I got This error by upgrading to TensorFlow 2.1: "AttributeError: 'Operation' object has no attribute '_graph'" .

[mehrpad]

My code is written under TensorFlow 2.0 and I run it on Ubuntu 18.04.04. The code is perfectly run with TensorFlow 2.0. But I need to use TensorFlow 2.1 because of some functionality of TensorFlow_addones library. I upgrade the Tensorflow within my virtual environment by the command " conda install tensorflow-gpu" and run the code. The code run for the first epoch and save the result but as soon as it starts next epoch, it stop and show this error.
If I downgrade the TensorFlow to 2.0, again the code run without any problem. Does anybody know what is the issue?

[gadagashwini-zz]

@hu26hase, Can you share the standalone code to analyze the reported issue. Thanks!

[mehrpad]

I figured it out that this Error is gone if I am not saving the model after each epoch.
I used tf.save_model.save(model, path). If I comment it then, then it will no error.

Then I use model.save(path), which is keras version of the saving model, it has no error.

[gadagashwini-zz]

@hu26hase, Closing this issue since its resolved. Please feel free to comment or open if you still face issue.Thanks

[gavins13]

I've reproduced this issue. It is a bug in TensorFlow-Keras. Please re-open as a bug.

Issue: Model cannot be used to .predict or call. 'Operation' object has no attribute '_graph'. Happens if you save a model that you later want to extract layers from. E.g. keras.model(inputs=prev_model.inputs, outputs=prev_model.layers[-1].output)

[gadagashwini-zz]

@gavinlive, Please provide the complete standalone code to reproduce the reported issue. Thanks