Dynamic Feature Acquisition with Arbitrary Conditional Flows

This is the official repository for ACFlow-DFA. Please see our paper for technical details.

Our method dynamically acquires new features to improve the predictions.

prerequisites

clone our repository

git clone

install required packages (virtual environment is preferred)

pip install -r requirements.txt

Prepare datasets

see data folder for code to preprocess all datasets we used. We cannot provide the original data due to the copyright issues. Please download from the corresponding repository (links are given in the comments.)

Experiments

Classification

• for synthetic dataset

(Train) python scripts/train.py --cfg_file=./exp/syn/params.json
(ACFlow DFA) python scripts/test_cls_dfa.py --cfg_file=./exp/syn/params.json
(ACFlow SFA) python scripts/test_cls_sfa.py --cfg_file=./exp/syn/params.json

• for gas dataset

(Train) python scripts/train.py --cfg_file=./exp/gas/params.json
(ACFlow DFA) python scripts/test_cls_dfa.py --cfg_file=./exp/gas/params.json
(ACFlow SFA) python scripts/test_cls_sfa.py --cfg_file=./exp/gas/params.json

Regression

• for housing dataset

(Train) python scripts/train.py --cfg_file=./exp/housing/params.json
(ACFlow DFA) python scripts/test_reg_dfa.py --cfg_file=./exp/housing/params.json
(ACFlow SFA) python scripts/test_reg_sfa.py --cfg_file=./exp/housing/params.json

• for whitewine dataset

(Train) python scripts/train.py --cfg_file=./exp/whitewine/params.json
(ACFlow DFA) python scripts/test_reg_dfa.py --cfg_file=./exp/whitewine/params.json
(ACFlow SFA) python scripts/test_reg_sfa.py --cfg_file=./exp/whitewine/params.json

Bayesian Network

• for synthetic datasets from Bayesian network repository (asia and sachs)

(Train) python scripts/train.py --cfg_file=./exp/asia/params.json
(ACFlow DFA) python scripts/test_reg_dfa.py --cfg_file=./exp/asia/params.json
(BN Learn) python scripts/bn_learn.py --cfg_file=./ep/asia/params.json
(BN:ACFlow) python scripts/test_bn_dfa.py --cfg_file=./exp/asia/params.json --gfile=./exp/asia/bn_learn/results.pkl
(BN:GT) python scripts/test_bn_dfa.py --cfg_file=./exp/asia/params.json --gfile=./data/asia/asia_bn_0.3.pkl

(Train) python scripts/train.py --cfg_file=./exp/sachs/params.json
(ACFlow DFA) python scripts/test_reg_dfa.py --cfg_file=./exp/sachs/params.json
(BN Learn) python scripts/bn_learn.py --cfg_file=./ep/sachs/params.json
(BN:ACFlow) python scripts/test_bn_dfa.py --cfg_file=./exp/sachs/params.json --gfile=./exp/sachs/bn_learn/results.pkl
(BN:GT) python scripts/test_bn_dfa.py --cfg_file=./exp/sachs/params.json --gfile=./data/sachs/sachs_bn_0.3.pkl

• for UCI datasets (boston housing and whitewine)

(BN Learn) python scripts/bn_learn.py --cfg_file=./exp/housing/params.json
(BN:ACFlow) python scripts/test_bn_dfa.py --cfg_file=./exp/housing/params.json --gfile=./exp/housing/bn_learn/results.pkl

(BN Learn) python scripts/bn_learn.py --cfg_file=./exp/whitewine/params.json
(BN:ACFlow) python scripts/test_bn_dfa.py --cfg_file=./exp/whitewine/params.json --gfile=./exp/whitewine/bn_learn/results.pkl

Time Series

• for digits dataset

(Train) python scripts/train.py --cfg_file=./exp/digits/params.json
(ACFlow DFA) python scripts/test_ts_dfa.py --cfg_file=./exp/digits/params.json
(Valid Prob.) python scripts/test_ts_prob.py --cfg_file=./exp/digits/params.json --split=valid
(Test Prob.) python scripts/test_ts_prob.py --cfg_file=./exp/digits/params.json --split=test
(Calibration) python scripts/calibrate.py --cfg_file=./exp/digits/params.json

• for pedestrian dataset

(Train) python scripts/train.py --cfg_file=./exp/pedestrian/params.json
(ACFlow DFA) python scripts/test_ts_dfa.py --cfg_file=./exp/pedestrian/params.json
(Valid Prob.) python scripts/test_ts_prob.py --cfg_file=./exp/pedestrian/params.json --split=valid
(Test Prob.) python scripts/test_ts_prob.py --cfg_file=./exp/pedestrian/params.json --split=test
(Calibration) python scripts/calibrate.py --cfg_file=./exp/pedestrian/params.json

Reference