caffe_custom_blob_multinomial_logistic_loss_layer

This repository contains an implementation of a multinomial logistic loss layer that accepts a probability blob of size NxCxHxW and a label blob size of Nx1xHxC.

The code is based on the implementation of softmax_loss_layer.hpp and softmax_loss_layer.cpp from https://github.com/TimoSaemann/caffe-segnet-cudnn5, which was a fork of BVLC/caffe.

This repository contains only a new layer in Caffe to be used in conjunction with Segnet.

The original Caffe's Multinomial Logistic Loss layer expects a label blob to have a size of Nx1x1x1 i.e. one label per one image. In my case, I would like to customise the probabilies obtained from the Softmax layer so I need a different Multinomial Logistic Loss layer, and that the use of SoftmaxWithLoss layer is not possible.

Build

1. Place the file 'custom_blob_multinomial_logistic_loss_layer.hpp' into $CAFFE_SEGNET_ROOT\include\caffe\layers.
   
2. Place the file 'custom_blob_multinomial_logistic_loss_layer.cpp' and 'custom_blob_multinomial_logistic_loss_layer.cu' into $CAFFE_SEGNET_ROOT\src\caffe\layers.
   
3. Open the terminal and cd to $CAFFE_SEGNET_ROOT and build the code:
   

make clean
make all

Usage

Below is an example use of this new layer after the Softmax layer. As you can see, this is equivalent to using SoftmaxWithLoss but allows a flexibility to modify the softmax results before computing loss.

layer {
  name: "softmax"
  type: "Softmax"
  bottom: "conv1_1_D"
  top: "softmax"
  softmax_param {engine: CAFFE}
}

# softmax values can be modified here before passing to the next layer

layer {
  name: "loss"
  type: "CustomBlobMultinomialLogisticLoss"
  bottom: "softmax"
  bottom: "label"
  top: "loss"
}