Issue when executing gen_anchors

[StevenPuttemans]

Hello @Jumabek

First of all, thanks for providing the necessary scripts to do some customizations of darknet.

I tried the gen_anchors script, with the following command python gen_anchors.py -filelist ~/Github/darknet/train.txt -num_clusters 5 -output_dir coconut_anchors/ but whenever running it, it seems I have an error through python

Traceback (most recent call last):
  File "gen_anchors.py", line 154, in <module>
    main(sys.argv)
  File "gen_anchors.py", line 128, in main
    w,h = line.split(' ')[3:]
ValueError: too many values to unpack

Any idea what could be the reason of this?

[StevenPuttemans]

Also, the file looks like this, which is the file called before this error is raised.

0 0.366279 0.707731 0.100000 0.100000 0 0.281418 0.061305 0.100000 0.100000 0 0.506687 0.026250 0.100000 0.100000 0 0.613862 0.108427 0.100000 0.100000 0 0.619609 0.019929 0.100000 0.100000 0 0.176829 0.143769 0.100000 0.100000 0 0.292912 0.164457 0.100000 0.100000 0 0.403247 0.166181 0.100000 0.100000 0 0.429107 0.239164 0.100000 0.100000 0 0.279694 0.252381 0.100000 0.100000 0 0.527375 0.239738 0.100000 0.100000 0 0.640872 0.244336 0.100000 0.100000 0 0.630528 0.351080 0.100000 0.100000 0 0.522491 0.351655 0.100000 0.100000 0 0.437440 0.360275 0.100000 0.100000 0 0.628229 0.439722 0.100000 0.100000 0 0.629953 0.524773 0.100000 0.100000 0 0.626505 0.643154 0.100000 0.100000 0 0.513871 0.637982 0.100000 0.100000 0 0.513871 0.527646 0.100000 0.100000 0 0.514445 0.450066 0.100000 0.100000 0 0.164761 0.527646 0.100000 0.100000 0 0.158440 0.624765 0.100000 0.100000 0 0.153842 0.714556 0.100000 0.100000 0 0.723336 0.640568 0.100000 0.100000 0 0.949755 0.640568 0.100000 0.100000 0 0.948031 0.512130 0.100000 0.100000 0 0.956651 0.421333 0.100000 0.100000 0 0.841717 0.429953 0.100000 0.100000 0 0.826201 0.512705 0.100000 0.100000 0 0.728508 0.522474 0.100000 0.100000 0 0.746323 0.356683 0.100000 0.100000 0 0.749771 0.243474 0.100000 0.100000 0 0.854935 0.344040 0.100000 0.100000 0 0.857808 0.257266 0.100000 0.100000 0 0.917573 0.227958 0.100000 0.100000 0 0.968719 0.256116 0.100000 0.100000 0 0.964121 0.346914 0.100000 0.100000 0 0.916424 0.023233 0.100000 0.100000 0 0.819306 0.027256 0.100000 0.100000 0 0.833672 0.103686 0.100000 0.100000 0 0.726784 0.100813 0.100000 0.100000 0 0.719314 0.017486 0.100000 0.100000

[StevenPuttemans]

Ok I solved it, seemed that my files had trailing whitespaces. That was solved rather quickly and then the script ran just fine. Executed on my data, finally got anchors that make sense.

Centroids =  [[ 0.09999999  0.09999999]
 [ 0.          0.        ]
 [ 0.          0.        ]
 [ 0.          0.        ]
 [ 0.          0.        ]]

Anchors =  [[ 0.          0.        ]
 [ 0.          0.        ]
 [ 0.          0.        ]
 [ 0.          0.        ]
 [ 1.29999981  1.29999981]]

centroids.shape (5, 2)

[StevenPuttemans]

Does it mean that once you get zeros, there is no use for adding multiple anchors?

[Jumabek]

Hi @StevenPuttemans , Yes, it seems for your dataset single anchor is enough. Meaning most of your objects probably have the same ratio of height to width.

[StevenPuttemans]

@Jumabek thanks for getting back to me! I will experiment with this new insight. Did you even tried adding manual anchors? For example I know objects have to be 100x100 pixels, so in principle a 1.0,1.0 anchor would be more fitting than the oversized 1.3,1.3?

Or is there a reason for the oversizing?

[StevenPuttemans]

As you can see, I managed to force it to learn square detections, however it is still not performing as I desire. Now my detections are very small. Any suggestions always welcome!

ps. edited - wron image

[Jumabek]

what is your input image size in cfg file? If it is not 416, then you have to change this line accordingly

[Jumabek]

Or you can just scale 1.3,1.3 3x and put anchor as 3.9,3.9 Another thing to try is having more than one anchors. For example make 4 anchors all of them are the scaled version of 1.3, 1.3

[StevenPuttemans]

Yeah I noticed. So it was 992 in that case. Changed it back to original and it worked. Now further finetuning the model. If it is 992 we need 2.0,2.0?

[StevenPuttemans]

Ok thnx for feedback!

[Jumabek]

No, you need to change 416 -> 992 in https://github.com/Jumabek/darknet_scripts/blob/master/gen_anchors/gen_anchors/gen_anchors.py#L44 line. and then compute the anchors again.