David Warde-Farley

I'd recommend looking at the implementation of the ilsvrc2010 converter, somehow that thing works. Also, I've found the throughput when working with JPEG-in-HDF5 to be... disappointing. I haven't had time...

One idea: is it possible to update an external shared variable from within the scan inner graph? You could have an update that stashes the batch statistics in an external...

This would be a useful thing to have in Theano regardless, though I don't know how you'd get the JPEG data on the GPU. I suppose you could cast the...

Note for myself: individual decode times are about halved with jpeg4py. Using `concurrent.futures.ThreadPoolExecutor` on Python 3 to map across threads brings it down to a negligible 50-100ms for a batch...

Not without loss, but it'll be a small enough difference to probably not matter.

You already won't get identical results depending on your JPEG decoder, I doubt re-saving with PIL will be a big deal.

There was talk of using Blocks Aggregators for this. it's a good idea, though we should think about how to store this information in a standardized way. @rizar thoughts? On...

@fvisin I think we should split this ticket up. We should try and tackle the offline preprocessor first, as it is potentially useful without the other two pieces. My next...

It seems like the `child_epoch_iterator` ought to be owned by the iterator, in order to at least make it possible for the data stream to remain stateless (there's no contract...

@rizar I suppose it's a question of whether we want the stream object or the iterator object to be stateful, or both. Often, the situation is straightforward enough that the...