hrnet / hrnet-fcos Goto Github PK

only a gpu is used, another one isn't used

The readme.md says we need to fix the image padding size when using SyncBN. But Since the process of calculating SyncBatchNorm is independent with the size of the image. And not fixing the padding size can speed the training up.
Does it because not fixing them will influence the last mAP?
Thanks a lot.

I create my own dataset as VOC-like format, and load noramlly, when I train the model for several epoch, error counters : 
  **File "HRNet-FCOS/maskrcnn_benchmark/structures/bounding_box.py", line 26, in __init__
 "bbox should have 2 dimensions, got {}".format(bbox.ndimension())**

ValueError: bbox should have 2 dimensions, got 1
I debug with bbox.size() and image_path, when this error appears, the result is torch.size([0]), but when I check the xml file in the image_path, the annotations is OK, so, how to fix this error, thanks a lot.
sincerely!

The given model weights can't be downloaded properly, would you mind sharing the Baiduyunpan
links ?

what is the differnece of hrnet.py between this repo and HRNet-MaskRCNN-Benchmark repo?