add 1 gamma parameter to grid search
script now supports mutliple kernels instead of only chi2

Adapt former script for benchmarking deepstrom classification in the end to end setting to keras (from tensorflow)
Create lazyfile for this experiment

No more horizontal flip for mnist and svhn datasets (horizontal flips on number = not good)
Now gamma is computed for deepfriedconvnet (when it was only done for deepstrom with rbf in the former version)

init classif_end_to_end_subsample_conv_hand_with_augment graph drawing + init classif_end_to_end_with_augment + add some tensorboard logging in end_to_end_with_2_layers_only_dense_with_augment

exp end_to_end_subsample_conv_hand_steph_model maj du fichier de log tensorboard + do not recompute repr after training

Transfert learning with batchnorm expe
End to end expe