Couldn‘t reproduce the results in the paper

[Dopamine0717]

Following readme construction, I only got AP 23.5 after training on Seed 1 Percent 1 Coco standard while 25.81 ± 0.28 is claimed in the paper. Here's my setting:

Dataset:
percent 1 seed 1 Coco standard

Pretrained baseline:
I just changed the number of gpu in the train_gpu2.sh to get the train_gpu4.sh and only got AP 12.3.

Training:
I use train_gpu4.sh with pretrained baseline and got AP 23.5

Config:
labelmatch_standard.py
samples_per_gpu = 1
total_iter = 250000
update_interval = 1000
test_interval = 2000
learning_rate = 0.005
lr_config = dict(policy='step',warmup='linear',warmup_iters=4000,warmup_ratio=0.001,step=[140000,200000])

Environment:
I used two envs, one with Python = 3.6, mmdet=2.10.0, pytorch=1.6.0, and another with Python = 3.9, mmdet = 2.25, torch = 1.12. Both got the same results.

Could you share your model weight and log of Label Match training to help use reproduce results?

[AceCoooool]

The log file 1-1-40k is uploaded. (It's the setting in the ablation study. The final results use longer iterations and stronger resize)