Add a switch for POST_NMS per batch/image during training

[rodrigoberriel]

This PR complements #676, adding the MODEL.RPN.FPN_POST_NMS_PER_BATCH switch:

• True (default): the code behaves AS IS, to keep it consistent with Detectron;
• False: POST_NMS is then applied per image, as "it should be":

maskrcnn-benchmark/maskrcnn_benchmark/modeling/rpn/inference.py

Lines 157 to 158 in 08fcf12

	# TODO resolve this difference and make it consistent. It should be per image,
	# and not per batch

I was wondering if I should add a comment on the README; something like this (changes in bold):

Furthermore, we set MODEL.RPN.FPN_POST_NMS_TOP_N_TRAIN 2000 as the proposals are selected for per the batch rather than per image in the default training. The value is calculated by 1000 x images-per-gpu. Here we have 2 images per GPU, therefore we set the number as 1000 x 2 = 2000. If we have 8 images per GPU, the value should be set as 8000. Note that this does not apply if MODEL.RPN.FPN_POST_NMS_PER_BATCH is set to False during training. See #672 for more details.

[fmassa]

LGTM, thanks a lot!

I think you can now change the README to take this new behavior into account, if you could send a PR it would be great!