[MAINT] Tensorflow 2.0 compatibility

[rbavery]

When Tensorflow 2.0 is out of beta, solaris will need to deal with these parts of the code, which have been renamed. The deprecation warnings appear when using tensorflow 1.14 but they turn into errors with tensorflow 2.0 beta.

W0816 21:32:30.784060 140233658926912 deprecation_wrapper.py:119] From /usr/local/lib/python3.6/dist-packages/solaris/nets/metrics.py:103: The name tf.keras.metrics.cosine_proximity is deprecated. Please use tf.keras.losses.cosine_similarity instead.

W0816 21:32:32.026067 140233658926912 deprecation_wrapper.py:119] From /usr/local/lib/python3.6/dist-packages/solaris/nets/_keras_losses.py:104: The name tf.keras.losses.cosine is deprecated. Please use tf.keras.losses.cosine_similarity instead.

[nrweir]

Thanks for this and sorry for the late response - will make sure to update once we're moving beyond 1.1x. If this looks like a bigger concern, I'll either pin to <2.0 (would prefer not to do that) or catch the error and pass the correct metric. I expect it'll take a while to get everyone transitioned to 2.0+, so we'll probably want to support 1.13 for a while.

[nrweir]

I just dug into this a bit and it looks like a more robust set of metrics are present as classes with a __call__() method than in functional form (https://www.tensorflow.org/api_docs/python/tf/keras/metrics). We may be better off switching from using the functional forms to using these classes; if an enterprising user would be interested in creating a PR for this, it's available ;)