search

YangHM / Convolutional-Prototype-Learning

An implementation (TensorFlow) of CPL and GCPL appeared in CVPR2018 paper: "Robust Classification with Convolutional Prototype Learning"
130 stars 32 forks source link

train model with imagenet dataset #1

Open zjupx opened 6 years ago

zjupx commented 6 years ago

thanks for this great job. your paper just report the minist\cifar scores. do you have train model for imagenet or other largescale dataset? can you share more experience for this largescale image classification task? thank you very much.

YangHM commented 5 years ago

thanks for this great job. your paper just report the minist\cifar scores. do you have train model for imagenet or other largescale dataset? can you share more experience for this largescale image classification task? thank you very much.

We don't train the model on ImageNet but train it on another large scale dataset: CASIA-OLHWDB, which has 3755 classes, 2697673 training samples and 224589 test samples. It seems the GCPL model can perform a little better than the traditional soft-max based CNN on this dataset. On large scale dataset, the initialization of the prototypes is important. From our experience, we first initialize the model, then compute the mean (average) for each class with the training data, then assign the mean to the prototype of the corresponding class for initialization.

deyituo commented 5 years ago

Have you compared with other softmax based loss such as cosface ?