This is an updated source code for paper Classification with Costly Features as a Sequential Decision-Making Problem wrote by Jaromír Janisch, Tomáš Pevný and Viliam Lisý: paper.

This version is enhanced with multiple options, namely:

• lagrangian optimization of lambda
• possibility of choosing an average or hard budget
• working with missing features (nans in the training data)
• reweighting of the dataset

For Classification with Costly Features using Deep Reinforcement Learning version, go to the master branch.

Prerequisites:

• cuda capable hardware
• ubuntu 16.04
• cuda 8/9
• python 3.6 (numpy, pandas, pytorch 0.4)

Usage:

• use tools tools/conv_*.py to prepare datasets; read the headers of those files; data is expected to be in ../data
• pretrained HPC models are in trained_hpc, or you can use tools/hpc_svm.py to recreate them; they are needed in ../data
• run python3.6 main.py [dataset] [target], choose dataset from config_datasets/
• set -target_type to lambda or cost, the latter automatically finds suitable lambda with lagrangian optimization (see this article)
• set -hard_budget for strict budget per sample (default is average budget)
• run python3.6 main.py --help for additional options
• the run will create multiple log files run*.dat
• you can use octave or matlab to analyze them with tools/debug.m
• you can also evaluate the agent on the test set with eval.py --dataset [dataset] --flambda [lambda]