addition of maxabs

example/dense_mnist.f90

@@ -43,11 +43,17 @@ program dense_mnist
         net, validation_images, label_digits(validation_labels)) * 100, ' %'
 
     block
-      real :: output_metrics(10,2) ! 2 metrics; 1st is default loss function (quadratic), other is Pearson corr.
+      real, allocatable :: output_metrics(:,:) ! 2 metrics; 1st is default loss function (quadratic), other is Pearson corr.
       output_metrics = net % evaluate(validation_images, label_digits(validation_labels), metric=corr())
       print *, "Metrics: quadratic loss, Pearson corr.:", sum(output_metrics, 1) / size(output_metrics, 1)
     end block
 
+    block
+      real, allocatable :: output_metrics(:,:) ! 3 metrics; 1st is default loss function (quadratic), others are Pearson corr.
+      output_metrics = net % evaluate(validation_images, label_digits(validation_labels), metrics=[corr(), corr()])
+      print *, "Metrics: quadratic loss, Pearson corr.:", sum(output_metrics, 1) / size(output_metrics, 1)
+    end block
+
   end do epochs
 
 contains

src/nf.f90

@@ -5,7 +5,7 @@ module nf
   use nf_layer_constructors, only: &
     conv2d, dense, flatten, input, maxpool2d, reshape
   use nf_loss, only: mse, quadratic
-  use nf_metrics, only: corr
+  use nf_metrics, only: corr, maxabs
   use nf_network, only: network
   use nf_optimizers, only: sgd, rmsprop, adam, adagrad
   use nf_activation, only: activation_function, elu, exponential,  &

src/nf/nf_metrics.f90

@@ -7,6 +7,7 @@ module nf_metrics
   private
   public :: metric_type
   public :: corr
+  public :: maxabs
 
   type, abstract :: metric_type
   contains
@@ -27,6 +28,12 @@ end function metric_interface
     procedure, nopass :: eval => corr_eval
   end type corr
 
+  type, extends(metric_type) :: maxabs
+    !! Maximum absolute difference
+  contains
+    procedure, nopass :: eval => maxabs_eval
+  end type maxabs
+
   contains
 
   pure module function corr_eval(true, predicted) result(res)
@@ -37,7 +44,7 @@ pure module function corr_eval(true, predicted) result(res)
     real, intent(in) :: predicted(:)
       !! Values predicted by the network
     real :: res
-      !! Resulting loss value
+      !! Resulting correlation value
     real :: m_true, m_pred
 
     m_true = sum(true) / size(true)
@@ -48,4 +55,18 @@ pure module function corr_eval(true, predicted) result(res)
 
   end function corr_eval
 
-end module nf_metrics
+  pure function maxabs_eval(true, predicted) result(res)
+    !! Maximum absolute difference function:
+    !!
+    real, intent(in) :: true(:)
+      !! True values, i.e. labels from training datasets
+    real, intent(in) :: predicted(:)
+      !! Values predicted by the network
+    real :: res
+      !! Resulting maximum absolute difference value
+
+    res = maxval(abs(true - predicted))
+
+  end function maxabs_eval
+
+end module nf_metrics