bbenligiray / rml-cnn

Semi-Supervised Robust Deep Neural Networks for Multi-Label Classification

RML-CNN

Code covering some of the experiments in the following paper:

Cevikalp, H., Benligiray, B., Gerek, O. N.. (2019). Semi-Supervised Robust Deep Neural Networks for Multi-Label Image Classification. In Pattern Recognition.

Use this to create a nus_wide.h5 and this to create a ms_coco.h5 file. Download resnet101_weights_tf.h5 from here. Put these in /rml-cnn, then run run.sh to run the experiments.

Alternatively, you can use the loss functions in ml_loss with any model/dataset you like, see main.py for reference. Note that your final layer's activation should be None for all loss functions (including softmax).