Open dmy1997 opened 8 months ago
Hi, This is an interesting finding. I was using Detectron2 interface when doing the training, you can see it in the config file, here, I used "freeze at: 1", this means the ResNet1 block should be fixed according to my understanding. Have you tried to re-train using my code and check the weights?
In the paper, R101 is frozen in stem and ResNet1 block. However, I print the weights and they are different from the in1k R101 model. Have they finetuned?