Huaisheng Zhu, Teng Xiao, Zhimeng Guo
- Install the packages listed in
requirements.txt - Download the two datasets into the same folder:
- CIFAR-10-C (Hendrycks and Dietterich) from this repository, which links to this shared storage.
- CIFAR-10.1 (Recht et al.) from this repository.
- Set the data folder to where the datasets are stored by editing:
--datarootargument inmain.py.--datarootargument inbaseline.py.datarootvariable inscript_test_c10.py.
- Run
script.shfor the main results (TTT-personalized), and./baseliness/script_baseline.shfor the baseline results. - The results are stored in the respective folders in
results/and./baseliness/results. - Once everything is finished, the results can be compiled and visualized with the following utilities:
show_table.pyparses the results into tables and prints them.show_plot.pymakes bar plots, and prints the tables in latex format; requires first runningshow_table.py.show_grad.pymakes the gradient correlation plot../baseliness/show_table.pyparses the baseline results into tables and prints them../baseliness/show_plot.pymakes bar plots for baseline results, and prints the tables in latex format; requires first runningshow_table.py../baseliness/show_grad.pymakes the gradient correlation plot for baselines.
- The whole training process may take a Tesla V100 several days. So please change the
CUDA_VISIBLE_DEVICESinscript.shand./baseliness/script_baseline.shto select a suitable GPU.