Question about the input of ROIAlign

[hangzhiyiwei]

Hi

I have a question about the input image feature size here
https://github.com/NVlabs/Bongard-HOI/blob/master/models/rn_bbox_encoder.py#L116

The input feature size for one image is 25611 (depthheightwidth). If so, how can ROIAlign crop proper ROI region feature? I have seen the output of this self.roi_pooler (with size 25677), and I found that many roi_feats are all zero outputs.
For others, for each channel of the whole 256 channels, only the upper left element has a number. An example shows below.
[ 0.4648, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000, 0.0000]

I am really confused about it. Sorry to bother. Any reply will be appreciated. Thanks very much.