This layer dimshuffles 4D input for interoperability between C01B and BC01 ops. BC01 (theano) -> C01B (cuda convnet)

This layer dimshuffles 4D input for interoperability between C01B and BC01 ops. C01B (cuda convnet) -> BC01 (theano)

UGLY HACK: pool.imap does not allow for extra arguments to be passed to the called function. This is a problem because we want to pass i

This layer expects the layer before to have 37 linear outputs. These are grouped per question, clipped, but NOT normalised, because it seems

divisive normalisation, weights are considered constant when differentiating, optimised for performance.

Like Input2DLayer, but the data is expected to be in c01b order instead of bc01.

This layer expects the layer before to have 37 linear outputs. These are grouped per question, clipped, and then normalised by dividing by the su

Max pooling across feature maps. This can be used to implement maxout. This is similar to the FilterPoolingLayer, but this version uses a dif

OLD implementation using T.maximum iteratively. This turns out to be very slow. Max pooling across feature maps. This can be used to impleme

pools filter outputs from the previous layer. If the pooling function is 'max', the result is maxout. supported pooling function: - '

This layer expects the layer before to have 37 linear outputs. These are grouped per question and then passed through a softmax each, to enco

Global pooling across the entire feature map, useful in NINs.

UGLY HACK: pool.imap does not allow for extra arguments to be passed to the called function. This is a problem because we want to pass i

Same ugly hack as before

Same ugly hack as before

Same ugly hack as before

Same ugly hack as before

Same ugly hack as before

This layer merges featuremaps that were separated by the MultiRotSliceLayer and flattens them in one go.

This layer cuts 4 square-shaped parts of out of the input, rotates them 0, 90, 180 and 270 degrees respectively so they all have the same ori

divisive normalisation, optimised for performance.

This layer merges featuremaps that were separated by the Rot90SliceLayer and flattens them in one go.

This layer cuts 4 square-shaped parts of out of the input, rotates them 0, 90, 180 and 270 degrees respectively so they all have the same ori

divisive normalisation with softplus function, optimised for performance.

This layer expects the layer before to have 37 linear outputs. These are grouped per question, rectified, squared and then normalised by dividing

Stochastic pooling implemented in Theano using reshapes, since the Pylearn2 class for it is way too slow. This only works for c01b, i.e.

This layer expects the layer before to have 37 linear outputs. These are grouped per question and then passed through a softmax each, to enco