tensorflow function predict_classes() removed in 2.6

[Astrocito98]

Hi, thanks for putting this up, im using this code for my MD project. Everything working well until the moment where i try to create a behavior classifier. Seems like the function predict_classes() has been removed from tensorflow in v2.6. When trying to call the function ClassifyBehaviors it shows the error

AttributeError: 'Sequential' object has no attribute 'predict_classes'

and points what code to use instead. When I try to correct the code on my own the function looks like this

ClassifyBehaviors <- function(t,model, model_parameters){
if(!IsTrackingData(t)){
stop("Object is not of type TrackingData")
}
t <- CreateTestSet(t, model_parameters$integration_period)
t$labels$classifications <- model %>% predict(t$train_x) %>% k_argmax()``
###before "t$labels$classifications <- model %>% predict_classes(t$train_x)###
t$labels$classifications <- c(rep(NA,model_parameters$integration_period), model_parameters$Feature_names[t$labels$classifications + 1], rep(NA,model_parameters$integration_period))
return(t)
}

That way it allows it to pass that last error, but then it shows

Error in model_parameters$Feature_names[t$labels$classifications + 1] :
invalid subscript type 'environment'
``

I'm stuck in trying to correct this error, hope you can update the code so it works with TF v2.6+

[lukasvonziegler]

Hi. Yes, unfortunately updates happen and this package here is not intended to be updated. this is, because it is also a repository that comes with a publication and should contain all the code that can accurately reproduce our results. However on my side i already fixed this issue for my own code going forward. Can you quickly try the updated function for ClassifyBehaviors and tell me if it works for you? I may fork this package in the future and maintain/update the forked version.

ClassifyBehaviors <- function(t,model, model_parameters){
if(!IsTrackingData(t)){
stop("Object is not of type TrackingData")
}
t <- CreateTestSet(t, model_parameters$integration_period)
t$labels$classifications <- model %>% predict(t$train_x)
t$labels$classifications <- c(rep(NA,model_parameters$integration_period),
model_parameters$Feature_names[apply(t$labels$classifications,MARGIN = 1, FUN = which.max)],
rep(NA,model_parameters$integration_period))
return(t)
}

[Astrocito98]

Hi, thanks for your answer. I first solved the problem by downgrading my TF version to v2.5. Ill problably keep working like this for my current project for consistency. However, I quickly tried with your upgraded version of the code and worked just fine. Havent find any other problem using the package.