Values in `.evaluate()`'s progress bar do not match final output about keras HOT 4 CLOSED

Actually the metric values shown in the progress bar are mean values, i.e. mean AUC of 32 steps.

from keras.

That metric in the progress bar is a training metric, computed on the training data. There are N different reasons a training metric might turn out different from the metric you see in a call to evaluate() (e.g. the state of the model changes at each step, and training behavior might be different from inference behavior).

If you want your displayed metrics to be equivalent to a call to evaluate() done at the end of the epoch, just pass validation_data=... in fit(). This will call evaluate() at the end of the epoch on the data you passed, and it will display the result in the progress bar under the name "val_auc".

from keras.

Actually the metric values shown in the progress bar are mean values, i.e. mean AUC of 32 steps.

But the mean of AUC over batches =/= Final AUC.

from keras.

That metric in the progress bar is a training metric, computed on the training data. There are N different reasons a training metric might turn out different from the metric you see in a call to evaluate() (e.g. the state of the model changes at each step, and training behavior might be different from inference behavior).

AFAIK, .evaluate just takes one dataset so it doesn't distinguish between train/test/val. I understand the progress bar I see during .fit reflects the changing state of the model, however the model weights should be frozen during evaluate , and the model should be in inference mode. So based on what you said I don't see why the metrics in .evaluate's progress bar shouldn't converge to the final metric.

Actually the metric values shown in the progress bar are mean values, i.e. mean AUC of 32 steps.

If this is true, and the progress bar shows a running average of the metric computed over batches, then it is understandable the answers are different. However I still think it's misleading and a better choice would be to show the result of the metric after incorporating each batch. (i.e. the progress bar should show metric.update_state(...).result() )

from keras.