RetineFace export to onnx has wired results

[jacqueline-weng]

I used the export_onnx.py script to transform the mxnet retina model to onnx. The onnx inference results are pretty wired. The confidence score for face box never exceeds 0.5. I am using onnx 1.5.0 and tensorrt 7.1.3 Following are some of the output results:

0.4027875 [ 0.00108999 0.04489411 -0.07810187 0.18495654 0.4027875 -0.15096149 -0.36747897 0.06386169 -0.3601119 -0.12819669 -0.07778522 -0.12063585 0.215042 0.0454467 0.22571692] 0.4105583 [ 0.00092068 -0.06066072 -0.08522762 0.22289114 0.4105583 -0.32271484 -0.44687378 -0.08884561 -0.39744833 -0.35895118 -0.1336861 -0.36705118 0.18532646 -0.18703444 0.22454621] 0.41528773 [ 0.00121907 0.06234938 -0.05923393 0.11546923 0.41528773 0.2724973 -0.3595028 0.44771698 -0.33609837 0.60410845 -0.02983215 0.26500404 0.26641363 0.38260424 0.29810804] 0.463866 [ 0.00109338 0.04545711 -0.0837696 0.15647165 0.463866 0.24408585 -0.37595078 0.41429448 -0.34585527 0.5555104 -0.05494411 0.21751963 0.23702072 0.33643827 0.2705367 ] 0.42746857 [ 0.00113435 0.01909502 -0.07017255 0.16814463 0.42746857 0.21636584 -0.3486702 0.40701234 -0.310448 0.5270049 -0.03101915 0.18199499 0.24490632 0.31391186 0.2837851 ] 0.40959314 [ 3.3570186e-04 -1.7408501e-02 -2.7924309e-02 1.9111551e-01 4.0959314e-01 -2.9661739e-01 -2.8980824e-01 -9.7867787e-02 -2.7061218e-01 -3.8381016e-01 4.2558089e-02 -2.8472298e-01 3.1775686e-01 -1.5051432e-01 3.2694706e-01] 0.40970972 [ 5.0232525e-04 1.1936285e-02 -5.0236136e-02 1.5036012e-01 4.0970972e-01 -2.6221037e-01 -3.2933250e-01 -6.3842371e-02 -3.0751497e-01 -3.3959898e-01 3.5137683e-04 -2.6569614e-01 2.8827393e-01 -1.3431580e-01 3.0214071e-01] 0.4259025 [ 0.00101786 0.03599988 -0.09109947 0.15826927 0.4259025 0.23911178 -0.35853773 0.40233937 -0.33194733 0.5406332 -0.04130673 0.20254968 0.22513627 0.31693083 0.2539889 ] 0.4221644 [ 0.00113109 0.03395934 -0.12123129 0.13252483 0.4221644 0.22949213 -0.36068708 0.39848888 -0.35619614 0.51799244 -0.0233437 0.18970846 0.22767223 0.3006508 0.23875538]