hyp values for custom dataset

[mhyeonsoo]

Hi, thanks for the great soucres.

I am trying to train the model with the custom dataset, and found that data config for model hyperparameter needs to be set for each dataset properly.

I assumed that you've found the best combinations with the experiments, but just curious about the way to do.
If there is any condition or specific way to setup hyperparameter, could you briefly tell me how to do that?

For now, I tried evolve argument from yolov repository, which is finding appropriate hyperparameter settings, but I could see the result like below, which does not look like being trained well.

The anchor values that I used is,

  - [115,  85, 123,  98, 136, 112]
  - [149, 124, 141, 149, 157, 155]
  - [194, 248, 298, 194, 544, 178]

apparently, in debugging process, my 'out' parameter in polygon_test.py has weird values.

out = polygon_non_max_suppression(out, conf_thres, iou_thres, labels=lb, multi_label=True, agnostic=False)    
t2 += time_synchronized() - t

which is,

[tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10)), tensor([], device='cuda:0', size=(0, 10))]

And I found one more thing,
When I saw the UCAS dataset's label text file, the order of the coordinates was

(bottom left, top left, top right, bottom right).

However, the order of custom dataset's coordinates label that was generated by Tutorial2 (https://github.com/XinzeLee/PolygonObjectDetection/blob/main/polygon-yolov5/Polygon-Tutorial2.ipynb),
was

(bottom right, bottom left, top left, top right).

can this matter for the model training performance? Just quick curiosity..!

Thanks!