Merge pull request #102 from BerkeleyLab/simplify-app

.github/workflows/deploy-docs.yml

@@ -23,8 +23,6 @@ jobs:
       run: |
         PATH=/home/linuxbrew/.linuxbrew/bin/:"$PATH"
         PATH=`brew --prefix ford`/bin:"$PATH"
-        echo $PATH
-        ford --version
         ford ford.md
         cp ./README.md ./doc/html
     - name: Upload Documentation

app/train-cloud-microphysics.f90

@@ -18,7 +18,7 @@ program train_cloud_microphysics
   !! Internal dependencies;
   use inference_engine_m, only : &
     inference_engine_t, mini_batch_t, input_output_pair_t, tensor_t, trainable_engine_t, rkind, NetCDF_file_t, &
-    training_configuration_t
+    training_configuration_t, shuffle
   use ubounds_m, only : ubounds_t
   implicit none
 
@@ -235,7 +235,26 @@ subroutine read_train_write(training_configuration, base_name, plot_unit, previo
         else
           close(network_unit)
           print *,"Initializing a new network"
-          trainable_engine = perturbed_identity_network(training_configuration, perturbation_magnitude=0.05)
+          block
+            character(len=len('YYYYMMDD')) date
+
+            call date_and_time(date)
+
+            associate(activation => training_configuration%differentiable_activation_strategy())
+              associate( &
+                model_name => string_t("Thompson microphysics"), &
+                author => string_t("Inference Engine"), &
+                date_string => string_t(date), &
+                activation_name => activation%function_name(), &
+                residual_network => string_t(trim(merge("true ", "false", training_configuration%skip_connections()))) &
+              )
+                trainable_engine = trainable_engine_t( &
+                  training_configuration, metadata = [model_name, author, date_string, activation_name, residual_network], &
+                  perturbation_magnitude=0.05 &
+                )
+              end associate
+            end associate
+          end block
         end if
 
         if (.not. allocated(end_step)) end_step = t_end
@@ -339,77 +358,12 @@ subroutine read_train_write(training_configuration, base_name, plot_unit, previo
 
   end subroutine read_train_write
 
-  subroutine shuffle(pairs)
-    type(input_output_pair_t), intent(inout) :: pairs(:)
-    type(input_output_pair_t) temp
-    real harvest(2:size(pairs))
-    integer i, j
-
-    call random_init(image_distinct=.true., repeatable=.true.)
-    call random_number(harvest)
-
-    durstenfeld_shuffle: &
-    do i = size(pairs), 2, -1
-      j = 1 + int(harvest(i)*i)
-      temp     = pairs(i) 
-      pairs(i) = pairs(j)
-      pairs(j) = temp
-    end do durstenfeld_shuffle
-
-  end subroutine
-
   pure function normalize(x, x_min, x_max) result(x_normalized)
     real(rkind), intent(in) :: x(:,:,:,:), x_min, x_max
     real(rkind), allocatable :: x_normalized(:,:,:,:)
     call assert(x_min/=x_max, "train_cloud_microphysics(normaliz): x_min/=x_max")
     x_normalized = (x - x_min)/(x_max - x_min)
   end function
 
