Implementation of temporal ensembling and Pi-model. Samuli Laine and Timo Aila, NVIDIA.
Released as part of ICLR 2017 paper submission "Temporal Ensembling for Semi-Supervised Learning".
Additional code (report.py, theano_utils.py, thread_utils.py) by Tero Karras, NVIDIA. Code in zca_bn.py adapted from Tim Salimans' repository at: https://github.com/TimSalimans/weight_norm/blob/master/nn.py
Package versions used when preparing the paper:
- Theano 0.9.0.dev4
- Lasagne 0.2.dev1
- CUDA toolkit 8.0, CUDNN 5105
To configure a training run, edit config.py. To execute, run train.py.