0

I have written a python script which calculates the median frequency balancing weights for each class during the pixel-wise segmentation. Then, I added a Python Layer to the caffe model definition, which sends the weights to the loss function. Based on this link, user mentions that SoftmaxWithLoss layer in caffe correspond to TensorFlow softmax_cross_entropy_with_logits. My question is how can I send the weights to SoftmaxWithLoss layer? What other Loss layers can be used with median frequency balancing? I used InfoGainLoss, but it does not converge. Your help is really appreciated.

S.EB
  • 1,966
  • 4
  • 29
  • 54

1 Answers1

0

If you want to weight the "SoftmaxLoss" (i.e., cross entropy loss) according to classes then "InfogainLoss" is what you need.
Note that infogain layer was upgraded in caffe a few months ago: it now integrates "Softmax" into the loss computation for robust gradient estimation.

Shai
  • 111,146
  • 38
  • 238
  • 371
  • Thanks for your comment, I applied `InfoGainLoss`, however, it is over-segmenting. I do not know what is the reason. My other question is that "`Softmax`" does not have any parameter that I can feed the weights by using that, could I to explain more? Thanks – S.EB Jan 23 '18 at 07:50