search

tsun / SSRT

Safe Self-Refinement for Transformer-based Domain Adaptation (CVPR 2022)
MIT License
39 stars 7 forks source link

Does the lambda function here multiply extra args.lr ? #5

Closed wjm-wjm closed 1 year ago

wjm-wjm commented 1 year ago

https://github.com/tsun/SSRT/blob/0bf39fd188d5f1ce12785ea94ae737eb55c3416a/trainer/train.py#L191

wjm-wjm commented 1 year ago
image

lambda function seems to a multiplicative factor for args.lr, so may not multiply with args.lr ?

tsun commented 1 year ago

Hi, after this line https://github.com/tsun/SSRT/blob/0bf39fd188d5f1ce12785ea94ae737eb55c3416a/trainer/train.py#L190 the lr of each parameter group is in fact 0.1/1.0, not args.lr. Therefore, args.lr in Lambda function is necessary. image

wjm-wjm commented 1 year ago

Thanks a lot! I didn't notice this self.parameter_ list function before. image