load_model_tf with custom loss require the dict() function

[Cdk29]

I put this here if someone else run into the trouble, it is more problem of documentation.

I created a model with a custom loss function :

smoothie <- function(y_true,y_pred){
    tf$losses$categorical_crossentropy(y_true, y_pred, label_smoothing=0.3)
}

Which make me think that the categorical_crossentropy function of the package keras could be updated too, to reflect the one of the python version of keras, which can use label_smoothing.

Running the following would create an error :

model <- load_model_tf("../input/efficientnetb7-label-smoothing-imagenet/EfficientB7_image_net/")

ValueError: ('Unknown loss function', ':smoothie')

And this is when things began to be puzzling. Looking at keras-team/keras#5916 in the main keras, it appeared to me that I must use custom_objects. But the documentation is totally clueless about how and using standard R data-structure did not worked (I tried a list without success). Finally looking at the solution provided on github, and the code of the function of the github repository, I remembered that there is some conversion to the dictionary format via reticulate, and indeed there is a function to do it. Using :

model <- load_model_tf("../input/efficientnetb7-label-smoothing-imagenet/EfficientB7_image_net/", custom_objects = dict(smoothie = "smoothie"))

Worked perfectly.

I would suggest an update to the documentation or a way to pass a list as an argument, which the first logical attempt of a R user.