* Update TimeDistributed docs to be a little more clear

* Address PR Review

* Fix in_top_k() for Theano when identical values appear in predictions

* Add test and update docstrings for in_top_k()

* DOC: embeddings, fixed indentation

* DOC: embeddings, clarified input_dim size description

* Update embeddings.py

Rmsprop with default learning rate (0.001) cannot converge in this example. 
Initialize learning rate to (0.0001) and add weight decay fix the problem.

See https://gist.github.com/StanislavVolodarskiy/60c770d8f9864487692c88fe6faae892

* Embedding visualization is added to TensorBoard callback.

* CI failure fix.

* Code review fixes

+ None or empty list for embeddings_layer_names implies monitoring
  of all layers of type Embedding
+ embeddings_metadata now can contain just a string with metadata
  filename if it's common for all the embedding layers.
+ Frequencies now takes 0-th epoch as first.

* Code review is in progress

* Add `exclude_optimizer` argument to `save_model`

* Change `exclude_optimiser` to `include_optimizer`

Missing preprocessing_function

load_model fails when a model has multiple output layers that have more
than one metric. Solve this problem by adding a clause that checks if
metrics are a list.
For more elaborate description see issue #3958

Include a unit test confirming that model with multiple outputs that
have more than one metric can indeed be saved and reloaded.

* get_file() with tar, tgz, tar.bz, zip and sha256, resolves #5861.

The changes were designed to preserve backwards compatibility while adding support
for .tar.gz, .tgz, .tar.bz, and .zip files.
sha256 hash is now supported in addition to md5.

* get_file() improve large file performance #5861.

* getfile() extract parameter fix (#5861)

* extract_archive() py3 fix (#5861)

* get_file() tarfile fix (#5861)

* data_utils.py and data_utils_test.py updated based on review (#5861)
# This is a combination of 4 commits.
# The first commit's message is:
get_file() with tar, tgz, tar.bz, zip and sha256, resolves #5861.

The changes were designed to preserve backwards compatibility while adding support
for .tar.gz, .tgz, .tar.bz, and .zip files.
Adds extract_archive() and hash_file() functions.
sha256 hash is now supported in addition to md5.
adds data_utils_test.py to test new functionality

# This is the 2nd commit message:

extract_archive() redundant open (#5861)

# This is the 3rd commit message:

data_utils.py and data_utils_test.py updated based on review (#5861)
test creates its own tiny file to download and extract locally.
test covers md5 sha256 zip and tar
_hash_file() now private
_extract_archive() now private

# This is the 4th commit message:

data_utils.py and data_utils_test.py updated based on review (#5861)
test creates its own tiny file to download and extract locally.
test covers md5 sha256 zip and tar
_hash_file() now private
_extract_archive() now private

* data_utils.py and data_utils_test.py updated based on review (#5861)

* data_utils.py get_file() cache_dir docs (#5861)

* data_utils.py address docs comments (#5861)

* get_file() comment link, path, & typo fix

* Fix fit_generator docs for validation_steps

* Remove trailing whitespace for pep8

- TensorFlow 1
- Theano 0.9 : also use "device=cuda" in theanorc to use new
"gpuarray" backend
- Miniconda 4.2.12 (latest conda installer with python 3.5)
- Simplified pip install for tensorflow and keras test dependencies

* Fix docstring relating to stacked recurrent layers

The docstring did not specify the need to use return_sequences=True when creating a stacked recurrent network. I have replaced the original example with a more descriptive one.

* expand comment on LSTM example

Comment expanded to explicitly state that the input size only needs to be defined for the first layer.

* Update recurrent.py

* review the docs

* fix pep8 issues