reproduce work&Segmentation Loss

[MartinPR307]

Hello, I am very interested in your work on HybridNets. Recently, I have been trying to reproduce this work, but following the training stages outlined, I have not been able to achieve the results published in the paper. The results I obtained are as follows:

And these are the results I obtained after loading the publicly available Hybridnets.pth for validation：

I don't understand why there is such a large discrepancy. What confuses me even more is the segmentation loss. The segmentation loss from my training is 0.24399, while the segmentation loss from your publicly available Hybridnets.pth is 2.11897. Although my segmentation loss is lower, the performance on both segmentation tasks is worse. Additionally, during the initial training stage when I freeze the segmentation, the segmentation loss is around 2.25. Why is this happening?
Looking forward to your reply.

[MartinPR307]

@datvuthanh @xoiga123