Why the anchors in cfg file are not same as trained anchors?

[PonyPC]

Why the anchors in cfg file are not same as trained anchors?

[yolo]
mask=0, 1, 2
anchors=18, 17, 12, 33, 26, 20, 44, 90, 85, 51, 58, 81, 100, 388, 381, 132, 139, 379

My data anchors:

  - [25,97, 95,33, 35,95]  # P3/8
  - [45,90, 86,52, 58,82]  # P4/16
  - [73,68, 52,135, 106,81]  # P5/32

[marcoslucianops]

The anchors are extracted from pt file during the conversion.

[PonyPC]

Got it. Thanks

[PonyPC]

Yolov5 has fixed this issue.
#7058

[PonyPC]

The anchors are not same as slightly.

25,99, 36,86, 97,33, 50,75, 75,56, 38,118, 51,104, 111,51, 71,82, 59,119, 103,74, 83,105
Deepstream-Yolo output:
[25, 99, 35, 85, 97, 32, 50, 75, 75, 55, 38, 118, 50, 103, 111, 50, 70, 82, 59, 119, 102, 74, 83, 104]

Should I ignore this? Does it impact the precision?
@marcoslucianops

[marcoslucianops]

which anchors are in cfg file?

[PonyPC]

[25, 99, 35, 85, 97, 32, 50, 75, 75, 55, 38, 118, 50, 103, 111, 50, 70, 82, 59, 119, 102, 74, 83, 104]
second line.

[PonyPC]

Hi, I have sent to your email. Please resolve two issues together. Thanks

[PonyPC]

The anchors in model which I sent you are:

tensor([[[ 26.85938,  79.68750],
         [ 75.12500,  31.82812],
         [ 51.09375,  64.68750],
         [ 29.70312, 114.75000]],

        [[ 43.28125,  85.43750],
         [ 76.43750,  51.43750],
         [107.37500,  41.90625],
         [ 42.84375, 118.87500]],

        [[ 68.06250,  77.68750],
         [ 61.03125, 112.12500],
         [104.62500,  69.81250],
         [ 89.87500,  97.68750]]], dtype=torch.float16)

It seems about round or floor conversion from float to integer

[marcoslucianops]

Please check with updated utils/gen_wts_yolov5.py file.

[PonyPC]

Okay, I'll try again

[PonyPC]

It runs well, thanks.