Javadoc
A preprocessor to allow RNN and CNN layers to be used together
For example, time series (video) input -> ConvolutionLayer, or conceivable GravesLSTM -> ConvolutionLayer
Functionally equivalent to combining RnnToFeedForwardPreProcessor + FeedForwardToCnnPreProcessor
Specifically, this does two things:
(a) Reshape 3d activations out of RNN layer, with shape [miniBatchSize, numChannels*inputHeight*inputWidth, timeSeriesLength])
into 4d (CNN) activations (with shape [numExamples*timeSeriesLength, numChannels, inputWidth, inputHeight])
(b) Reshapes 4d epsilons (weights.*deltas) out of CNN layer (with shape
[numExamples*timeSeriesLength, numChannels, inputHeight, inputWidth]) into 3d epsilons with shape
[miniBatchSize, numChannels*inputHeight*inputWidth, timeSeriesLength] suitable to feed into CNN layers.
Note: numChannels is equivalent to depth or featureMaps referenced in different literature