about [Gaussian_yolo]

[Code-Fight]

Hi, @AlexeyAB
Thanks for you share.
I found you add "[Gaussian_yolo]"
Can I train now？Any precautions？

Trhanks

[AlexeyAB]

@Code-Fight Hi, yes.

1. just use [Gaussian_yolo] instead of [yolo] in cfg-file

2. And for [Gaussian_yolo] should be:
   filters = (classes + 8 + 1) * <numbers in mask>

   instead of (for [yolo]):
   filters = (classes + 4 + 1) * <numbers in mask>]

[Code-Fight]

Hi @AlexeyAB ，Thanks for your reply.
I got it
However, when I went to train myself , the network could not converge and fluctuated around 10 rounds.
Excuse me, what is the reason?
Single card, 1080ti, lr=0.001
Thanks

[sctrueew]

@AlexeyAB Hi,

Is Gaussian_yolo layer support for the all cfg's?
SPP, TinyPRN and etc.

[AlexeyAB]

@zpmmehrdad Yes.
@Code-Fight Show chart.png

[Code-Fight]

Hi @AlexeyAB ,Thanks
This is my cfg and chart.png

sword_full_size_Gaussian.cfg.txt

[AlexeyAB]

Try to set learning_rate=0.0001 instead of learning_rate=0.001 and train

[Code-Fight]

Hi @AlexeyAB I got it , I will train again
Thank you very much

[sctrueew]

@AlexeyAB Hi,

I just changed [Yolo] to [Gussian_Yolo] and When I use [Gussian_Yolo] for Yolov3_Tiny, I got this error:
l.outputs == params.inputs, file ....\src\parser.c, line 431

Thanks

[AlexeyAB]

@zpmmehrdad

For [Gaussian_yolo] should be:
filters = (classes + 8 + 1) * <numbers in mask>

instead of (for [yolo]):
filters = (classes + 4 + 1) * <numbers in mask>]

[Code-Fight]

HI ,@AlexeyAB
I train again.
learning_rate=0.0001 instead of learning_rate=0.001
but, still the same problem.

What is the reason?

Thanks

[AlexeyAB]

This is normal for [Gaussian_yolo] layer. Check the mAP.
Or try to use even learning_rate=0.00001

[Code-Fight]

@AlexeyAB Thanks , I will try again

[Code-Fight]

Hi @AlexeyAB ,I train again with "map" .
I found the loss increase and "map" increase too.
Do you know what the reason is?
Thank you very much

[AlexeyAB]

@Code-Fight I don't know, may be delta of uncertainty is increasing https://github.com/AlexeyAB/darknet/blob/master/src/gaussian_yolo_layer.c#L182-L185

[Code-Fight]

@AlexeyAB I got it
Thanks for your help
Thanks

[sctrueew]

@Code-Fight Hi,

What cfg did you use?

[Code-Fight]

Hi @zpmmehrdad
this cfg yolov3.cfg
and changed [Yolo] to [Gussian_Yolo]

[sctrueew]

@Code-Fight Thanks, Did you check [Gussian_Yolo] with SPP?

[Code-Fight]

@zpmmehrdad no, i didn't

[sctrueew]

@Code-Fight Hi,

Could you share the mAP and FPS with [Gussian_Yolo] and with [Yolo] on your dataset?

[litingsjj]

@AlexeyAB Have you tried gaussian with yolov2?

[lq0104]

Hi @AlexeyAB ,I train again with "map" .
I found the loss increase and "map" increase too.
Do you know what the reason is?
Thank you very much

hi @Code-Fight would you show me your final learning rate, thx~

[Code-Fight]

@lq0104 Hi，this is my cfg

batch=64
subdivisions=16
width=416
height=416
channels=3
momentum=0.9
decay=0.0005
angle=0
saturation = 1.5
exposure = 1.5
hue=.1
flip=0
learning_rate=0.00001
burn_in=1000

[lq0104]

Yesterday I train my dataset with gaussian layer, as a result after hundreds of iterators the loss diverge to nan with learning rate=0.001, thanks again and i will try the new learning rate. :）

[Code-Fight]

@zpmmehrdad HI,sorry,my reply too late.
[Gussian_Yolo] FPS 55+ 1080ti