This repository contains an accompanying source code and experiments' results for this paper.
Folder experiments contains the code for different agents. It's built around 5 main python files:
parameters.pycontains definitions for the model: hyperparameters, seed, definition of the environment/insturctions dataset/layouts, and etc.train.pycontains logic for training the agent. The training is assumed to be started by callingstart_training. All the parameters are taken fromparameters.pyvisualize.pycontains logic for visualising the agent. All the parameters are taken fromparameters.pyevaluation.pycontains logic for evaluating the agent. All the parameters are taken fromparameters.pymodel.pycontains the model of the network. All the parameters are taken fromparameters.py
In order to train, evaluate, or visualise any of the agents, you need to import the target function/module. Few examples:
- To visualise the Dueling DQN + Gated-Attention agent
from experiments.dueling_dqn_ga import visualize
- To start training of the Dueling DQN + Concatenation agent
from experiments.dueling_dqn_cat import train
train.start_training()
- To benchmark the Dueling DQN + Concatenation agent
from experiments.dueling_dqn_cat import evaluate
evaluate.benchmark_all()
Every experiment is saved into its own folder, which is defined in terms of the parameters in parameters.py. All the training data (tensorboard logs), benchmarks (benchmarks.json), and the best model (best.model) are inside of it.
If you want to run many experiments with different parameters, consider the file runner_hypothesis_synonyms.py. It shows how you could do it (but it's a bit messy).
It's recommended to use Linux-based system, as the most Windows systems are not friendly with long paths.
Just execute build_run.sh, and it will build a container for running training or benchmarking jobs (visualiztion should be run locally).