Instantiating a network from h5 with correct size

Author: milancurcic

test/test_parse_keras_model.f90

@@ -3,17 +3,24 @@ program test_parse_keras_model
   use iso_fortran_env, only: stderr => error_unit
   use nf_datasets, only: download_and_unpack, keras_model_dense_mnist_url
   use nf_io_hdf5, only: get_h5_attribute_string
+  use nf, only: layer, network, dense, input
   use json_module
 
   implicit none
 
   character(:), allocatable :: model_config_string
   character(*), parameter :: test_data_path = 'keras_dense_mnist.h5'
   type(json_core) :: json
-  type(json_value), pointer :: model_config, layers, next_layer, layer
-  character(:), allocatable :: class_name, layer_type
+  type(json_value), pointer :: &
+    model_config, layer_list, this_layer, layer_config
+  character(:), allocatable :: class_name, layer_type, activation
+  real, allocatable :: tmp_array(:)
+
+  type(layer), allocatable :: layers(:)
+  type(network) :: net
+
   logical :: found
-  integer :: n, num_layers
+  integer :: n, num_layers, num_elements
   logical :: file_exists
   logical :: ok = .true.
 
@@ -24,18 +31,44 @@ program test_parse_keras_model
     get_h5_attribute_string(test_data_path, '.', 'model_config')
 
   call json % parse(model_config, model_config_string)
-  call json % get(model_config, 'config.layers', layers)
+  call json % get(model_config, 'config.layers', layer_list)
+
+  num_layers = json % count(layer_list)
 
-  num_layers = json % count(layers)
+  layers = [layer ::]
 
+  ! Iterate over layers
   do n = 1, num_layers
-    call json % get_child(layers, n, layer)
-    !print *, 'Layer', n
-    !call json % print(layer)
-    call json % get(layer, 'class_name', layer_type)
-    !print *, layer_type
+
+    ! Get pointer to the layer
+    call json % get_child(layer_list, n, this_layer)
+
+    ! Get type of layer as a string
+    call json % get(this_layer, 'class_name', layer_type)
+
+    ! Get pointer to the layer config
+    call json % get(this_layer, 'config', layer_config)
+
+    ! Get size of layer and activation if applicable;
+    ! Instantiate neural-fortran layers at this time.
+    if (layer_type == 'InputLayer') then
+      call json % get(layer_config, 'batch_input_shape', tmp_array, found)
+      num_elements = tmp_array(2)
+      layers = [layers, input(num_elements)]
+    else if (layer_type == 'Dense') then
+      call json % get(layer_config, 'units', num_elements, found)
+      call json % get(layer_config, 'activation', activation, found)
+      layers = [layers, dense(num_elements, activation)]
+    else
+      error stop 'This layer is not supported'
+    end if
+
+    print *, n, layer_type, num_elements, activation
   end do
 
+  net = network(layers)
+  call net % print_info()
+
   if (.not. num_layers == 3) then
     ok = .false.
     write(stderr, '(a)') 'Keras dense MNIST model has 3 layers.. failed'