-  pure function e(j,n) result(unit_vector)
-    integer, intent(in) :: j, n
-    integer k
-    real, allocatable :: unit_vector(:)
-    unit_vector = real([(merge(1,0,j==k),k=1,n)])
-  end function
-
-  function perturbed_identity_network(training_configuration, perturbation_magnitude) result(trainable_engine)
-    type(training_configuration_t), intent(in) :: training_configuration
-    real(rkind), intent(in) :: perturbation_magnitude
-    type(trainable_engine_t) trainable_engine
-
-    ! local variables:
-    integer k, l
-    real, allocatable :: identity(:,:,:), w_harvest(:,:,:), b_harvest(:,:)
-    character(len=len('YYYMMDD')) date
-
-    call date_and_time(date)
-
-    associate(n=>training_configuration%nodes_per_layer(), activation=>training_configuration%differentiable_activation_strategy())
-      associate(n_max => maxval(n), layers => size(n))
-
-        identity = reshape( [( [(e(k,n_max), k=1,n_max)], l = 1, layers-1 )], [n_max, n_max, layers-1])
-        allocate(w_harvest, mold = identity)
-        allocate(b_harvest(size(identity,1), size(identity,3)))
-        call random_number(w_harvest)
-        call random_number(b_harvest)
-
-        associate( &
-          w => identity + perturbation_magnitude*(w_harvest-0.5)/0.5, &
-          b => perturbation_magnitude*(b_harvest-0.5)/0.5, &
-          activation_name => activation%function_name(), &
-          residual_network => string_t(trim(merge("true ", "false", training_configuration%skip_connections()))), &
-          model_name => string_t("Thompson microphysics"), &
-          author => string_t("Inference Engine"), &
-          date_string => string_t(date) &
-        )
-          trainable_engine = trainable_engine_t( &
-            nodes = n, weights = w, biases = b, differentiable_activation_strategy = activation, &
-            metadata = [model_name, author, date_string, activation_name, residual_network] &
-          )   
-        end associate
-      end associate
-    end associate
-  end function
-
 end program train_cloud_microphysics
 #endif // __INTEL_FORTRAN

example/fit-polynomials.f90

@@ -56,7 +56,7 @@ program train_polynomials
         integer e, b
         do e = 1,num_epochs
           call random_number(random_numbers)
-          call shuffle(input_output_pairs, random_numbers)
+          call shuffle(input_output_pairs)
           mini_batches = [(mini_batch_t(input_output_pairs(bins(b)%first():bins(b)%last())), b = 1, size(bins))]
           call trainable_engine%train(mini_batches, cost, adam=.true., learning_rate=1.5)
           print *,sum(cost)/size(cost)

example/learn-addition.f90

@@ -75,7 +75,7 @@ program train_polynomials
         integer e, b
         do e = 1,num_epochs
           call random_number(random_numbers)
-          call shuffle(input_output_pairs, random_numbers)
+          call shuffle(input_output_pairs)
           mini_batches = [(mini_batch_t(input_output_pairs(bins(b)%first():bins(b)%last())), b = 1, size(bins))]
           call trainable_engine%train(mini_batches, cost, adam=.true., learning_rate=1.5)
           print *,sum(cost)/size(cost)

example/learn-exponentiation.f90

@@ -75,7 +75,7 @@ program train_polynomials
         integer e, b
         do e = 1,num_epochs
           call random_number(random_numbers)
-          call shuffle(input_output_pairs, random_numbers)
+          call shuffle(input_output_pairs)
           mini_batches = [(mini_batch_t(input_output_pairs(bins(b)%first():bins(b)%last())), b = 1, size(bins))]
           call trainable_engine%train(mini_batches, cost, adam=.true., learning_rate=1.5)
           print *,sum(cost)/size(cost)

example/learn-microphysics-procedures.f90

@@ -84,7 +84,7 @@ program learn_microphysics_procedures
 
         do e = previous_epoch + 1, previous_epoch + max_num_epochs
           call random_number(random_numbers)
-          call shuffle(input_output_pairs, random_numbers)
+          call shuffle(input_output_pairs)
           mini_batches = [(mini_batch_t(input_output_pairs(bins(b)%first():bins(b)%last())), b = 1, size(bins))]
           call trainable_engine%train(mini_batches, cost, adam=.true., learning_rate=1.5)
           call system_clock(counter_end, clock_rate)

