Keras correlation coefficient as metric in R

Question

Does anyone know how to use the correlation coefficient or squared correlation coefficient as a metric in keras in R (although other languages may provide clues). This is for a CNN that functions similar to an image compressor/decompressor. Thank you!

Answer 1

Also struggling with the same problem. Here's my go at a custom metric using keras backend functions.

  metric_correlation <- 
custom_metric("correlation", 
              function(y_true, y_pred){
                k_sqrt(k_sum((y_pred-k_mean(y_pred))*(y_true-k_mean(y_true)))/length(y_true))/
                  k_sqrt(k_var(y_true))/k_sqrt(k_var(y_pred))
              })
                                          
  model %>% compile(loss = 'mean_squared_error', optimizer = 'adam',metrics = metric_correlation)

Something is amiss though - correlation drops as the epochs increase