predict_classes() deprecated in R, what to use instead?

This is a convoluted way, but I managed to get it to work. I first had to get the array from the tf object, which turned out to be a python array, so then had to convert it to R object. And then I could compare/do calculations on that against another r object.

This is the prediction on the test data, output (prednn.tf) is a tensor:

prednn.tf <- modelnn %>% predict(x_test) %>% k_argmax() 

Now convert from tensor to an array using np_array:

prednn.array <- np_array(prednn.tf)

Now to a r-object:

prednn.array <- reticulate::py_to_r(prednn.array) 

Now compare the prediction against actual and calculate the accuracy:

accuracy(prednn.array, g_test) 

Note - following is the accuracy function:

accuracy <- function(pred, truth)
  mean(drop(pred) == drop(truth))

All this because predict_classes() is deprecated. There might be an easier way, but this worked for me.

You can use

cnn_pred <- cnn_model %>%
    predict(x_test) %>% k_argmax()

In my example, seems an even easier method was to do this:

change accuracy function from:

accuracy <- function(pred, truth)
  mean(drop(pred) == drop(truth))

to:

accuracy <- function(pred, truth)
  mean(drop(as.numeric(pred)) == drop(truth))

and then the following works:

modelnn %>% predict(x_test) %>% k_argmax() %>% accuracy(g_test)

so it seems, as.numeric() will take the tensor and extract the values. Hope this helps!