Invalid dists

[nrpr93]

Hi,

I get NaN dists for some reason. The scripts can´t read all the annotations and he can get all the information in there, but we can´t give me the dists correctly. Can someone help me with this?

Regards, NR

[TheMikeyR]

@nrpr93 I've experienced the same issue regarding getting NaN dists, I solved mine by reducing the clusters 1 or 3 instead of 5.

[nrpr93]

Now it´s work with one cluster, but it just run two iterations, and the final anchor doesn´t seem very right. I assume that´s because the number of iterations.

[TheMikeyR]

Are you testing with the voc data?

[nrpr93]

No, I´m testing with my own data.

[TheMikeyR]

Maybe there are some issues with your own data, you could try to do the voc example in the readme and see if you can reproduce the same results.

Could you tell me about your data and how it is formatted?

[nrpr93]

The format of my data it´s pretty simple I have 15 classes, and I use the same tpe of annotation that yolo wants: , my 'width' and 'height' it´s always the same for all the images, just the 'x' and 'y' change in all.

0 0.53 0.35 0.0625 0.08333333333333333 1 0.53 0.490909090909 0.0625 0.08333333333333333 2 0.476666666667 0.486363636364 0.0625 0.08333333333333333 3 0.583333333333 0.490909090909 0.0625 0.08333333333333333 4 0.453333333333 0.604545454545 0.0625 0.08333333333333333 5 0.61 0.604545454545 0.0625 0.08333333333333333 6 0.446666666667 0.731818181818 0.0625 0.08333333333333333 7 0.613333333333 0.75 0.0625 0.08333333333333333 8 0.53 0.581818181818 0.0625 0.08333333333333333 9 0.493333333333 0.672727272727 0.0625 0.08333333333333333 10 0.563333333333 0.677272727273 0.0625 0.08333333333333333 11 0.48 0.877272727273 0.0625 0.08333333333333333 12 0.566666666667 0.877272727273 0.0625 0.08333333333333333 13 0.47 1.05 0.0625 0.08333333333333333 14 0.57 1.04545454545 0.0625 0.08333333333333333

[TheMikeyR]

How much data do you have for each class ?

[TheMikeyR]

Is that all the bounding boxes you have? @nrpr93 Because that is probably your issue. I've used ~60k frames with ~180k detections to find my anchor boxes, so I believe the program can't generate anchor boxes if you only have one detection per class, you need to collect a bigger dataset if that is the case.

[nrpr93]

Of course I don´t have just one detection for class, I have 18k for each class. I just show a example of a file with 15 detetion.

[TheMikeyR]

Okay sorry, I don't know the issue then, maybe @Jumabek can help?

[Jumabek]

Hi @nrpr93 , could you upload your annotation file here (One would be sufficient). I want to test your format (I suspect it has to do some trailing white spaces or something)

if you cannot update then please email [email protected]

[Jumabek]

Sorry I just read that you guyz are getting NaN. @TheMikeyR, @nrpr93 can you send me your annotations. No need for the images, just annotation. I want to fix the code using your annotations if you do not mind. BTW sorry I am replying late, I was on a vocation

[nrpr93]

I believe the problem it´s because my width and height in all objects are the same. So that´s a division calulation problem.

Here´s one example of my annotation: 0 0.493750 0.325000 0.09 0.9 1 0.490625 0.450000 0.09 0.9 2 0.440625 0.450000 0.09 0.9 3 0.540625 0.450000 0.09 0.9 4 0.425000 0.562500 0.09 0.9 5 0.562500 0.566667 0.09 0.9 6 0.418750 0.687500 0.09 0.9 7 0.559375 0.687500 0.09 0.9 8 0.493750 0.537500 0.09 0.9 9 0.465625 0.625000 0.09 0.9 10 0.525000 0.625000 0.09 0.9 11 0.453125 0.804167 0.09 0.9 12 0.528125 0.804167 0.09 0.9 13 0.440625 0.966667 0.09 0.9 14 0.534375 0.954167 0.09 0.9 1.txt

[TheMikeyR]

@Jumabek don't worry I believe my error is due to the cluster size was too high compared to the size of my annotations ~100x100 so I couldn't go up to 5 in clusters, but I had no issues using voc data, if you still want my data just give me a thumps up and I will send it 😄

[Jumabek]

@nrpr93 now I see it. It seems all of your object have the same width and height (0.09,0.9). purpose of generating anchors (a.k.a clustering) is to group the objects that are closer to each other in terms of their shape).

Since in your case you have same shape but different class objects, you should choose only one cluster.

if the code fails for some other reason here I calculate the centroid for you:

input_height = 416; //height of your network input
input_width = 416; //width of your net input, depending on you it could be different
downsampling_factor = 32;

//then you will have 1 cluster/anchor which is 
anchor_w = 0.09*input_weight/downsampling_factor
anchor_h = 0.9*input_height/downsampling_factor