Sometimes IOU's are empty which breaks the code. about simple-faster-rcnn-pytorch

[sidml]

@cartvector You may try to reduce the anchor base size. It may fix the issue.
You may go to simple-faster-rcnn-pytorch-master\model\utils\bbox_tools.py and look for generate_anchor_base function. The default value for base_size is 16. You may want to reduce it to 4 or 8.

[joeybose]

This also corresponds with no anchors inside the image? I.e. inside_index is 0 in line 206 in creator_tool.py. Maybe for this special case clipping the anchors to image size is the way to go?

[chenyuntc]

Maybe adjust anchor size to fit your dataset? And could it be that the bbox instead of anchor is empty?

Btw, I suggest you to run it in IPython with %run. Then when it interrupt, you can go into the code that breaks with %debug.

[joeybose]

So I clipped the anchors whenever inside_index returns 0 and now it trains atleast. But I think this is a small bug regardless as if none of your anchors are inside the image something should be done.

[chenyuntc]

ious = bbox_iou(anchor, bbox)
argmax_ious = ious.argmax(axis=1)

I think it seems more likely that the bbox is mepty.

[cramaiah]

@joeybose How did you clip the anchors?

[joeybose]

It's been a while so I can't remember the exact solution anymore. But I have a fork of this repo which kinda works so maybe look there?

[cartvector]

@chenyuntc I debug this project in pycharm,this problem happened:
"ValueError: attempt to get argmax of an empty sequence"
in "ious = bbox_iou(anchor, bbox)
argmax_ious = ious.argmax(axis=1)
max_ious = ious[np.arange(len(inside_index)), argmax_ious]"
What can I do to solve this problem?

[cartvector]

it works！ thank you @sidml