Domino Valdano created MADLIB-1450:
--------------------------------------
Summary: Remove dependency on keras
Key: MADLIB-1450
URL: https://issues.apache.org/jira/browse/MADLIB-1450
Project: Apache MADlib
Issue Type: Improvement
Components: Deep Learning
Reporter: Domino Valdano
Fix For: v1.18.0
Presently, we require that users install both standalone "multi-backend"
`keras` and `tensorflow` python modules in order to use the DL features of
madlib.
But there is no good reason for the keras requirement, as keras has come
included with tensorflow since 2017, starting with version `1.4.0`:
[https://github.com/tensorflow/tensorflow/tree/v1.4.0/tensorflow/python/keras]
In fact, the multi-backend version of keras is a dead project, no longer
maintained by anyone.
As it says on [https://github.com/keras-team/keras]:
*Multi-backend Keras has been discontinued. At this time, we recommend that
Keras users who use multi-backend Keras with the TensorFlow backend switch to
{{tf.keras}}*
Keras 2.2.5 was the last release to support backends other than tensorflow, now
it is just maintained as a part of the tensorflow module.
Because madlib 1.17.0 does not import `tensorflow` directly, users still must
have the old multi-backend keras module installed in order to use our DL
features.
In order to fix this, dropping the keras dependency, in 1.18.0 we should switch
from using:
`import keras`
`from keras.backend import tf`
to:
`import tensorflow as tf`
`from tensorflow.python import keras`
We can continue using both `tf` and `keras` as before, without touching
anything but the import statements.
Another motivation for this change is that `keras 2.2.5` no longer includes
`keras.tf`, so our current method of importing only works up through `keras
2.2.4`.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
