GitLab

* Set use_bias=False for conv layers in Keras ResNet model

* Removed bias regularizer from Conv2D layers (which should have no effect after bias is removed)

* Setting default data format based on available devices.

* Handle data format in Keras ResNet model properly for better performance on GPU; Compile only the training graph when skip_eval flag is True

* Added data format fix to Keras Cifar model; Removed unnecessary import

* Add a comment to the skip_eval flag per Priya's request

- kears_cifar_main returns dict of stats after run.
- unit tests added for keras_common.
- lint issues fixed in all 4 files.

 - Average of 5 runs single gpu is 93%+ range 92.5 to 93.4 top_1