Keras tensorflow : Get predictions and their associated ground truth labels after model.evaluate() or model.predict()

I am new to Tensorflow and Keras so the answer is perhaps simple, but I have a batched and prefetched tensorflow dataset (of type tf.data.TFRecordDataset) which consists in images and their label (int type) , and I apply a classification model on it.

        `y_pred = model.predict(tf_test_dataset)`

I would like to compare classification results from ground truth for each image and put this in a dictionary. To do that I tried to create this dictionary with y_pred for each image, and its Ground truth label :

        iterator = iter(tf_test_dataset)
        result_dict = {}
        for image_number in range(len(y_pred)):
            result_dict[image_number]= \
                {
                "Score": y_pred[i],
                "GroundTruth": iterator.get_next()[1].numpy()
            }

But I am not sure if it is the right way to get predictions and their associated ground truth label. (Is it easier using model.evaluate ?). Moreover when I do

    iterator = iter(tf_test_dataset)
    print(iterator.get_next()[1].numpy())

and

    for i in tf_test_dataset:
       print(i[1].numpy())

the results (the lists with the labels) are different...

Thanks for your help

Start with the shown tf_test_dataset iteration and model.predict() calls; determine why labels differ between iterator.get_next() and the for loop, and how predictions should be paired with labels. Done when the issue explains a reproducible way to align each prediction with its ground-truth label, including whether model.evaluate() addresses that need.