* Skip newlines in TTY AND ipykernel.

After the addition of tty check output was ruined in jupyter. Go back to
the old ways when running jupyter notebooks and ipython.

* Trigger travis

* Only check for ipykernel once when initiating progbar

* Add colon

* Enable loading CuDNNLSTM weights into an LSTM layer

* Convert the kernels during weight loading

* Simplify reporthook-related logic

* More of the same

* Print time per step

* pep8

* Improve synchronized shuffle

* pep8

* Added TensorFlow Dataset API example.

* Made most recommended changes to dataset API example.

* Update cifar10.py

* Update cifar100.py

* Increase file save hash size

Issue #8249 runs into a problem where the max hash size is low, resulting in files being potentially overwritten.

The behavior of this functionality is as such:
As batch size increases, the probability that any given file will have a conflict in the hash reduces. However, when the author wants to save images with a batch size of one, the probability of a conflict increases since {index} is always 1. This seems to just be an unforeseen bug because a batch size of one is uncommon.

* Change hash size to 1e7

For example, if len(x) is 3 and the batch size is 2, 3//2 == 1 but there really is 2 batches.

* Add missing parameter name to K.identity()

* Remove unneeded parenthesis

Resolves https://github.com/fchollet/keras/issues/8232

* Simplify - remove K.cast()

* pep8

* Simplify, clarify, sync style examples/mnist_net2net.py

* Pass globals as arguments into functions

* Pass epochs as arguments into functions

* Remove default epochs value

* Add a model.check_trainable_weights_consistency

This will raise a UserWarning when the user modifies model.trainable
and tries to print a model summary or launch a fit without having
called .compile.

Calling .compile() is necessary because trainable weights are collected
in compile (model._collected_trainable_weights).

* Fix comments and cosmetics on count_params

* Make Model.check_trainable_weights_consistency private

Also fix its docstring

* Fix docstring of test_trainable_weights_count_consistency

* Expose and document utils.layer_utils.print_summary as utils.print_summary

Fixes #8105

* Fix docs autogen when running on python 3.6.2

When a class inherits from object without defining __init__,
AttributeError is raised instead of TypeError. So we need to catch
both exceptions to be compatible.

* Fix process_XXX_docstring functions to work with models function

The process_function_docstring would fail to parse the “# Argument” 
and “# Returns” of class functions. This was because it expected the
“# Argument” string to be preceeded by exactly a newline and 4 spaces,
but for class methods, that would be a newline and 8 spaces.

This fixes the problem by expecting a newline and some number of space.

* Quick fix cudnn rnns

* Tests passing.

* Normalize device names in multi_gpu_model (#8213)

* added _normalize_device_name function to training_utils.py --> convert "/device:GPU:0" to "/gpu:0"
* call _normalize_device_name() after obtaining available devices in multi_gpu_model() before verifying against target_devices.
* Successfully run tests in tests/keras/utils/multi_gpu_test.py

* Fix PEP8 violations in keras/utils/training_utils.py using autopep8

* Added support for passing external constants to RNN, which will pass them on to the cell

* Added class for allowing functional composition of RNN Cells, supporting constants

* put back accidentally commented out recurrent tests

* added basic example of functional cell

* new class AttentionRNN

* restored RNN layer

* renamed constants to attended in FunctionRNNCell, avoided duplicating outputs in wrapped model

* minor clean-up of docs

* Minor cleanup & improvments in docs, fixed PEP breaking formatting in attention test

* removed FunctionalRNNCell and AttentionRNN, added back support for constants in RNN

* fixed PEP8 violations

* fixed minor review comments

* added test case for when both inital_state and constants are passed to RNN.__call__

Rename random_seed => seed for consistency with the rest of the code. Maybe resolves https://github.com/fchollet/keras/issues/8227, although maybe GeneratorEnqueuer.seed should be eliminated altogether, because currently it's not used anywhere in Keras code.

* Fix samplewise normalization in ImageDataGenerator

Normalize pixels uniformly by the image mean and std

* dont specify axes for samplewise image normalization

* Sync Sequential.compile() with Model.compile()

* Remove loss_weights in Sequential.compile()

It may be confusing for some people for a model with only one output. If you need this feature, use functional Model API.

* Put Sequential.loss_weights attribute back. 

its removal breaks tests and stuff.

Same as https://github.com/fchollet/keras/commit/89e6eb01f200ef6fa8db926ecfee68b37b229fbc, caused by https://github.com/fchollet/keras/pull/8063

* add support for pandas DataFrame

* multiple updates according to @fchollet's review

* DataFrame should be handled correctly if list/dict is passed as model inputs/outputs

* Clarify docs for `shuffle` arg of `fit_generator`

* Clarify docs on `shuffle` arg of `model.fit_generator`

* Clarify docs for `shuffle` arg of `fit_generator`

* Clarify docs for `shuffle` arg of `fit_generator`