Variants-ResNet-Prototxt-Caffe

Variants-ResNet-Prototxt-Caffe copied to clipboard

How could this generator be adapted to a custom data set?

@achaiah Sorry for the late reply. This generator is currently implemented to replicate the results of the original paper on the CIFAR dataset. I haven't implemented the generator for ImageNet. The CIFAR ResNet accepts 32*32 inputs, and outputs 10 scores for 10 classes. So if you want to adapt the CIFAR ResNet to your own dataset, you should change 10 to the number of classes of your dataset. The size of images should be similiar to CIFAR, or you'd better consider the ImageNet ResNet-34/50/101/152.

@Coldmooon : by the way do you happen to still have the preprocessed datasets for CIFRAR10? or share the script as well for a quick and hassle free benchmark comparing your results and ours?

@Coderx7 : The ZCA whitened CIFAR dataset (no pad) can be downloaded from https://gist.github.com/mavenlin/e56253735ef32c3c296d#descriptions.

The ZCA whitened CIFAR dataset ( padded with 8 pixels for each side ) can be found here, https://github.com/flukeskywalker/highway-networks

For the 4-pixel padding case, the script is at https://github.com/lisa-lab/pylearn2/blob/master/pylearn2/scripts/datasets/make_cifar10_gcn_whitened.py and refer to https://gist.github.com/hiwonjoon/8f91034cc1168f2d2dd5