deprecated call on predict_classes in R notebook R02

[EdwinB12]

preds_classes <- model %>%
  predict_classes(x = pred_images)
preds_classes

gives a depreciated call error.

Reported by a course participant and reproduced. Full traceback below:

Warning message in predict_classes(., x = pred_images):
“`predict_classes()` is deprecated and and was removed from tensorflow in version 2.6.
Please update your code:
  * If your model does multi-class classification:
    (e.g. if it uses a `softmax` last-layer activation).

      model %>% predict(x) %>% k_argmax()

  * if your model does binary classification
    (e.g. if it uses a `sigmoid` last-layer activation).

      model %>% predict(x) %>% `>`(0.5) %>% k_cast("int32")
”
Error in py_get_attr_impl(x, name, silent): AttributeError: 'Sequential' object has no attribute 'predict_classes'

Traceback:

1. model %>% predict_classes(x = pred_images)
2. predict_classes(., x = pred_images)
3. do.call(object$predict_classes, args)
4. object$predict_classes
5. `$.python.builtin.object`(object, "predict_classes")
6. py_get_attr_or_item(x, name, TRUE)
7. py_get_attr(x, name)
8. py_get_attr_impl(x, name, silent)

[ubdbra001]

Just had a look at this, the fix suggested by the error message works. But if you want to access the predicted classes as an array of integers (which the following cells suggest you do), then you'll want to add as.integer() as the final step in the pipe.

Happy to fix it, just want to make sure this is what you want it to do

[EdwinB12]

Yes fairly sure that is the desired behaviour @ubdbra001.

[ubdbra001]

Also need to update this text:

This prediction approach can be useful if you want to know the probability distribution for each prediction across all classes. However, if you are only interested in the predicted class, there is a shorthand keras function that can return predicted classes instead, predict_classes.

The code following it will be: preds_classes <- model %>% predict(x = pred_images) %>% k_argmax() %>% as.integer()