FocalLoss

FocalLoss copied to clipboard

May I ask why the definition of ReLU6 is added? I actually do not see the rationale of adding it regarding focal loss definition. Thanks.

kazenokizi

All the boxes are calculated in focal loss, function MineHardExamples should be modified to cover all boxes. Another tricky way is set _pos_neg_ratio_ into a very large number, but takes...

basaltzhang

How about the result with SSD?

35

I'm glad to see your work with focal loss, have you got some better performance with focal loss than ohem in ssd? Moreover, have you test focal loss with your...

bailvwangzi

Details of the code

1

Hello, I would like to ask the specific details of code, label_value and pt specifically refers to what, label_value is the anchor corresponding label, pt is the probability of the...

xiang1226

which version of ssd

1

Thanks for such a nice work. But I met some problems and I want to ask which version of ssd these codes match. Please give me the link. Thank you...

yqkgithup

I got loss = nan .....

5

Solving VGG_VOC0712_SSD_300x300_train I1112 16:14:25.373922 12274 solver.cpp:295] Learning Rate Policy: multistep I1112 16:14:26.191864 12274 solver.cpp:243] Iteration 0, loss = 393.183 I1112 16:14:26.191922 12274 solver.cpp:259] Train net output #0: mbox_loss = 463.676...

jiachen0212

About α

1

Hi, The original paper introduces a weighting factor α for class 1 and 1-α for class -1, but in your code I just see α, I think it is wrong,...

igoodboy

In my code: for (map::iterator it = match_indices.begin(); it != match_indices.end(); ++it) { const int label = it->first; // Get positive indices. int num_pos = 0; for (int m =...

kingstarcraft

About focal loss layer

1

Hi, I'm wondering have you changed a two-stage net like faster-rcnn or FPN to a one-stage net like the paper said? Do I just need to change the softmax loss...

zhanglonghao1992