Implement BPL - Githubissues

zhanwenchen / relaug

MIT License

0 stars 0 forks source link

Implement BPL #114

Closed zhanwenchen closed 1 year ago

zhanwenchen commented 1 year ago

What about

[ ] Base_LR
[ ] dtype matching the original trained model?
[ ] Maybe we shouldn't change the float vs long in model_motifs.py.

zhanwenchen commented 1 year ago

Is mR @ 50 supposed to drop in the first few epochs of BPL? Confirm with GSC. In GSC (45934586_motif_baseline_predcls_4GPU_riv_1_predcov_bottomk_0016000_bpl_sa):

went up from 22 to 32 right away
base_lr is 0.001 but rl_factor (present in relation_train_net which makes effective lr 0.04
dtype is float32 but original was float32 too I think
also used both relaug methods

zhanwenchen commented 1 year ago

The issue was the missing code changes in VGDataset