Source Code for: Generative Imputation and Stochastic Prediction

Citation

M. Kachuee, K. Kärkkäinen, O. Goldstein, S. Darabi, M. Sarrafzadeh, 
Generative Imputation and Stochastic Prediction, 
IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020.

Project Structure

• train.py: main entry point of the program, used to run different experiments
• utils.py: a set of utility functions
• test_cifar.sh: a simple script to run GI experiment for the CIFAR-10 dataset
• README.md: this readme file
• environment.py: conda environment settings
• data.py: load and preprocessing for different datasets
• otherwork/: source code from other work (with some modifications to integrate)
• models/: pytorch model architectures and modules
• imputers/: training process for different imputers

Note: as shown in the example below, always fix the hash seed. We use python hashing to fingerprint samples.

Command example

export PYTHONHASHSEED=0
python3 train.py --exp "ENS_EPS2000" --dataset cifar10 --data_dir ~/Database/Image/ \
   --objective bce --lr_d 0.0005 --lr_g 0.0005 --lr_patience 0.25 \
   --missing_type mcar_rect --missing_rate 0.20 --hint_rate 0.0 --alpha 0.0 \
   --device cuda:0 --epoches 2000 --eval_freq 0.05 --batch_size 64 \
   --train_predictor --n_samples 128  --aug_noise_std 0.0 \
   --result_dir ./run_outputs/ --dump_ens