Training and evaluating self-supervised deep neural networks.
These models were all trained entirely without labeled data using colorization as a proxy task:
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)
- Colorization as a Proxy Task for Visual Understanding, Larsson, Maire, Shakhnarovich, CVPR 2017 (https://arxiv.org/abs/1703.04044)