mkirchmeyer/adaptation-imputation

Unsupervised domain adaptation with non-stochastic missing data

Unsupervised Domain Adaptation with non-stochastic missing data

This repository contains the code of the following paper "Unsupervised Domain Adaptation with non-stochastic missing data" accepted at ECML2021 journal track Data Mining and Knowledge Discovery.

Dependencies

In order to run, the code requires the following Python modules specified in requirements.txt

• Numpy
• Matplotlib
• POT (Python Optimal Transport library)
• PyTorch
• sklearn

Quickstart

• Install Miniconda then source ~/.bashrc
• Create conda environment: conda create --name adaptation-imputation python=3.6 -y
• Activate environment source activate adaptation-imputation
• Install the requirements in this environment pip install -r requirements.txt.
• Install the package pip install -e . at the root

Run an experiment

• cd adaptation-imputation/experiments/launcher
• python ../../orchestration/launcher.py --experiment "dann_mnist_usps" --gpu_id=1

The experiment argument is defined in adaptation-imputation/experiments/__init__.py. All hyperparameters are stored in a .py file.

Both DANN and DeepJDOT extensions to missing data are in this repository.

• with no suffix, it will run the model with full data (for digits) else it will run the model on missing data
• with "ignore" suffix, it will run the model ignoring the missing component
• with "zeroimput" suffix, it will run the model with zero imputation for the missing component
• with "imput" suffix, it will run the model with conditional generation of the missing component

gpu_id specifies which gpu machine to use. Jobs can be run on CPU but training time will be long.

Notes

Figures are saved in folder figures and logs in a seperate results folder created when the job is launched

Utils functions are saved in utils

Datasets

Digits dataset

Digits datasets will download as part of the training script. Code is taken from existing github repos and credit is given in the .py files.

Criteo dataset

Preprocessed data is available here. If necessary follow steps below to regenerate the data.

Regenerate Criteo data

• Download Criteo Kaggle dataset from here into data folder

• Run python data_preprocessing.py in data folder

• Run following UNIX commands:

  • Define UNIX function for seeding:

  get_seeded_random() { seed="$1" openssl enc -aes-256-ctr -pass pass:"$seed" -nosalt </dev/zero 2>/dev/null ; }

  • shuf -o total_source_shuffled.txt < total_source.txt --random-source=<(get_seeded_random 42)
  • sed -n -e '1,1183117p' total_source_shuffled.txt > total_source_data.txt
  • rm total_source_shuffled.txt total_source.txt

• Training script is then ready to be run

Citation

@article{Kirchmeyer2021, author = {Kirchmeyer, Matthieu and Gallinari, Patrick and Rakotomamonjy, Alain and Mantrach, Amin}, doi = {10.1007/s10618-021-00775-3}, isbn = {1573-756X}, journal = {Data Mining and Knowledge Discovery}, title = {Unsupervised domain adaptation with non-stochastic missing data}, url = {https://doi.org/10.1007/s10618-021-00775-3 }, year = {2021}}