* implemented CRF
* added test for CRF
* added a chunking example for CRF
* changed to use up2date tensorflow
* added conll2000 data
* minimize package dependency
* using logsumexp in keras instead
* Added improved_wgan.py to examples
Implements the improved WGAN with gradient penalty in Keras.
* Fixed PEP8 issues and one docstring mistake.
* Added a note to the discriminator to avoid BatchNormalization, as per the paper.
* Minor docstring correction to clarify description for wasserstein_loss function.
* Fixed a missing line in the generator
* Minor cleanup to make generating the label vectors a bit more elegant.
* Fixing typo in previous commit.
* Moved the model out of a main() function. Model now correctly respects
keras.backend.image_data_format
* Aaaand another PEP8 cleanup >.>
* Gradient penalty is now correctly based on the gradient w.r.t. the input averaged samples, rather than the weights of the discriminator.
* Gradient is now correctly calculated on the output of the discriminator, not the
K.mean() of the output (i.e. the loss).
* add srelu activation
* update srelu style
* make srelu and pelu work with any initializer
* add missing config line
* make initializations better match paper
It looks like in 1d57dd3c63 a spelling error ("kenel" not "kernal") was introduced into densenet.py that causes an argument error from keras when trying to use the network. This pull request corrects that.
* Bugfix for DenseNetFCN + Activation parameter + revert to Keras 1 model
* Correct ordering of parameters (input_shape) for DenseNet models
* Updated to Keras 2 API
* Changed image_dim_ordering to image_data_format
* Corrected "th" to channels_first and "tf" to channels_last
* Fixed " to '
* PEP 8 fixes
* Fix PEP 8 again
Frédéric Branchaud-Charron