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