training loss fluctuates about libfacedetection.train

[fengyuentau]

Since our network is light-weight, it is hard for the light-weight network to further optimize itself as the training keeps going, which means it may reach local optimum.

[nhatanhphung]

I was not trying to load the pretrained model to further optimize, the training started from scratch and the loss would stay very high, and IoU would fluctuate between 0.11 and 0.16 mostly

[fengyuentau]

I was not trying to load the pretrained model to further optimize, the training started from scratch and the loss would stay very high, and IoU would fluctuate between 0.11 and 0.16 mostly

What I am trying to say is the model's complexity limits its learning ability. But, of course, you can always think of some optimization to help it learn better.

[nhatanhphung]

when I tried to train the model with the default settings with the widerface data, the result is very wrong, low IoU, high loss, testing gives 5-6 bounding boxes for 1 singular face image.

[fengyuentau]

Can you provide your full train log and the test image?

[nhatanhphung]

I attached my train log earlier, I stopped after 40 epoch as there was no change

training_log.txt

The image I used was from google
https://user-images.githubusercontent.com/20840293/129165111-ce17073f-c2a8-4cd2-b6bf-7ca047c3831d.jpg

Edit: I also included the weight and the result image using libfacedetection c++ code
yunet_final.zip
https://user-images.githubusercontent.com/20840293/129166485-bfad18f9-15ea-4888-9d10-a2f661201218.png

[fengyuentau]

Thanks for your log and the test image.

1. Since you stopped after 40 epochs, the model is not well trained. Please follow default settings and have it trained 500 epochs.
2. This model by default priorbox setting can detect faces of from ~10x10 pix to ~300x300 pix, but the face in the image (840x840 pix) you provide is way larger than this. You can try shrinking the image to 1/2 or even 1/3 and then run the detection.