w/o & w dcn, the network have different trained results

Epiphqny / VisTR

[CVPR2021 Oral] End-to-End Video Instance Segmentation with Transformers

https://arxiv.org/abs/2011.14503

Apache License 2.0

738 stars 96 forks source link

Closed huapohen closed 3 years ago

huapohen commented 3 years ago

When I remove the dcn layer, the loss descend slowly

huapohen commented 3 years ago

Thank you for sharing this great work!

huapohen commented 3 years ago

I might understand