Learning to Teach with Dynamic Loss Functions
This repo contains the simple demo for the NeurIPS-18 paper: Learning to Teach with Dynamic Loss Functions.
@inproceedings{wu2018learning,
title={Learning to teach with dynamic loss functions},
author={Wu, Lijun and Tian, Fei and Xia, Yingce and Fan, Yang and Qin, Tao and Jian-Huang, Lai and Liu, Tie-Yan},
booktitle={Advances in Neural Information Processing Systems},
pages={6466--6477},
year={2018}
}
Description
- Please note this is only a simple demo for the Mnist experiments based on Lenet.
- Please note the algorithm in the demo is little different to the paper, but the main spirit is same.
- The code is based on the Theano framework, which is somehow too old to directly apply this code.
Detailed Critical Codes
Refer to loss_lenet_light_dynamic.py
for the detailed demo codes. The general comments are here:
- Define the loss parameters and loss computation graph
- Teacher model paramters
- Define the updates of model training according to the loss function (jointly train model update)
- Define the reverse mode training
- Train student model with fixed loss function
- Detailed reverse model training
The 'reverse model training' defines the updates of the teacher model, and the last one is the detailed reverse model update chains.