search

HyeonwooNoh / caffe

modified version of caffe which support DeconvNet and DecoupledNet
Other
102 stars 69 forks source link

Multinomial Logistical Loss #6

Closed affromero closed 8 years ago

affromero commented 8 years ago

Is there any chance to use MultinomialLogisticalLossLayer for this architecture?

My problem relies on I have tiny objects to classify in each image and using softmaxlayer, the network always "learn" to predict background label, which indeed is around 95% of the image. My problem does not admit edge-boxes extracting object proposals. Someone advice me to use logistic loss.

HyeonwooNoh commented 8 years ago

Did you mean using SigmoidCrossEntropyLoss? I guess using MultinomialLogisticLoss doesn't help since this loss layer is equal to softmaxlayer except that the input have to have probability distribution already. Note that "SoftmaxLayer + MultinomialLogisticLossLyaer" is equal to "SoftmaxLossLayer".