[WIP] [Feature] YOLO v4

[RyanXLi]

This commit implements YOLO v4. This is a work-in-progress.

[codecov]

Codecov Report

Merging #3781 (c0e6e9e) into master (7a404a2) will increase coverage by 0.02%.
The diff coverage is 61.81%.

@@            Coverage Diff             @@
##           master    #3781      +/-   ##
==========================================
+ Coverage   61.07%   61.10%   +0.02%     
==========================================
  Files         217      215       -2     
  Lines       15377    15378       +1     
  Branches     2629     2608      -21     
==========================================
+ Hits         9392     9397       +5     
+ Misses       5515     5512       -3     
+ Partials      470      469       -1     

Flag	Coverage Δ
unittests	61.05% <61.81%> (+0.02%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
mmdet/core/__init__.py	100.00% <ø> (ø)
mmdet/core/bbox/coder/yolo_bbox_coder.py	57.89% <ø> (ø)
mmdet/datasets/builder.py	59.72% <ø> (ø)
mmdet/datasets/coco.py	45.63% <ø> (ø)
mmdet/models/dense_heads/rpn_test_mixin.py	83.87% <0.00%> (ø)
mmdet/ops/mish.py	0.00% <0.00%> (ø)
mmdet/models/dense_heads/yolo_head.py	33.50% <31.03%> (-0.21%)	⬇️
mmdet/models/necks/yolo_spp_pan_neck.py	59.25% <59.25%> (ø)
mmdet/models/necks/yolo_neck.py	62.29% <60.00%> (-1.17%)	⬇️
mmdet/utils/collect_env.py	40.00% <66.66%> (+17.08%)	⬆️
... and 30 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 7a404a2...be11113. Read the comment docs.

[callzhang]

Any updates on this feature? Any instruction on how to run would be great!

[PCRKTY]

@RyanXLi Is this done? How to use mosaic augmentation?

[fcakyon]

@RyanXLi is yolov4 support abandoned?

[RyanXLi]

@RyanXLi is yolov4 support abandoned?

Hi! Due to changes in my work this will not be actively maintained by me, but it may be updated by other members of the team in the future. Meanwhile, please feel free to check out other implementations, e.g. this one.

[tycoer]

d
Hi, have you trained 'yolov4_bce_giou_608_63e.py' ?
I get the log:
2021-12-27 13:56:03,213 - mmdet - INFO - Epoch [1][50/1833] lr: 1.220e-04, eta: 2 days, 12:35:00, time: 1.889, data_time: 0.956, memory: 23666, loss_cls: 382427.1763, loss_conf: 61.2742, loss_xy: 3.4589, loss_wh: 0.0000, loss: 382491.9084, grad_norm: 1943079.3325
2021-12-27 13:56:46,100 - mmdet - INFO - Epoch [1][100/1833] lr: 1.445e-04, eta: 1 day, 20:01:23, time: 0.858, data_time: 0.031, memory: 23666, loss_cls: 395967.3309, loss_conf: 65.1521, loss_xy: 3.4666, loss_wh: 0.0000, loss: 396035.9494, grad_norm: 1516712.2675
2021-12-27 13:57:29,117 - mmdet - INFO - Epoch [1][150/1833] lr: 1.671e-04, eta: 1 day, 14:31:22, time: 0.860, data_time: 0.032, memory: 23666, loss_cls: 353452.6903, loss_conf: 61.7409, loss_xy: 3.5157, loss_wh: 0.0000, loss: 353517.9522, grad_norm: 1460956.4975

The loss is very big and loss_wh is 0?

May I have your train log?