Stochastic trainer with increasing minibatches

[accosmin]

Relevant papers:

• adaptive minibatch size: http://www.optimization-online.org/DB_FILE/2011/11/3226.pdf
• geometrically increasing minibatch size: http://arxiv.org/pdf/1104.2373v4.pdf

[accosmin]

New parameter for stochastic trainer ~ ratio:

• 1 [default] - fixed minibatch size of 8xT aka 8xlogical_cpus)
• R [1,16] - geometrically increasing from 8xT to min(Rx8xT, N/4) (N = size of training fold)

Need to adjust the epoch size accordingly.

[accosmin]

Removed the parameter as it is not improving the performance.