Convergence speed of the network?

[twmht]

I tried to train the network with WIDERFACE, But I found that it's difficult to converge.

after the several epochs, both the precision and recall are less than 1%, did you have the similar problem?

it seems that the paper work is trained at least 200 epochs, did you train such a long epochs?

I tried to mask the small face of the training data (https://github.com/zeusees/FaceBoxes), but it does not speed up the training.

[XiaXuehai]

The number of training epochs is 300 which is same as the paper. In the paper, the author did not mention the mask. so I did not try it.

[xtanitfy]

@twmht I find the loss is still about 3.0 after 100 epochs. I don't know if you are also this problem. Have you finished the training? Is that normal? Looking forward your reply! Thanks!

[laidy3]

@twmht I find the loss is still about 3.0 after 100 epochs. I don't know if you are also this problem. Have you finished the training? Is that normal? Looking forward your reply! Thanks! I also meet the problem. Have u solved it?

[xtanitfy]

It seems like the weights have been finished.You can use the demo to test pictures by using the weights.

[laidy3]

@twmht How about train from sratch?

[twmht]

@laidy3

I trained from scatch, do I need to use a pretrained weight?

[laidy3]

@twmht I think it doesn’tneed. How about the preformence of traing from scatch?

[twmht]

@laidy3

since the convergence is slow, So I stop training early at 1xx epochs. I forgot the performance, it's very low.

[twmht]

@laidy3

Did you tried to train from scratch? how about the performance?

[laidy3]

@twmht Xavier init, 12 hours, batch size 128, 145 epochs, val loss= 5.22

[XiaXuehai]

150-300 epoch of my training loss.

[lucasjinreal]

@XiaXuehai @laidy3 So, how does the inference result