onnx tensorrt model infer results are different in some special shapes.

[congjianting]

Hi,

i found two weird things in trt model inferencing process. Any one knows why?

1. "dict(type='Pad', size_divisor=32)", is not work in trt model, it can lead different results(eg. 10 pixels shifting in det box-x.) between onnx and trt. but when i used "dict(type='Pad', size_divisor=128)," the result is same with onnx result, the shifting is gone.

2. when i trained a model only 1 fg class num, i found
   "rcnn=dict(
   score_thr=0.3," ....)
   is not working, i debuged and found some zeros det boxes, their scores is near to 0.5.

eg.
tensor([[[3.2900e+02, 1.1648e+02, 7.1000e+02, 1.5277e+02, 9.9482e-01],
[0.0000e+00, 0.0000e+00, 0.0000e+00, 0.0000e+00, 5.0489e-01], ... # this score is weird.

[congjianting]

by the way, the model structure is cascade-res50-fpn-maskrcnn