How to get a smooth ROC curve from CNN with keras functional API?

Question

I use the following code:

fpr, tpr, t = roc_curve(true_categories, predicted_categories)

I know the predicted_categories should be probability or confidence.

But as predict_proba() seems not working for keras Functional API, how should I get the predicted_categories right?

I tried

predicted_categories = tf.argmax(y_pred, axis=1)

and the ROC looks like this:

ROC with a sharp angle

and this

predicted_categories =  = tf.reduce_max(y_pred, axis=1, keepdims=True)

the ROC looks like this:

an ROC with AUC ~50%, which should be 85%

Both of them are not what I expected...

Any suggestion is appreciated!