vita-group/l2o-training-techniques

[NeurIPS 2020 Spotlight Oral] "Training Stronger Baselines for Learning to Optimize", Tianlong Chen*, Weiyi Zhang*, Jingyang Zhou, Shiyu Chang, Sijia Liu, Lisa Amini, Zhangyang Wang

Training Stronger Baselines for Learning to Optimize

Code for this paper Training Stronger Baselines for Learning to Optimize.

Tianlong Chen*, Weiyi Zhang*, Jingyang Zhou, Shiyu Chang, Sijia Liu, Lisa Amini, Zhangyang Wang

Overview

With many efforts devoted to designing more sophisticated L2O models, we argue for another orthogonal, under-explored theme: the training techniques for those L2O models. We show that even the simplest L2O model could have been trained much better.

• Curriculum Learning

  We first present a progressive training scheme to gradually increase the optimizer unroll length, to mitigate a well-known L2O dilemma of truncation bias (shorter unrolling) versus gradient explosion (longer unrolling).

• Imitation Learning

  We further leverage off-policy imitation learning to guide the L2O learning , by taking reference to the behavior of analytical optimizers.

Our improved training techniques are plugged into a variety of state-of-the-art L2O models, and immediately boost their performance, without making any change to their model structures.

Experiment Results

Training the L2O-DM baseline to surpass the state-of-the-art

Training state-of-the-art L2O models to boost more performance

Ablation study of our proposed techniques

Imitation Learning v.s. Self-Improving

Reproduce Details

Experimental details on L2O-DM and RNNProp are refer to this README.

Experimental details on L2O-Scale are refer to this README.

Citation

If you use this code for your research, please cite our paper:

l2o-training-techniques's Issues

How many 1080Ti GPUs are needed for reproduce the experiment results?

Hi,

Thanks for sharing your codes. If possible, could you let us know the number of GPUs used for the experiments, e.g., training L2O-DM-CL-IL with N_train=500?