Fix the error when save adv model and use modelcheckpoint callback

[review-notebook-app[bot]]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[google-cla[bot]]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

For more information, open the CLA check for this pull request.

[csferng]

@wangbingnan136 , could you take a look and sign the Contributor License Agreement in the above comment?

[wangbingnan136]

@wangbingnan136，你能看一下并在上面的评论中签署贡献者许可协议吗？

It looks like I have already signed

[wangbingnan136]

I think the checks were done,sorry this is my first time make a contribution to tensorflow~

[wangbingnan136]

ok,I got it!

[wangbingnan136]

It is done

[csferng]

Thanks for the fix! Let me polish the example notebook a bit before merging the PR.

[wangbingnan136]

Okk

[wangbingnan136]

I found that when we load the model，the "fit" and "predict" methods works as expected.But the "evalaute" has some bugs.I will fix this problem when I am not busy.

[csferng]

Sorry for the delay. I made some updates on the notebook and will proceed to merge this PR.

Regarding the bug on evaluate, what I encountered is that the accuracy number was off for the loaded model (0.09 vs. 0.98). The root cause turns out to be that metrics=['acc'] is interpreted differently. (Keras has some internal heuristic to distinguish what "accuracy" means to a model.) What we expect is tf.keras.metrics.SparseCategoricalAccuracy, which is the case for the adv_model. But the loaded model somehow got tf.keras.metrics.CategoricalAccuracy, so the evaluation result became different. To workaround this issue, I changed metrics=['acc'] to metrics=[tf.keras.metrics.SparseCategoricalAccuracy()], which then produced consistent result before and after saving/loading.

[wangbingnan136]

Sorry for the delay. I made some updates on the notebook and will proceed to merge this PR.

Regarding the bug on evaluate, what I encountered is that the accuracy number was off for the loaded model (0.09 vs. 0.98). The root cause turns out to be that metrics=['acc'] is interpreted differently. (Keras has some internal heuristic to distinguish what "accuracy" means to a model.) What we expect is tf.keras.metrics.SparseCategoricalAccuracy, which is the case for the adv_model. But the loaded model somehow got tf.keras.metrics.CategoricalAccuracy, so the evaluation result became different. To workaround this issue, I changed metrics=['acc'] to metrics=[tf.keras.metrics.SparseCategoricalAccuracy()], which then produced consistent result before and after saving/loading.

I see,great job!