Closed jaymefosa closed 9 months ago
https://github.com/HyelinNAM/ContrastiveDenoisingScore/blob/310b15d96f7583623d1e8f3e511fcb37cbf6c969/pipeline_cds.py#L141C45-L141C45
It kind of reminds me of the orthogonal delta rotation network that sets a super high learning rate. Was there any source/reason for this choice?
Hello @jaymefosa
Most of the configs include learning rate are inherited from official implementation of DDS. (https://github.com/google/prompt-to-prompt/blob/main/DDS_zeroshot.ipynb)
https://github.com/HyelinNAM/ContrastiveDenoisingScore/blob/310b15d96f7583623d1e8f3e511fcb37cbf6c969/pipeline_cds.py#L141C45-L141C45
It kind of reminds me of the orthogonal delta rotation network that sets a super high learning rate. Was there any source/reason for this choice?