Inconsistency in LR Schedule Between Paper and Code

[charon517-517]

Thank you for your excellent work! I’ve been working on reproducing the results of ReLA on the GRefCOCO validation set (SwinB version) and came across a discrepancy regarding the learning rate decay schedule.

According to the Supplementary Material for GRES https://openaccess.thecvf.com/content/CVPR2023/supplemental/Liu_GRES_Generalized_Referring_CVPR_2023_supplemental.pdf, the learning rate is initially set to 1e-5 and decreases by a factor of 10 at the 10,000th and 140,000th iterations. However, the default config file https://github.com/henghuiding/ReLA/blob/1cf1e9fe46cd69fea44e1cbe7917e599e895e438/configs/referring_R50.yaml#L56 specifies the first decay step at 110,000 iterations, which differs significantly from the paper.

Could you please clarify which learning rate schedule should be followed for the best results?

Thank you for your time and assistance!