Jie Zhang

df / dx is needed for BP

margin won't change the direction. I think the weight initialization makes this.

@bkj Hi, I'm not the author of the paper. The official implement released by author is [wy1iu/LargeMargin_Softmax_Loss](https://github.com/wy1iu/LargeMargin_Softmax_Loss). I personally tried `norm(w) * norm(x) * (cos(theta) / m)` and it plays...

you write a wrong layer type. ``` layer { name: "loss" type: "OrdinalRegressionLoss" ordinal_regression_loss_param { k: 4 } bottom: "fc8bisi" bottom: "label" top: "loss" } ```

There are two kind of weight. `inter weight` represents the importance of every task and `outer weight` balances the training samples of every task. In this layer implementation, `inter weight`...

something like below, this layer won't give the final output label (which always a label integer for every data) and can only be used in training. You should implement yourself...

this layer is just for test the accuracy. remove this layer can still train the network.

The training details can be found in [kongsicong/Age_recognition_OR](https://github.com/kongsicong/Age_recognition_OR).

If classifier of C0 ~ Ci gives label 1, then the output label is i, we predict age of [0, 99]. It's different from original paper. I recommend you can...

I don't know much detail of the training. Maybe @kongsicong can help this.