Skip to content

mathematicallfs/Uniform-LGI

Repository files navigation

From Optimization Dynamics to Generalization Bounds via Łojasiewicz Gradient Inequality

This repository contains the official codes that used for generating numerical results in the following paper:

Fusheng Liu, Haizhao Yang, Soufiane Hayou, Qianxiao Li, From Optimization Dynamics to Generalization Bounds via Łojasiewicz Gradient Inequality, Transactions on Machine Learning Research, 2022.

Generating the numerical results for Figure 1 & Figure 2

Saving results

For non-analytic and non-lgi models, modify the function in get_lgi_counter.py. Then run the following commands to save results for these two models and linear regression models respectively.

python get_lgi_counter.py
python get_lgi_synthetic.py

For neural network models, change different models in get_lgi.py. Then run the following commands to save results with respect to different models

python get_lgi.py

Plotting Figure 1 & 2

Modifying 'name, K0, s' in plot_lgi.py, then

python plot_lgi.py

Generating the numerical results for Figure 3

Run the following commands to save results

python get_lgi_corruption.py
python get_vary_size.py

Then

python plot_corruption.py
python plot_vary_size.py

About

Official code for ''From Optimization Dynamics to Generalization Bounds via Łojasiewicz Gradient Inequality'' (TMLR)

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages