* suppress serde warning

* add test

* pep8

* python2

* theano fix

* theano fix

* pop before call

* few fixes

* fix wrappers

* mooore fix

move self.l1 and self.l2 outside K.sum may increase the performance of calculating the norm.
In deep learning, the norm should not be very large (a large norm will make loss function useless), so move self.l1 and self.l2 outside K.sum will not cause overflow problem

* allow len1

* add test

* pep8

* rem unusedvar

* skip theano

* typo

* fix

* pep8

In certain use cases, load_weights() would cause OSError if called repeatedly on many different files. The proposed addition will now make sure the HDF5 files are closed before exiting.

* Added weighted metrics to save/load

* Patch to enable backwards-compatability with versions that did NOT save weighted_metrics

* Line wrap to fit PEP8 style

If building TF with XLA option enabled, an additional
`xla_gpu:0` device will be exposed to users which is a domain
for running graph in XLA mode. Should skip this device to
avoid the number of multi_gpu improperly inferred.

When we dilate a kernel of size N by D we take N - 1 steps of size D and finally sample 1 more pixel, hence (N - 1) * D + 1.

This reverts commit 08f6bdeb.

* Sync keras optimizers with tf.keras optimizers.

Changes:
1) epsilon has been removed from argument list, using default instead
   for future support of mixed precision. passing epsilon is still
   supported.
2) decay has been removed from argument list, for future support of
   learning rate decay objects. passing decay is still supported.
3) all lr has been changed to learning_rate. passing lr is still
   supported.
4) add initial_accumulator_value to Adagrad.
5) add momentum and centered to RMSprop.
6) Adam optimize won't have vhat if amsgrad is not present. set_weights
   is override to maintain backward compatibility.
7) Adadelta adds iterations to its weight list. set_weights is override
   to maintain backward compatibility.
8) Adagrad adds iterations to its weight list. set_weights is override
   to maintain backward compatibility.
9) RMSprop adds iterations to its weight list. set_weights is override
   to maintain backward compatibility.
10) Nadam changes default learning rate from 0.002 to 0.001
11) Adamax changes default learning rate from 0.002 to 0.001
12) Adadelta changes default learning rate from 1.0 to 0.001
13) Adagrad changes default learning rate from 0.01 to 0.001

* Slightly refactor SGD.

* Remove blank spaces.

* Fix some pep8 errors.

* Fix some error.

* Restore learning rate to previous default.

For Adadelta, Adamax, Adagrad

* Adjust some more lr.

* Nit

* Use layer.dtype as the default dtype when adding a new weight, and add dtype to network initialization.

* Fix CNTK test.

* don't reuse names in introductory example

I would try to be as explicit as possible in the first example of the functional API

* Update functional-api-guide.md

rename symbolic tensors to output

* Fixed plot_model for expand_nested=True

* Fixed plotting direct link from submodel to submodel

* Simplified expression

* add cnn seq2seq example

* PEP8 - Lines were too long

* Style fixes.

Noticed some changes that hasn't been applied to the ReadMe Doc. This change will refactor those artifacts.

__sudo: required__ no longer is...  [Travis are now recommending removing the __sudo__ tag](https://blog.travis-ci.com/2018-11-19-required-linux-infrastructure-migration).

"_If you currently specify __sudo:__ in your __.travis.yml__, we recommend removing that configuration_"

* Updated readme for windows users

* Fix phrasing