Open Cascol-Chen opened 1 year ago
I think a pretrained vision transformer trained while stochastic depth is on would help.
w?hat do you mean by the concept of stochastic depth?
By stochastic depth, I mean the drop path rate in vision transformer. A model with stochastic depth could be trained using the config upernet_swin.py. However, since the training and the test time adaptation doesn't share a same mmseg repository, it is hard to trained a network in SHIFT-TTA-train_source_model and apply adaptation on the network in SHIFT-Continuous_Test_Time_Adaptation.
We want to verify our algorithm on shift segmentation tta. However, our algorithm requires stochastic depth while training and the codebase seems complicated. Could you kindly show us how and where can I easiest modify the train script and the model structure so that the requirements could be fullfilled.