input_shape)
+    class(input2d_layer), intent(in out) :: self
+    integer, intent(in) :: input_shape(:)
+  end subroutine init
+
+  pure module subroutine set(self, values)
+    class(input2d_layer), intent(in out) :: self
+    real, intent(in) :: values(:,:)
+    self % output = values
+  end subroutine set
+
+end submodule nf_input2d_layer_submodule 

src/nf/nf_layer_constructors.f90

@@ -35,10 +35,10 @@ module function input1d(layer_size) result(res)
         !! Resulting layer instance
     end function input1d
 
-    module function input3d(layer_shape) result(res)
-      !! 3-d input layer constructor.
+    module function input2d(dim1, dim2) result(res)
+      !! 2-d input layer constructor.
       !!
-      !! This layer is for inputting 3-d data to the network.
+      !! This layer is for inputting 2-d data to the network.
       !! Currently, this layer must be followed by a conv2d layer.
       !! An input layer must be the first layer in the network.
       !!
@@ -50,10 +50,29 @@ module function input3d(layer_shape) result(res)
       !! ```
       !! use nf, only :: input, layer
       !! type(layer) :: input_layer
-      !! input_layer = input([28, 28, 1])
+      !! input_layer = input(28, 28)
+      !! ```
+      integer, intent(in) :: dim1, dim2
+        !! First and second dimension sizes
+      type(layer) :: res
+        !! Resulting layer instance
+    end function input2d
+
+    module function input3d(dim1, dim2, dim3) result(res)
+      !! 3-d input layer constructor.
+      !!
+      !! This is a specific function that is available
+      !! under a generic name `input`.
+      !!
+      !! Example:
+      !!
+      !! ```
+      !! use nf, only :: input, layer
+      !! type(layer) :: input_layer
+      !! input_layer = input(28, 28, 1)
       !! ```
-      integer, intent(in) :: layer_shape(3)
-        !! Shape of the input layer
+      integer, intent(in) :: dim1, dim2, dim3
+        !! First, second and third dimension sizes
       type(layer) :: res
         !! Resulting layer instance
     end function input3d

src/nf/nf_layer_constructors_submodule.f90

@@ -5,6 +5,7 @@
   use nf_dense_layer, only: dense_layer
   use nf_flatten_layer, only: flatten_layer
   use nf_input1d_layer, only: input1d_layer
+  use nf_input2d_layer, only: input2d_layer
   use nf_input3d_layer, only: input3d_layer
   use nf_maxpool2d_layer, only: maxpool2d_layer
   use nf_reshape_layer, only: reshape3d_layer
@@ -81,16 +82,28 @@ module function input1d(layer_size) result(res)
   end function input1d
 
 
-  module function input3d(layer_shape) result(res)
-    integer, intent(in) :: layer_shape(3)
+  module function input2d(dim1, dim2) result(res)
+    integer, intent(in) :: dim1, dim2
     type(layer) :: res
     res % name = 'input'
-    res % layer_shape = layer_shape
+    res % layer_shape = [dim1, dim2]
     res % input_layer_shape = [integer ::]
-    allocate(res % p, source=input3d_layer(layer_shape))
+    allocate(res % p, source=input2d_layer([dim1, dim2]))
+    res % initialized = .true.
+  end function input2d
+
+
+  module function input3d(dim1, dim2, dim3) result(res)
+    integer, intent(in) :: dim1, dim2, dim3
+    type(layer) :: res
+    res % name = 'input'
+    res % layer_shape = [dim1, dim2, dim3]
+    res % input_layer_shape = [integer ::]
+    allocate(res % p, source=input3d_layer([dim1, dim2, dim3]))
     res % initialized = .true.
   end function input3d
 
+
   module function maxpool2d(pool_size, stride) result(res)
     integer, intent(in) :: pool_size
     integer, intent(in), optional :: stride
@@ -119,6 +132,7 @@ module function maxpool2d(pool_size, stride) result(res)
 
   end function maxpool2d
 
+
   module function reshape(output_shape) result(res)
     integer, intent(in) :: output_shape(:)
     type(layer) :: res

test/test_conv2d_layer.f90

@@ -29,7 +29,7 @@ program test_conv2d_layer
     write(stderr, '(a)') 'conv2d layer defaults to relu activation.. failed'
   end if
 
-  input_layer = input([3, 32, 32])
+  input_layer = input(3, 32, 32)
   call conv_layer % init(input_layer)
 
   if (.not. conv_layer % initialized) then
@@ -51,7 +51,7 @@ program test_conv2d_layer
   allocate(sample_input(1, 3, 3))
   sample_input = 0
 