example/learn-multiplication.f90

@@ -75,7 +75,7 @@ program train_polynomials
         integer e, b
         do e = 1,num_epochs
           call random_number(random_numbers)
-          call shuffle(input_output_pairs, random_numbers)
+          call shuffle(input_output_pairs)
           mini_batches = [(mini_batch_t(input_output_pairs(bins(b)%first():bins(b)%last())), b = 1, size(bins))]
           call trainable_engine%train(mini_batches, cost, adam=.true., learning_rate=1.5)
           print *,sum(cost)/size(cost)

example/learn-power-series.f90

@@ -77,7 +77,7 @@ program train_polynomials
         integer e, b
         do e = 1,num_epochs
           call random_number(random_numbers)
-          call shuffle(input_output_pairs, random_numbers)
+          call shuffle(input_output_pairs)
           mini_batches = [(mini_batch_t(input_output_pairs(bins(b)%first():bins(b)%last())), b = 1, size(bins))]
           call trainable_engine%train(mini_batches, cost, adam=.true., learning_rate=1.5)
           print *,sum(cost)/size(cost)

example/learn-saturated-mixing-ratio.f90

@@ -83,7 +83,7 @@ program train_saturated_mixture_ratio
 
         do e = previous_epoch + 1, previous_epoch + max_num_epochs
           call random_number(random_numbers)
-          call shuffle(input_output_pairs, random_numbers)
+          call shuffle(input_output_pairs)
           mini_batches = [(mini_batch_t(input_output_pairs(bins(b)%first():bins(b)%last())), b = 1, size(bins))]
           call trainable_engine%train(mini_batches, cost, adam=.true., learning_rate=1.5)
           call system_clock(counter_end, clock_rate)

example/train-and-write.f90

@@ -63,7 +63,7 @@ program train_and_write
         integer e, b
         do e = 1,num_epochs
           call random_number(random_numbers)
-          call shuffle(input_output_pairs, random_numbers)
+          call shuffle(input_output_pairs)
           mini_batches = [(mini_batch_t(input_output_pairs(bins(b)%first():bins(b)%last())), b = 1, size(bins))]
           call trainable_engine%train(mini_batches, cost, adam=.true., learning_rate=1.5)
           print *,sum(cost)/size(cost)

src/inference_engine/input_output_pair_m.f90

@@ -41,10 +41,9 @@ elemental module function expected_outputs(self) result(my_expected_outputs)
       type(tensor_t) :: my_expected_outputs
     end function
 
-    pure module subroutine shuffle(pairs, random_numbers)
+    module subroutine shuffle(pairs)
       implicit none
       type(input_output_pair_t), intent(inout) :: pairs(:)
-      real, intent(in) :: random_numbers(2:)
     end subroutine
 
   end interface

src/inference_engine/input_output_pair_s.f90

@@ -21,14 +21,15 @@
 
   module procedure shuffle
     type(input_output_pair_t) temp
+    real harvest(2:size(pairs))
     integer i, j
 
-    call assert(size(random_numbers) >= size(pairs)-1, "input_output_pair_s(shuffle): size(random_numbers) >= size(pairs)-1")
+    call random_number(harvest)
 
     durstenfeld_shuffle: &
     do i = size(pairs), 2, -1
-      j = 1 + int(random_numbers(i)*i)
-      temp     = pairs(i)
+      j = 1 + int(harvest(i)*i)
+      temp     = pairs(i) 
       pairs(i) = pairs(j)
       pairs(j) = temp
     end do durstenfeld_shuffle

src/inference_engine/trainable_engine_m.f90

@@ -8,6 +8,7 @@ module trainable_engine_m
   use kind_parameters_m, only : rkind
   use tensor_m, only :  tensor_t
   use mini_batch_m, only : mini_batch_t
+  use training_configuration_m, only : training_configuration_t
   implicit none
 
   private
@@ -51,6 +52,14 @@ pure module function construct_from_inference_engine(inference_engine) result(tr
       type(trainable_engine_t) trainable_engine
     end function
 
