Evolve missing dfl hyperparameter in train_dual.py and train.py

[mkrupczak3]

Running training with the --evolve flag, the system outputs the following error:

Traceback (most recent call last):
  File "/hpcuser/yolov9/train_dual.py", line 644, in <module>
    main(opt)
  File "/hpcuser/yolov9/train_dual.py", line 604, in main
    g = np.array([meta[k][0] for k in hyp.keys()])  # gains 0-1
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/hpcuser/yolov9/train_dual.py", line 604, in <listcomp>
    g = np.array([meta[k][0] for k in hyp.keys()])  # gains 0-1
                  ~~~~^^^
KeyError: 'dfl'

The root cause is that the dfl hyperparameter is absent in the evolve section of train.py and train_dual.py.

In the file hyp.scratch-high.yaml the dfl hyperparameter appears just after obj_pw:

yolov9/data/hyps/hyp.scratch-high.yaml

Line 13 in 3f27231

dfl: 1.5 # dfl loss gain

However, in both train_dual.py and train.py this hyperparameter is missing under the keys for evolve:

yolov9/train_dual.py

Line 556 in 3f27231

'iou_t': (0, 0.1, 0.7), # IoU training threshold

yolov9/train.py

Line 546 in 3f27231

'iou_t': (0, 0.1, 0.7), # IoU training threshold

[WongKinYiu]

Currently YOLOv9 does not support evolve, you may need do some modification.

For example, modify loss gain to adaptive with hyp.
https://github.com/WongKinYiu/yolov9/blob/main/utils/loss_tal_dual.py#L247-L249

[KoniHD]

Any updates if --evolve works now?
Because for me I can run one generation and after that the script crashes with pretty much the same error.

[mkrupczak3]

I was able to get evolve to run with the changes in #150, though with what the author stated I'm not sure if it produced truly optimal results.

[KoniHD]

Thank you! I just saw it sorry.
Just curious. I am a newbie to the object detection topic. What do the 3 numbers in

'dfl': (1, 1.0, 1.75), # dfl loss gain

refer to?

[mkrupczak3]

See here:

yolov9/train.py

Line 532 in 5b1ea9a

# Hyperparameter evolution metadata (mutation scale 0-1, lower_limit, upper_limit)