Using iter_size while manually feeding data into the network

[kareem ahmed]

I am currently training a triplet network, where each batch consists of 3 images, two of which are similar with the third being dissimilar. I have 2 different classes, and have therefore reasoned that it would be much easier to read lmdbs of each of the classes to create batches in python on the fly, as opposed to having to create one humongous lmdb containing all the triplet batches. I wish to utilise the iter_size option, but found that such a thing was not trivial given the manner in which the data is fed. My question is, does there exist an easy way through which this can be achieved? I had though about feeding iter_size batches into the network, calling net.forward_backward_all(), accumulating the diffs, and then calling solver.apply_update(), would that be the correct way to do it?