Merge pull request #139 from BerkeleyLab/normalize-in-inference

Feature: map input to unit range & output tensors from unit range in inference_engine_t infer procedure

cloud-microphysics/app/train-cloud-microphysics.f90

@@ -238,15 +238,15 @@ subroutine read_train_write(training_configuration, base_name, plot_unit, previo
         else
           close(network_unit)
 
-          print *,"Calculating inputs tensor component ranges."
+          print *,"Calculating input tensor component ranges."
           input_range = tensor_range_t( &
             layer  = "inputs", &
             minima = [minval(pressure_in), minval(potential_temperature_in), minval(temperature_in), &
               minval(qv_in), minval(qc_in), minval(qr_in), minval(qs_in)], &
             maxima = [maxval(pressure_in), maxval(potential_temperature_in), maxval(temperature_in), &
               maxval(qv_in), maxval(qc_in), maxval(qr_in), maxval(qs_in)] &
           )
-          print *,"Calculating outputs tensor component ranges."
+          print *,"Calculating output tensor component ranges."
           output_range = tensor_range_t( &
             layer  = "outputs", &
             minima = [minval(dpt_dt), minval(dqv_dt), minval(dqc_dt), minval(dqr_dt), minval(dqs_dt)], &
@@ -297,10 +297,10 @@ subroutine read_train_write(training_configuration, base_name, plot_unit, previo
             ] &
           ), lon = 1, size(qv_in,1))], lat = 1, size(qv_in,2))], level = 1, size(qv_in,3))], time = start_step, end_step, stride)]
 
-        print *,"Normalizing inputs tensors"
+        print *,"Normalizing input tensors"
         inputs = input_range%map_to_unit_range(inputs)
 
-        print *,"Normalizing outputs tensors"
+        print *,"Normalizing output tensors"
         outputs = output_range%map_to_unit_range(outputs)
 
         print *, "Eliminating",int(100*(1.-keep)),"% of the grid points that have all-zero time derivatives"

cloud-microphysics/fpm.toml

@@ -5,7 +5,7 @@ author = "Damian Rouson, Tan Nguyen, Jordan Welsman, David Torres, Brad Richards
 maintainer = "rouson@lbl.gov"
 
 [dependencies]
-assert = {git = "https://github.com/sourceryinstitute/assert", tag = "1.5.0"}
-sourcery = {git = "https://github.com/sourceryinstitute/sourcery", tag = "4.5.3"}
+assert = {git = "https://github.com/sourceryinstitute/assert", tag = "1.6.0"}
+sourcery = {git = "https://github.com/sourceryinstitute/sourcery", tag = "4.6.1"}
 inference-engine = {path = "../"}
 netcdf-interfaces = {git = "https://github.com/LKedward/netcdf-interfaces.git", rev = "d2bbb71ac52b4e346b62572b1ca1620134481096"}

src/inference_engine/inference_engine_m_.f90

@@ -23,7 +23,7 @@ module inference_engine_m_
   type inference_engine_t
     !! Encapsulate the minimal information needed to perform inference
     private
-    type(tensor_range_t) inputs_range_, outputs_range_
+    type(tensor_range_t) input_range_, output_range_
     type(string_t) metadata_(size(key))
     real(rkind), allocatable :: weights_(:,:,:), biases_(:,:)
     integer, allocatable :: nodes_(:)
@@ -59,13 +59,13 @@ module inference_engine_m_
 
   interface inference_engine_t
 
-    pure module function construct_from_padded_arrays(metadata, weights, biases, nodes, inputs_range, outputs_range) &
+    pure module function construct_from_padded_arrays(metadata, weights, biases, nodes, input_range, output_range) &
       result(inference_engine)
       implicit none
       type(string_t), intent(in) :: metadata(:)
       real(rkind), intent(in) :: weights(:,:,:), biases(:,:)
       integer, intent(in) :: nodes(0:)
-      type(tensor_range_t), intent(in), optional :: inputs_range, outputs_range
+      type(tensor_range_t), intent(in), optional :: input_range, output_range
       type(inference_engine_t) inference_engine
     end function
 

src/inference_engine/inference_engine_s.F90

@@ -36,9 +36,12 @@
 
     associate(w => self%weights_, b => self%biases_, n => self%nodes_, output_layer => ubound(self%nodes_,1))
 
