Add support for multichannel images

NVIDIA / DIGITS

Deep Learning GPU Training System

https://developer.nvidia.com/digits

BSD 3-Clause "New" or "Revised" License

4.12k stars 1.38k forks source link

Add support for multichannel images #600

Open IgorX2 opened 8 years ago

IgorX2 commented 8 years ago

By default, DIGITS converts RGBA images to RGB. It would be useful to add support for 4-channel PNG (or maybe TIFF) images. Most of the code for 4-channel PNG images is already there.

I tried with modifications to the image.py in utils, but the modifications need to be done in other places too. I can add this functionality if someone points me to the files that need to be modified for this to be done as simply as possible.

lukeyeager commented 8 years ago

Supporting 4-channel image data might not take too much work. One complication is that we'd need to disable JPEG encoding when saving to LMDBs. Unfortunately, I don't have time to scope it all out right now.

FYI, eventually we want to accept many more data types with something like https://github.com/NVIDIA/DIGITS/issues/197#issuecomment-136896978.

lukeyeager commented 7 years ago

@jmancewicz you were looking into this last week, right? Any updates?

jmancewicz commented 7 years ago

The change to image_to_array to also accept 4 channels was added with #1188