self-supervision icon indicating copy to clipboard operation
self-supervision copied to clipboard

Training and evaluating self-supervised deep neural networks

Self-supervision

Training and evaluating self-supervised deep neural networks.

  • http://people.cs.uchicago.edu/~larsson/color-proxy

Models

These models were all trained entirely without labeled data using colorization as a proxy task:

  • AlexNet <http://people.cs.uchicago.edu/~larsson/color-proxy/models/alexnet.caffemodel.h5>_
  • VGG-16 <http://people.cs.uchicago.edu/~larsson/color-proxy/models/vgg16.caffemodel.h5>_
  • ResNet-152 <http://people.cs.uchicago.edu/~larsson/color-proxy/models/resnet152.caffemodel.h5>_

The models are saved in Caffe HDF5 files. Our evaluation code in TensorFlow can load these files. The AlexNet model uses LRN (applied before pooling)

Paper

  • Colorization as a Proxy Task for Visual Understanding, Larsson, Maire, Shakhnarovich, CVPR 2017 (https://arxiv.org/abs/1703.04044)