+    module function perturbed_identity_network(training_configuration, perturbation_magnitude, metadata) result(trainable_engine)
+      implicit none
+      type(training_configuration_t), intent(in) :: training_configuration
+      type(string_t), intent(in) :: metadata(:)
+      real(rkind), intent(in) :: perturbation_magnitude
+      type(trainable_engine_t) trainable_engine
+    end function
+
   end interface
 
   interface

src/inference_engine/trainable_engine_s.f90

@@ -235,4 +235,41 @@
     inference_engine = inference_engine_t(metadata = self%metadata_, weights = self%w, biases = self%b, nodes = self%n)
   end procedure
 
+  module procedure perturbed_identity_network
+
+    integer k, l
+    real, allocatable :: identity(:,:,:), w_harvest(:,:,:), b_harvest(:,:)
+
+    associate(n=>training_configuration%nodes_per_layer())
+      associate(n_max => maxval(n), layers => size(n))
+
+        identity = reshape( [( [(e(k,n_max), k=1,n_max)], l = 1, layers-1 )], [n_max, n_max, layers-1])
+        allocate(w_harvest, mold = identity)
+        allocate(b_harvest(size(identity,1), size(identity,3)))
+        call random_number(w_harvest)
+        call random_number(b_harvest)
+
+        associate( &
+          w => identity + perturbation_magnitude*(w_harvest-0.5)/0.5, &
+          b => perturbation_magnitude*(b_harvest-0.5)/0.5, &
+          activation => training_configuration%differentiable_activation_strategy() &
+        )
+          trainable_engine = trainable_engine_t( &
+            nodes = n, weights = w, biases = b, differentiable_activation_strategy = activation, metadata = metadata &
+          )
+        end associate
+      end associate
+    end associate
+
+  contains
+
+    pure function e(j,n) result(unit_vector)
+      integer, intent(in) :: j, n
+      integer k
+      real, allocatable :: unit_vector(:)
+      unit_vector = real([(merge(1,0,j==k),k=1,n)])
+    end function
+
+  end procedure
+
 end submodule trainable_engine_s

src/inference_engine/training_configuration_m.f90

@@ -74,13 +74,13 @@ elemental module function learning_rate(self) result(rate)
       real(rkind) rate
     end function
 
-    module function differentiable_activation_strategy(self) result(strategy)
+    pure module function differentiable_activation_strategy(self) result(strategy)
       implicit none
       class(training_configuration_t), intent(in) :: self
       class(differentiable_activation_strategy_t), allocatable :: strategy
     end function
 
-    module function nodes_per_layer(self) result(nodes)
+    pure module function nodes_per_layer(self) result(nodes)
       implicit none
       class(training_configuration_t), intent(in) :: self
       integer, allocatable :: nodes(:)

test/trainable_engine_test_m.f90

@@ -390,7 +390,7 @@ function perturbed_identity_converges() result(test_passes)
     type(tensor_t), allocatable :: inputs(:)
     type(trainable_engine_t)  trainable_engine
     type(bin_t), allocatable :: bins(:)
-    real(rkind), allocatable :: cost(:), random_numbers(:)
+    real(rkind), allocatable :: cost(:)
     integer, allocatable :: neurons(:)
     integer, parameter :: num_pairs = 6
     integer, parameter :: num_epochs = 148
@@ -409,11 +409,8 @@ function perturbed_identity_converges() result(test_passes)
       end associate
       bins = [(bin_t(num_items=num_pairs, num_bins=num_bins, bin_number=bin), bin = 1, num_bins)]
 
-      allocate(random_numbers(2:size(input_output_pairs)))
-
       do epoch = 1,num_epochs
-        call random_number(random_numbers)
-        call shuffle(input_output_pairs, random_numbers)
+        call shuffle(input_output_pairs)
         mini_batches = [(mini_batch_t(input_output_pairs(bins(bin)%first():bins(bin)%last())), bin = 1, size(bins))]
         call trainable_engine%train(mini_batches, cost, adam=.true., learning_rate=1.5)
       end do