Thread for minor results

[shinya7y]

I use this issue for posting minor results, including disappointing results.

[shinya7y]

GFLv2 on Manga109-s

settings:
GFLv1 from COCO
GFLv2 from COCO
UniverseNet-20.08 (GFLv1) from ImageNet
UniverseNet-20.08 +GFLv2 from ImageNet

AP AP50 AP75 APS APM APL body face frame text
0.673 0.906 0.715 0.179 0.389 0.744 0.617 0.457 0.922 0.694
0.672 0.904 0.713 0.180 0.390 0.743 0.618 0.454 0.918 0.699
0.689 0.922 0.733 0.199 0.426 0.758 0.643 0.476 0.930 0.707
0.689 0.923 0.734 0.198 0.434 0.758 0.645 0.475 0.930 0.706

[shinya7y]

Optimizers on Manga109-s

base config: universenet50_2008_fp16_4x4_mstrain_480_960_1x_manga109s.py

settings:

optimizer = dict(type='SGD', lr=0.08, momentum=0.9, weight_decay=0.0001)
optimizer = dict(_delete_=True, type='AdamW', lr=0.0001, betas=(0.9, 0.999), weight_decay=0.0001)
optimizer = dict(_delete_=True, type='AdamW', lr=0.0008, betas=(0.9, 0.999), weight_decay=0.0001)
optimizer = dict(_delete_=True, type='AdamW', lr=0.0008, betas=(0.9, 0.999), weight_decay=0.05)

AP AP50 AP75 APS APM APL body face frame text
0.699 0.925 0.743 0.205 0.436 0.771 0.666 0.480 0.937 0.712
0.691 0.921 0.732 0.198 0.425 0.763 0.654 0.468 0.932 0.708
0.701 0.925 0.746 0.205 0.442 0.769 0.664 0.483 0.939 0.717
0.701 0.926 0.745 0.210 0.441 0.770 0.663 0.487 0.939 0.714

[shinya7y]

Gradient accumulation on Manga109-s

settings:
4 GPUs * 4 samples_per_gpu (without gradient accumulation)
1 GPU * 4 samples_per_gpu * 4 cumulative_iters

AP AP50 AP75 APS APM APL body face frame text
0.699 0.925 0.743 0.205 0.436 0.771 0.666 0.480 0.937 0.712
0.698 0.924 0.745 0.199 0.442 0.769 0.667 0.480 0.937 0.709