Code is based on https://github.com/lisa-lab/pylearn2/blob/master/pylearn2/expr/normalize.py

Adds gaussian noise. Note that this is not entirely correct, the correct way would be to do it before downsampling, so the regular image and

Adds gaussian noise. Note that this is not entirely correct, the correct way would be to do it before downsampling, so the regular image and

new version, using Pool.imap instead of Pool.map, to avoid the data structure conversion from lists to numpy arrays afterwards.

by default, only the identity transform is in the augmentation list, so no augmentation occurs (only ds_transforms are applied).

The Krizhevskhy max pooling layer only supports square input. This function provides a workaround that uses Theano's own max pooling op, flan

sparsity: fraction of the weights to each output unit that should be nonzero

Generator that runs a slow source generator in a separate thread. buffer_size: the maximal number of items to pre-generate (length of the buf