why loss of landmark task does not descend?

[gobigrassland]

I have trained your codes many times, but the loss of landmark task does not converge. I don' t known what is wrong. When i only train face classification and regression of bounding boxes, losses of these tasks both descend. Why?

[foreverYoungGitHub]

Do you generate the training data and labels by yourself?

If so, what's the learning rate of your code?

[gobigrassland]

I check and modify the code generating training data. Then i re-generate the training data and labels and train again. Although the loss of landmark seems to converge, the detected landmarks of faces are not correct.

[KangolHsu]

i prepare my data and label like this:

image_path face face_box_cords landmarks img1.jpg 1 0.12 0.21 0.34 0.27 [0.1,0.2,0.3,0.4,0.5,0.1,0.2,0.3,0.4,0.5]; //positive face img2.jpg 1 0.02 0.12 0.24 0.17 [0.1,0.2,0.3,0.4,0.5,0.1,0.2,0.3,0.4,0.5]; // partial face img3.jpg 0 0 0 0 0 [ 0, 0 , 0, 0 , 0 , 0 , 0 , 0 , 0 , 0 ]; // no face ...... ......

i want to konw whether the image data (eg:img1.jpg ) is just a face image or face part with background part?

[blankWorld]

@gobigrassland Hi,I met the same question . Have you solved the problem ? Although my landmarks loss converge at last it descend so slowly and it does badly on my landmark validation set.