search

CaoWGG / TensorRT-CenterNet

tensorrt5 , centernet , centerface, deform conv, int8
767 stars 158 forks source link

the difference between before and after model trans. #73

Closed KevenLee closed 4 years ago

KevenLee commented 4 years ago

Suppose pytorch model as A , it's output as AC , tensorRT model as B , it's output as BC.

Then, I observed that there is a big gap between the output layers AC and BC , although the final detection result looks normal. For example, model A heatmap hm[i, j] = 0.75, model B heatmap hm[i, j] = 0.53, although the values of a and b are very different, they are both local maximums. So , detection results seems normal. What do you think of this problem? Thanks.

janismdhanbad commented 4 years ago

Hi, could you please let me know how you extracted hm values? Thanks

KevenLee commented 4 years ago

Hi, could you please let me know how you extracted hm values? Thanks

I‘m very sorry that, I do something wrong with feature extraction. I have fixed it

Jinreejing commented 2 years ago

with feature extraction. I ha

hello, could you please let me know is there a dcnv2 op in your model?

Jinreejing commented 2 years ago

with feature extraction. I ha

hello, could you please let me know is there a dcnv2 op in your model?

@KevenLee