This is the code used to create results for paper Classification with Costly Features in Hierarchical Deep Sets.

Directory structured as follows:

• code: contains the code and datasets
• vis_tool: javascript tool to visualize the agent's behavior
• experiments: scripts to perform the experiments and produce the plots

In the code folder, there are three algorithms rl (the main agent), rw (random sampling), and mil (the hmil agent with all information) in their corresponding directories. The data folder contains preprocessed datasets.

In each directory, the main.py is the main script. Run agent_rl as: python main.py [dataset] [target lambda] i.e. python main.py recipes 0.001

Run agent_rw as: python main.py [dataset] [target budget] i.e. python main.py recipes 5.0

Run agent_mil as: python main.py [dataset] i.e. python main.py recipes

Note that toy_b dataset needs to be run with an additional argument -dataseed 0.

With a trained rl model, you can create a json for visualization in the vis_tool with agent_rl/eval_vis.py tool. Run as: python eval_vis.py [dataset] -model [model file]

The code requires numpy, pytorch and sklearn libraries.