+
       allocate(a(maxval(n), input_layer:output_layer))
 
-      a(1:n(input_layer),input_layer) = inputs%values()
+      associate(normalized_inputs => self%input_range_%map_to_unit_range(inputs))
+        a(1:n(input_layer),input_layer) = normalized_inputs%values()
+      end associate
 
       feed_forward: &
       do l = input_layer+1, output_layer
@@ -47,7 +50,9 @@
         end associate
       end do feed_forward
 
-      outputs = tensor_t(a(1:n(output_layer), output_layer))
+      associate(normalized_outputs => tensor_t(a(1:n(output_layer), output_layer)))
+        outputs = self%output_range_%map_from_unit_range(normalized_outputs)
+      end associate
 
     end associate
 
@@ -128,22 +133,22 @@ pure subroutine set_activation_strategy(inference_engine)
     block
       integer i
 
-      if (present(inputs_range)) then
-        inference_engine%inputs_range_ = inputs_range
+      if (present(input_range)) then
+        inference_engine%input_range_ = input_range
       else
         associate(num_inputs => nodes(lbound(nodes,1)))
           associate(default_minima => [(0., i=1,num_inputs)], default_maxima => [(1., i=1,num_inputs)])
-            inference_engine%inputs_range_ = tensor_range_t("inputs", default_minima, default_maxima)
+            inference_engine%input_range_ = tensor_range_t("inputs", default_minima, default_maxima)
           end associate
         end associate
       end if
 
-      if (present(outputs_range)) then
-        inference_engine%outputs_range_ = outputs_range
+      if (present(output_range)) then
+        inference_engine%output_range_ = output_range
       else
         associate(num_outputs => nodes(ubound(nodes,1)))
           associate(default_minima => [(0., i=1,num_outputs)], default_maxima => [(1., i=1,num_outputs)])
-            inference_engine%outputs_range_ = tensor_range_t("outputs", default_minima, default_maxima)
+            inference_engine%output_range_ = tensor_range_t("outputs", default_minima, default_maxima)
           end associate
         end associate
       end if
@@ -157,7 +162,7 @@ pure subroutine set_activation_strategy(inference_engine)
   module procedure construct_from_json
 
     type(string_t), allocatable :: lines(:), metadata(:)
-    type(tensor_range_t) inputs_range, outputs_range
+    type(tensor_range_t) input_range, output_range
     type(layer_t) hidden_layers, output_layer
     type(neuron_t) output_neuron
     real(rkind), allocatable :: hidden_weights(:,:,:)
@@ -189,9 +194,9 @@ pure subroutine set_activation_strategy(inference_engine)
 
     associate(prototype => tensor_range_t("",[0.],[1.]))
       associate(num_lines => size(prototype%to_json()))
-        inputs_range = tensor_range_t(lines(l:l+num_lines-1))
+        input_range = tensor_range_t(lines(l:l+num_lines-1))
         l = l + num_lines
-        outputs_range = tensor_range_t(lines(l:l+num_lines-1))
+        output_range = tensor_range_t(lines(l:l+num_lines-1))
         l = l + num_lines
       end associate
     end associate
@@ -342,7 +347,7 @@ function get_key_value(line) result(value_)
     character(len=17) :: single_value
     integer, parameter :: &
       outer_object_braces = 2, hidden_layer_outer_brackets = 2, lines_per_neuron = 4, inner_brackets_per_layer  = 2, &
-      output_layer_brackets = 2, metadata_outer_braces = 2, inputs_range_object = 5, outputs_range_object = 5
+      output_layer_brackets = 2, metadata_outer_braces = 2, input_range_object = 5, output_range_object = 5
 
     call assert_consistency(self)
 
