/rethinking_lupi

Experiments on knowledge transfer techniques in the LUPI paradigm

Primary LanguageJupyter Notebook

Rethinking knowledge transfer in Learning Using Privileged Information (LUPI)

This code reproduces some of the experiments from Lopez et al. 2016 and Collier et al. 2022, two papers that propose various knowledge transfer techniques for the LUPI paradygm. Additionally, we conduct a real-life experiment using an open-source dataset from IJCAI15 competition.

Synthetic experiments synthetic/

The notebooks gen_dist_exp.ipynb and tram_exp.ipynb contain experiments with synthetic data for Generalized distillation Lopez et al. 2016 and TRAM Collier et al. 2022.

Experiments with stadardized datasets

MNIST: scaling beyond /mnist

This code reproduces the MNIST experiment from Lopez et al. 2016 and extends its training epochs to beyond the original 50 epochs. The code has been ported to work with Python 3.9, which required changing some of the requirements. For training, download the MNIST dataset and put it to mnist/data, cd to mnist/ folder and run

python mnist_varying_size.py

Sarcos: generalized distillation vs flat zero predictions /sarcos

This code reproduces the experiment from Lopez et al. 2016 and extends it with a naive baseline of predicting flat zeros. The code is from 2016 and requires Python 2.7.18 to run. Because the original code did not have requirements specified we have added the highest working requirements we could find. sarcos/requirements.txt contains the necessary dependencies.

For training, download the Sarcos dataset and put it to sarcos/data, cd to sarcos/ folder and run

python sarcos.py

Real-world experiment with IJCAI15 competition data bandit_data/

We compare Generalized distillation and TRAM on the Repeat Buyers Prediction dataset, a large-scale public dataset from the IJCAI-15 competition. The data provides users’ activity logs of an online retail platform, including user-related features, information about items at sale, and implicit multi-behavioral feedback such as click, add to cart, and purchase.

For training:

  1. download data from https://tianchi.aliyun.com/dataset/42
  2. copy user_info_format1.csv and user_log_format1.csv to bandit_data/data/IJCAI15/
  3. cd bandit_data/
  4. run python train.py