LogLoss on Kaggle

FuriouslyCurious · March 24, 2017, 9:39am

Kaggle competitions rely on Log Loss (which is torch.nn.BCELoss):

LogLoss= −1/n∑i=1 to n [yi log(ŷ i) + (1−yi) log(1−ŷ i)],

where:
  n   : is the number of patients in the test set
  ŷi  : is the predicted probability of the image belonging to a patient with cancer
  yi  : is 1 if the diagnosis is cancer, 0 otherwise

Can someone please explain how this loss differs from torch.nn.MultiLabelSoftMarginLoss?

torch.nn MLSM Loss:
loss(x, y) = - sum_i (y[i] log( exp(x[i]) / (1 + exp(x[i])))
                      + (1-y[i]) log(1/(1+exp(x[i])))) / x:nElement()

Thanks!

FuriouslyCurious · March 31, 2017, 2:58am

Hi Torchies,

Bumping an old question… if someone has a paper / presentation comparing different loss functions and their effects, it will really help me out.

Thanks!