Two word embedding mapping compatible with OpenAI gym.
Requirements:
- Python 3.5+
- OpenAI Gym
- NumPy
- Gensim
$ conda env create -f gym-embedding-world/environment.yml
$ source embedding-world
$ pip install -e gym-embedding-world/.
!git clone "https://github.com/SamirMoustafa/gym-embedding-world.git"
!pip install -e gym-embedding-world/.
!mv gym-embedding-world gym-embedding-world-org
!cp -r gym-embedding-world-org/embedding_world /content
!ls embedding_world
$ python >>> import gym
$ python >>> import embedding_world
$ python >>> env = gym.make('embedding_world-v0')
$ python >>> env.set_paths(embedding_from_file="... YOUR EMBEDDING PATH TO MAP FROM IT ...",
embedding_to_file ="... YOUR EMBEDDING PATH TO MAP TO IT .....")
$ python >>> env.production_is_off()
$ python >>> env.set_sentences('... YOUR SENTENCE TO TRANSLATE FROM IT ...',
'... YOUR SENTENCE TO TRANSLATE TO IT .....')
$ python >>> state, reward, done, info = env.step('dim(0)+1')
Embedding world is a simple environment based on OpenAI gym, that loads two-word embedding e.g. Stanfrod GloVe or facebook fastText models with N-dimension and moves from one word(s) embedding-location to the other embedding using an agent actions such that actions that could be taken are 2N + 1 actions {dimension(i)+1, dimension(i)-1} ∪ {pickup} ∀ i in range from 1 to N
which deterministically cause the corresponding state transitions but actions that would take an agent of the grid leave a state unchanged. The reward is negative for all transition until the goal is reached. The terminal state(goal) is represented in a vector/s.
This environment has been built as part of a graduation project at University of Alexandria, Department of Computer Science
Please use this bibtex if you want to cite this repository in your publications:
@misc{embedding_world,
author = {Samir Moustafa},
title = {Embedding Environment for OpenAI Gym},
year = {2019},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/SamirMoustafa/gym-embedding-world}}
}