Balanced accuracy values in tensorflow vs Caret/SKLEARN values

[DohaNaga]

Slightly different balanced accuracy values calculated within tensorflow than the ones calculated by the regular confusion matrix from the caret package. Possible cause is k_round.
The balanced accuracy used for training the models was calculated as follows:

Tensorflow

balanced_acc <- custom_metric("balanced_acc",function(y_true,y_pred){
  y_pred_pos = k_round(k_clip(y_pred, 0, 1))
  y_pred_neg = 1 - y_pred_pos
  
  y_pos = k_round(k_clip(y_true, 0, 1))
  y_neg = 1 - y_pos
  
  tp = k_sum(y_pos * y_pred_pos)
  tn = k_sum(y_neg * y_pred_neg)
  
  fp = k_sum(y_neg * y_pred_pos)
  fn = k_sum(y_pos * y_pred_neg)
  
  sensi = (tp/(tp + fn + k_epsilon()))
  specifi = (tn/(tn + fp + k_epsilon()))
  
  return((sensi + specifi )/ 2 )
  
})


##or 

balanced_acc <- custom_metric("balanced_acc",function(y_true,y_pred){

  tp = k_sum(k_round(k_clip(y_true * y_pred, 0, 1)))
  tn =  k_sum(k_round(k_clip((1 - y_true) * (1 - y_pred), 0, 1)))
  
  fp = k_sum(k_round(k_clip((1 - y_true) * y_pred, 0, 1)))
  fn = k_sum(k_round(k_clip(y_true * (1 - y_pred), 0, 1)))
  
  sensi = (tp/(tp + fn + k_epsilon()))
  specifi = (tn/(tn + fp + k_epsilon()))
  
  return((sensi + specifi )/ 2 )
  
})

The balanced accuracy used for evaluating the models was calculated as follows:

caret

  conf.matrix <-  caret::confusionMatrix(
    factor(pred_model, levels = 0:1),
    factor(y_test, levels = 0:1),
    positive = "1"
  )
  
  BalancedAccuracy <- conf.matrix$byClass[11]