Deep Residual Learning for CIFAR10 with keras and tensorflow, http://arxiv.org/abs/1512.03385

I have finish two versions of Res-56 in that paper
train_keras.py is the implementation with a prue keras, which has a val_acc of 0.9071
train_keras_and_tensorflow.py is the implementation with keras in architecture part and tensorflow in training part, which has a val_acc of 0.8892 

I think there should be some thing wrong in train_keras_and_tensorflow.py but I can't figure it out