@@ -360,7 +365,7 @@ function get_key_value(line) result(value_)
       associate(num_lines => &
         outer_object_braces &
         + metadata_outer_braces + size(key) &
-        + inputs_range_object + outputs_range_object &
+        + input_range_object + output_range_object &
         + hidden_layer_outer_brackets + (num_hidden_layers)*(inner_brackets_per_layer + neurons_per_layer*lines_per_neuron) &
         + output_layer_brackets + num_outputs*lines_per_neuron &
       )
@@ -392,24 +397,24 @@ function get_key_value(line) result(value_)
         lines(line) = string_t('    },')
 
         block
-          type(string_t), allocatable :: inputs_range_json(:), outputs_range_json(:)
+          type(string_t), allocatable :: input_range_json(:), output_range_json(:)
 
           line = line + 1
-          inputs_range_json = self%inputs_range_%to_json() 
-          associate(last_line => ubound(inputs_range_json,1))
-            call assert(last_line==inputs_range_object, "inference_engine_s(to_json): inputs_range object line count")
-            inputs_range_json(last_line) = inputs_range_json(last_line) // ","
-            lines(line:line+inputs_range_object-1) = inputs_range_json
-            line = line + inputs_range_object-1
+          input_range_json = self%input_range_%to_json() 
+          associate(last_line => ubound(input_range_json,1))
+            call assert(last_line==input_range_object, "inference_engine_s(to_json): input_range object line count")
+            input_range_json(last_line) = input_range_json(last_line) // ","
+            lines(line:line+input_range_object-1) = input_range_json
+            line = line + input_range_object-1
           end associate
 
           line = line + 1
-          outputs_range_json = self%outputs_range_%to_json() 
-          associate(last_line => ubound(outputs_range_json,1))
-            call assert(last_line==outputs_range_object, "inference_engine_s(to_json): outputs_range object line count")
-            outputs_range_json(last_line) = outputs_range_json(last_line) // ","
-            lines(line:line+outputs_range_object-1) = outputs_range_json
-            line = line + inputs_range_object-1
+          output_range_json = self%output_range_%to_json() 
+          associate(last_line => ubound(output_range_json,1))
+            call assert(last_line==output_range_object, "inference_engine_s(to_json): output_range object line count")
+            output_range_json(last_line) = output_range_json(last_line) // ","
+            lines(line:line+output_range_object-1) = output_range_json
+            line = line + input_range_object-1
           end associate
         end block
 

src/inference_engine/tensor_range_s.f90

@@ -73,23 +73,15 @@
 
   module procedure map_to_unit_range
     associate(tensor_values => tensor%values())
-      call assert(all(tensor_values>=self%minima_) .and. all(tensor_values<=self%maxima_), &
-        "tensor_range_s(map_to_unit_range): unnormalized range")
       associate(normalized_values => (tensor_values - self%minima_)/(self%maxima_ - self%minima_))
-        call assert(all(normalized_values>=0.) .and. all(normalized_values<=1.), &
-          "tensor_range_s(map_to_unit_range): normalized range")
         normalized_tensor = tensor_t(normalized_values)
       end associate
     end associate
   end procedure
 
   module procedure map_from_unit_range
     associate(tensor_values => tensor%values())
-      call assert(all(tensor_values>=0.).and.all(tensor_values<=1.), &
-        "tensor_range_s(map_from_unit_range): normalized input")
       associate(unnormalized_values => self%minima_ + tensor_values*(self%maxima_ - self%minima_))
-        call assert(all([unnormalized_values>=self%minima_, unnormalized_values<=self%maxima_]), &
-          "tensor_range_s(map_to_unit_range): unnormalized range")
         unnormalized_tensor = tensor_t(unnormalized_values)
       end associate
     end associate

src/inference_engine/trainable_engine_s.F90

@@ -68,7 +68,9 @@
 
       allocate(a(maxval(n), input_layer:output_layer)) ! Activations
 
-      a(1:n(input_layer),input_layer) = inputs%values()
+      associate(normalized_inputs => self%input_range_%map_to_unit_range(inputs))
+        a(1:n(input_layer),input_layer) = normalized_inputs%values()
+      end associate
 
       feed_forward: &
       do l = 1,output_layer
@@ -77,7 +79,9 @@
         )
       end do feed_forward
 
-      outputs = tensor_t(a(1:n(output_layer),output_layer))
+      associate(normalized_outputs => tensor_t(a(1:n(output_layer), output_layer)))
+        outputs = self%output_range_%map_from_unit_range(normalized_outputs)
+      end associate
 
     end associate