Converting frame.pixels to usable form is too slow in Python

[lydiatliu]

See for example the following line (86) from run_mission.py
#image = Image.frombytes('RGB', (frame.width, frame.height), str(bytearray(frame.pixels)) )

str(bytearray(frame.pixels)) by itself takes 0.056 seconds per call on average and takes up > 50% of my network training time. In particular, it seems that bytearray() is the slow operation.

Is there a way to get around this slowness in python if I still want to use Pillow for image processing? Thanks!

[lydiatliu]

@DaveyBiggers @timhutton

[DaveyBiggers]

I tried a few different methods for the ALE_HAC.py and found nothing that was particularly performant. We need to do some digging.

[DaveyBiggers]

The boost python vector_indexing_suite that we use to expose vectors as python objects apparently doesn't implement the buffer protocol, which is what things like numpy, PIL etc use for fast memory access (eg to get at the raw bytes without needing to copy).
Hopefully we can use something else to expose the unsigned char vector - if necessary we can create our own Python/C wrapper for that one object manually (without using boost python), though I'm sure there are easier ways...