* Fix CSV formatting for windows with python 2

* Fix pep8 whitespace

* Fix quote style

* Update docs and add contributing page

* Add space for pep8

Remove excess preceding 4 spaces in Model.fit(..) docstring line
corresponding to verbose parameter.

* Update image.md

Enhancements for _flow_from_directory.  Classes and class_mode None

* Update image.md

Reworked it based on the comments

* Update image.md

* Update image.md

* Update image.md

typos

* Fix docstring

* add huber loss function (for robust regression)

* rename huber to logcosh (PR comments were correct), fix PEP8 whitespace checks

* logcosh loss: change from lambda to fn def'n, add text coverage

* Add a symbol to avoid indent

* Fix typo

* Add test for documentation

* Changes according to review

* Changes according to review

* Fix documentation and add Travis task

* Style fixes.

* Fix line length

* PEP8

Closes: #3688

* Added support for the new pydot API to fix find_graphviz error

* Simplified pydot installation checking

* Workaround for pydot generic Exception raising

* Removed hacky workaround for pyplot Exception, included comment

* Bug fix in recurrent layer

* Add test to recurent layer

* fix specify state

* Added documentation for `reset_states`

* Remove unneeded check

* Update Documentation

* pep8

* Fix when initial_states is a tensor

* modify tests for non-list initial states.

* use initial_state instead of initial_states

* pep8

* change get_initial_states to get_initial_state in ConvLSTM2D

* Check for Keras Tensors in Recurrent

* check if initial_state is passed to call

* pep8

* Move state_spec definition to __init__

* Fix reset states

* fix masking when specifying state

* added masking test for RNNs with specified state

* pep8

* remove unnecessary blank line

* Update layer_utils.py

Model is not sequential if there is a "merge" layer somewhere in the graph. So if a layer has multiple input layers ("inbound_layers"), the whole model is no longer sequential...

* Explanation of changed condition

Added a comment to explain the check for sequentiality in a model:
A model is not sequential if it has multiple nodes or if a layer has multiple inbound_layers

* Update the value of 'steps_per_epoch'

* Update the docstring of fit_generator to steps_per_epoch * batch_size

* Update the value of 'steps_per_epoch'

* Update the docstring of fit_generator: when 'steps_per_epoch' batches have been seen

* add warnings when advanced activations were passed into Activation

* fix import issue

* warning message beautify

* adopt user-friendly message

Updated the HDF5Matrix class to support inferred slice indeces such as data[:10] or data[19120:]. (#6299)