Question about pretrain checkpoint and its pred res

[GlennLoveZhen0124]

Hi, I have downloaded your trained model weights (Openlane benchmark), with backbone Efficientnet-B7. But when I load the weight and do inference on an openlane image, the output feature map - pred_seg_bev_map - always keep exactly the same, no matter what image I input to the model, Could you please help me to use it properly?

[RicardLee]

Hi, do you mean that the different inputs output the same feature map?

[GlennLoveZhen0124]

@RicardLee
yes, different image output same map with exactly same values in every locations, i check the first element, and the sum of the feature map elements, they always keep the same (the input tensor I checked are diffrenent for sure)
only pred_seg_bev_map keep the same all the time. other feature maps are OK

[RicardLee]

Hi, we reproduce this problem. In previous experiments, we don't evaluate the output of the seg_bev_map head. The loss of this branch doesn't converge and we believe that the mutil-task performance of PersFormer desires more exploration.

[GlennLoveZhen0124]

Thanks for your reply ~