-  input_layer = input([1, 3, 3])
+  input_layer = input(1, 3, 3)
   conv_layer = conv2d(filters, kernel_size)
   call conv_layer % init(input_layer)
 

test/test_conv2d_network.f90

@@ -11,7 +11,7 @@ program test_conv2d_network
 
   ! 3-layer convolutional network
   net = network([ &
-    input([3, 32, 32]), &
+    input(3, 32, 32), &
     conv2d(filters=16, kernel_size=3), &
     conv2d(filters=32, kernel_size=3) &
   ])
@@ -48,7 +48,7 @@ program test_conv2d_network
     call random_number(sample_input)
 
     cnn = network([ &
-      input(shape(sample_input)), &
+      input(1, 5, 5), &
       conv2d(filters=1, kernel_size=3), &
       conv2d(filters=1, kernel_size=3), &
       dense(1) &
@@ -84,7 +84,7 @@ program test_conv2d_network
     y = [0.1234567]
 
     cnn = network([ &
-      input(shape(x)), &
+      input(1, 8, 8), &
       conv2d(filters=1, kernel_size=3), &
       maxpool2d(pool_size=2), &
       conv2d(filters=1, kernel_size=3), &
@@ -119,7 +119,7 @@ program test_conv2d_network
     y = [0.12345, 0.23456, 0.34567, 0.45678, 0.56789, 0.67890, 0.78901, 0.89012, 0.90123]
 
     cnn = network([ &
-      input(shape(x)), &
+      input(1, 12, 12), &
       conv2d(filters=1, kernel_size=3), & ! 1x12x12 input, 1x10x10 output
       maxpool2d(pool_size=2), &           ! 1x10x10 input, 1x5x5 output
       conv2d(filters=1, kernel_size=3), & ! 1x5x5 input, 1x3x3 output

test/test_flatten_layer.f90

@@ -25,7 +25,7 @@ program test_flatten_layer
     write(stderr, '(a)') 'flatten layer is not initialized yet.. failed'
   end if
 
-  input_layer = input([1, 2, 2])
+  input_layer = input(1, 2, 2)
   call test_layer % init(input_layer)
 
   if (.not. test_layer % initialized) then
@@ -68,7 +68,7 @@ program test_flatten_layer
   end if
 
   net = network([ &
-    input([1, 28, 28]), &
+    input(1, 28, 28), &
     flatten(), &
     dense(10) &
   ])

test/test_get_set_network_params.f90

@@ -9,7 +9,7 @@ program test_get_set_network_params
 
   ! First test get_num_params()
   net = network([ &
-    input([3, 5, 5]), & ! 5 x 5 image with 3 channels
+    input(3, 5, 5), & ! 5 x 5 image with 3 channels
     conv2d(filters=2, kernel_size=3), & ! kernel shape [2, 3, 3, 3], output shape [2, 3, 3], 56 parameters total
     flatten(), &
     dense(4) & ! weights shape [72], biases shape [4], 76 parameters total
@@ -45,7 +45,7 @@ program test_get_set_network_params
 
   ! Finally, test set_params() and get_params() for a conv2d layer
   net = network([ &
-    input([1, 3, 3]), &
+    input(1, 3, 3), &
     conv2d(filters=1, kernel_size=3) &
   ])
 

test/test_input3d_layer.f90

@@ -10,7 +10,7 @@ program test_input3d_layer
   real, allocatable :: output(:,:,:)
   logical :: ok = .true.
 
-  test_layer = input([3, 32, 32])
+  test_layer = input(3, 32, 32)
 
   if (.not. test_layer % name == 'input') then
     ok = .false.

test/test_insert_flatten.f90

@@ -9,7 +9,7 @@ program test_insert_flatten
   logical :: ok = .true.
 
   net = network([ &
-    input([3, 32, 32]), &
+    input(3, 32, 32), &
     dense(10) &
   ])
 
@@ -19,7 +19,7 @@ program test_insert_flatten
   end if
 
   net = network([ &
-    input([3, 32, 32]), &
+    input(3, 32, 32), &
     conv2d(filters=1, kernel_size=3), &
     dense(10) &
   ])
@@ -32,7 +32,7 @@ program test_insert_flatten
   end if
 
   net = network([ &
-    input([3, 32, 32]), &
+    input(3, 32, 32), &
     conv2d(filters=1, kernel_size=3), &
     maxpool2d(pool_size=2, stride=2), &
     dense(10) &