def build_document_components(config):
"""Returns the document components of the simulator."""
# Create a finite pool of restaurants as candidates.
restaurants = [
Restaurant(idx, type_)
for idx, type_ in enumerate(config.restaurant_types)
]
# There is no need to sample restaurants repeatedly. We just want to use
# the same set over the course of the simulation. The SingletonSampler is
# useful for this.
return restaurants, recsim_samplers.SingletonSampler(
restaurants, Restaurant)
def create_gym_environment(env_config):
"""Returns a RecSimGymEnv with specified environment parameters.
Args:
env_config: an `EnvConfig` object.
Returns:
A RecSimGymEnv object.
"""
user_ctor = functools.partial(User, **attr.asdict(env_config.user_config))
initial_embeddings = utils.load_embeddings(env_config)
genre_history = utils.load_genre_history(env_config)
dataset = utils.Dataset(
env_config.data_dir,
user_ctor=user_ctor,
movie_ctor=Movie,
embeddings=initial_embeddings,
genre_history=genre_history,
genre_shift=env_config.genre_shift,
bias_against_unseen=env_config.bias_against_unseen)
document_sampler = recsim_samplers.SingletonSampler(dataset.get_movies(),
Movie)
user_sampler = recsim_samplers.UserPoolSampler(
seed=env_config.seeds.user_sampler,
users=dataset.get_users(),
user_ctor=user_ctor,
partitions=env_config.train_eval_test,
partition_seed=env_config.seeds.train_eval_test)
user_model = UserModel(
user_sampler=user_sampler,
seed=env_config.seeds.user_model,
slate_size=env_config.slate_size,
)
env = MovieLensEnvironment(
user_model,
document_sampler,
num_candidates=document_sampler.size(),
slate_size=env_config.slate_size,
resample_documents=False,
)
reward_aggregator = average_ratings_reward
return recsim_gym.RecSimGymEnv(env, reward_aggregator)
def create_gym_environment(env_config):
"""Returns a RecSimGymEnv with specified environment parameters.
Args:
env_config: an `EnvConfig` object.
Returns:
A RecSimGymEnv object.
"""
user_ctor = functools.partial(User, **attr.asdict(env_config.user_config))
initial_embeddings = movie_lens_utils.load_embeddings(env_config)
dataset = movie_lens_utils.Dataset(env_config.data_dir,
user_ctor=user_ctor,
movie_ctor=Movie,
embeddings=initial_embeddings)
document_sampler = recsim_samplers.SingletonSampler(
dataset.get_movies(), Movie)
if env_config.user_sampler_ctor == recsim_samplers.SequentialUserSampler:
user_sampler = env_config.user_sampler_ctor(
seed=env_config.seeds.user_sampler,
users=dataset.get_users(),
user_ctor=user_ctor)
else:
user_sampler = env_config.user_sampler_ctor(
seed=env_config.seeds.user_sampler,
users=dataset.get_users(),
user_ctor=user_ctor,
partitions=env_config.train_eval_test,
partition_seed=env_config.seeds.train_eval_test)
user_model = UserModel(
user_sampler=user_sampler,
seed=env_config.seeds.user_model,
)
env = MovieLensEnvironment(user_model,
document_sampler,
num_candidates=document_sampler.size(),
slate_size=1,
resample_documents=False)
reward_aggregator = functools.partial(
multiobjective_reward,
lambda_non_violent=env_config.lambda_non_violent)
return recsim_gym.RecSimGymEnv(env, reward_aggregator)
def _initialize_from_config(self, env_config):
self.working_dir = tempfile.mkdtemp(dir='/tmp')
self.initial_embeddings = movie_lens_utils.load_embeddings(env_config)
self.genre_history = movie_lens_utils.load_genre_history(env_config)
user_ctor = functools.partial(movie_lens.User,
**attr.asdict(env_config.user_config))
self.dataset = movie_lens_utils.Dataset(
env_config.data_dir,
user_ctor=user_ctor,
movie_ctor=movie_lens.Movie,
genre_history=self.genre_history,
embeddings=self.initial_embeddings,
genre_shift=env_config.genre_shift,
bias_against_unseen=env_config.bias_against_unseen)
self.document_sampler = recsim_samplers.SingletonSampler(
self.dataset.get_movies(), movie_lens.Movie)
self.user_sampler = recsim_samplers.UserPoolSampler(
seed=env_config.seeds.user_sampler,
users=self.dataset.get_users(),
user_ctor=user_ctor)
self.user_model = movie_lens.UserModel(
user_sampler=self.user_sampler,
seed=env_config.seeds.user_model,
slate_size=env_config.slate_size,
)
env = movie_lens.MovieLensEnvironment(
self.user_model,
self.document_sampler,
num_candidates=self.document_sampler.size(),
slate_size=env_config.slate_size,
resample_documents=False)
env.reset()
reward_aggregator = movie_lens.average_ratings_reward
self.env = recsim_gym.RecSimGymEnv(env, reward_aggregator)
def _initialize_from_config(self, env_config):
self.working_dir = tempfile.mkdtemp(dir=FLAGS.test_tmpdir)
self.initial_embeddings = movie_lens_utils.load_embeddings(env_config)
user_ctor = functools.partial(movie_lens.User,
**attr.asdict(env_config.user_config))
self.dataset = movie_lens_utils.Dataset(
env_config.data_dir,
user_ctor=user_ctor,
movie_ctor=movie_lens.Movie,
embeddings=self.initial_embeddings,
)
self.document_sampler = recsim_samplers.SingletonSampler(
self.dataset.get_movies(), movie_lens.Movie)
self.user_sampler = recsim_samplers.UserPoolSampler(
seed=env_config.seeds.user_sampler,
users=self.dataset.get_users(),
user_ctor=user_ctor,
)
self.user_model = movie_lens.UserModel(
user_sampler=self.user_sampler,
seed=env_config.seeds.user_model,
)
env = movie_lens.MovieLensEnvironment(
self.user_model,
self.document_sampler,
num_candidates=self.document_sampler.size(),
slate_size=1,
resample_documents=False,
)
env.reset()
reward_aggregator = functools.partial(
movie_lens.multiobjective_reward,
lambda_non_violent=env_config.lambda_non_violent,
)
self.env = recsim_gym.RecSimGymEnv(env, reward_aggregator)