Adanet?

[jarlva]

Hi All, It seems that more and more, Tensorflow-derived projects could benefit from each other. For example, using the tensorflow/AdaNet AutoML to find the best TF model, In addition, ability to speed up training on colab TPU's by using the built in function to convert tf.keras models to TPU optimized ones. Would it be possible to convert the project to tf.keas to use with AdaNet to rip the above benefits?

[germain-hug]

Hi,
I don't have much experience with AdaNet, although looking at the tf.keras documentation it seems that converting the models alone should be fairly easy. I'm quite busy right now but if you want to give it a try for DDQN for instance go ahead! Would love to create a separate branch or such for this idea

[jarlva]

I already tried a simple test: inserting tensorflow in front the imports (from keras.XX to from tensorflow.keras.XX) but wishful thinking. I got esoteric API errors which I'm not familiar with, which indicate that the tensorflow.keras is not 1-to-1 with keras. I can give it a try and let you examine it when you have some time?