Transformer+Semantic, Transformer+Visual, and Transformer+Semantic+Visual PredCls Not as Good as Before

[zhanwenchen]

GSC good BPL results were based on 0.04.

However, previous good results were based on LR=0.001.

So because of poor performance of transformers with LR=1e-3 (even though it's labeled as 1e-4) (#81) on Transformer baseline PredCls and Transformer+Semantic PredCls, I changed it to 0.16 (#115) and that bounced. Then I changed it to 0.016.

• It seems to do better than 1e-3 on Transformer baseline PredCls and Transformer+Semantic PredCls but not be as good for +visual PredCls and +semantic+visual PredCls. In fact, smaller is better for anything visual for PredCls?
• For SGCls, bigger (0.016) LR is better for Transformer Baseline SGCls, and better (but not great) Transformer+Semantic SGCls. +Visual is better but +Semantic+Visual is worse. Perhaps for SGCls, smaller is better is good for anything +semantic?
• For SGGen, bigger LR (0.016) is better for baseline but terrible for all other augmentations.

Instead of 0.016, try smaller LRs for +visual and +semantic+visual like 0.005 (5e-3) or 0.008.

For Motif and VCTree, LR wasn't the cause for the bad BPL results.

PredCls:

• +Semantic+Visual: 21.88 (2023-03-26182152_transformer_none_semantic_visual_predcls_4GPU_labx_1e-4,10000) => 19.21 (2023-04-10064353_transformer_none_semantic_visual_predcls_2GPU_lab1_0.016,50000)