Interogations concerning compactloss

[Sylv-Lej]

Hi,

I've been studied DOC for a while now and i find something weird :

is original_loss a custom loss or is it an implementation of the paper's loss ?

As written in the original paper :

K.sum((y_pred -K.mean(y_pred,axis=0))**2,axis=[1])

must be the average sample variance

but

K.var(y_pred, axis=1)

is quite different causing very different loss than expected

Am I missing something ?

[shinmura0]

As written in the original paper : K.sum((y_pred -K.mean(y_pred,axis=0))**2,axis=[1])

In implementation(from_preprocessing_to_training.ipynb), the code is same above. And this code is based on the paper.

K.var(y_pred, axis=1)

Where is this? I cannot search this code in the implementation.

[Sylv-Lej]

I understand this

you wrote this : lc = 1/(classes*batchsize) * batchsize**2 * K.sum((y_pred -K.mean(y_pred,axis=0))**2,axis=[1]) / ((batchsize-1)**2)

i understand this part that is the one written in the paper 1/(classes*batchsize) * K.sum((y_pred -K.mean(y_pred,axis=0))**2,axis=[1])

but not the multiplication by this batchsize**2 / ((batchsize-1)**2)

and why is (y_pred -K.mean(y_pred,axis=0))**2 squared ?

[Sylv-Lej]

K.var(y_pred, axis=1)

Has never been in your code, i wrote it cause of

[shinmura0]

I remembered a day of implementation. My code was based on this equation( appendix).

And your proposal

K.var(y_pred, axis=1

is possible to do so.