def main(gamename, seed, algo, task_id):
env = gym.make(gamename)
print("start running task", task_id)
t0 = time.time()
# set seed
env.seed(seed)
torch.manual_seed(seed)
np.random.seed(seed)
state_dim, action_dim, action_lim = check_env(env)
replay_buffer = Data(args.buffersize)
if algo == 'ddpg':
trainer = DDPGTrainer(state_dim, action_dim, action_lim, replay_buffer)
elif algo == 'td3':
trainer = TD3Trainer(state_dim, action_dim, action_lim, replay_buffer)
else:
print("error algo")
return
frame_count = 0
timestep_since_eval = 0
reward_list = []
evaluations = []
for episode in range(args.max_episode):
trainer.init_episode()
obs = env.reset()
obs = torch.from_numpy(obs).reshape(1, state_dim).float()
reward_episode = 0
actor_loss_l = []
critic_loss_l = []
for i in range(args.T):
if frame_count < args.start_timesteps:
action = env.action_space.sample()
else:
action = trainer.get_exploration_action(obs)
new_obs, r, done, _ = env.step(action)
new_obs = torch.from_numpy(new_obs).reshape(1, state_dim).float()
reward_episode += r
sequence = [
obs,
torch.from_numpy(action).reshape(1, action_dim).float(),
reward_episode, new_obs, 0 if done else 1
]
replay_buffer.push(sequence)
obs = new_obs
frame_count += 1
timestep_since_eval += 1
if done:
break
trainer.optimize(i)
reward_list.append(reward_episode)
if timestep_since_eval > args.eval_freq:
timestep_since_eval %= args.eval_freq
evaluations.append(evaluate_policy(env, trainer))
trainer.save_model(gamename, evaluations, seed)
if frame_count > args.Tmax:
break
trainer.save_model(gamename, evaluations, seed)
return "Over"
def execute(self, p_conn):
debug_on_event = p_conn.get_debug_on_event()
data = Data()
data.set_data("debug_on_event", debug_on_event)
self.emit_("debug_on_event_update_request", data)
def get_movies_with_similar_genres(movie_id: int, n: int = 5, popularity_bias: bool = False
, user_bias: bool = False, movies: pd.DataFrame = None):
# Get all movies and split them into the base movie and the rest
if n is None:
n = 5
# Use the preferred movie df
if movies is None:
all_movies = Data.movie_meta()[Column.genres.value]
else:
all_movies = movies[Column.genres.value]
# get the base out of the df and remove it from the rest
base_genres = eval(all_movies.loc[movie_id])
all_movies = all_movies.drop(movie_id)
# count similar genres
all_movies = all_movies.apply(
lambda row: count_elements_in_set(row, base_genres)
)
# remove all movies which have no genre in common
filtered_movies_sum = all_movies[all_movies > 0]
# if user_bias is true
if user_bias:
# reduce the amount of movies to n * 10 movies
top_n_mul_ten = filtered_movies_sum.nlargest(n * 10)
ratings = Data.ratings()
# group by movie
ratings_grouped = ratings.groupby(str(Column.movie_id))
# calculate mean rating and number of ratings for each movie
# (select rating to remove first level of column index. before: (rating: (mean, count)), after: (mean, count) )
measures: pd.DataFrame = ratings_grouped.agg(['mean', 'count'])[str(Column.rating)]
# merging mean, count and genre sum into one DataFrame
measures_movies = pd.merge(measures, pd.DataFrame(top_n_mul_ten), left_index=True, right_index=True)
if popularity_bias:
# give more weight to the number of ratings (~popularity)
# by raising the avg ratings to some power (to preserve some notion of good vs. bad ratings)
# and multiplying the count back in
# additionally multiply the genre back in
# to prevent good rated movies with little correlation to the genres
results = measures_movies.eval('(mean ** 3) * count * genres')
else:
# multiply genre to prevent good rated movies with little correlation to the genres
results = measures_movies.eval('mean * genres')
else:
results = filtered_movies_sum
# breakpoint()
return results
def _check_dbwhere(self, p_curdbwhere):
#FIXME: documentacion
predbwhere = self.__predbwhere
if p_curdbwhere != predbwhere:
data = Data()
data.set_data("where", p_curdbwhere)
self.emit_("where_changed", data)
self.__predbwhere = p_curdbwhere
def _check_dbstack(self, p_curdbstack):
#FIXME: documentacion
predbstack = self.__predbstack
if p_curdbstack != predbstack:
data = Data()
data.set_data("stack", p_curdbstack)
self.emit_("stack_changed", data)
self.__predbstack = p_curdbstack
def execute(self, p_conn):
# FIXME: re-implementar este metodo haciendo
# uso de 'p_conn.is_file_in_loadpath'
funcname = self.__funcname
if funcname[1]:
real = "['%s', filemarker(), '%s']" % (funcname[0], funcname[1])
else:
real = "'%s'" % funcname[0]
dbstatus = p_conn.dbstatus(real)
lines = []
if dbstatus:
file_ = dbstatus[0]["file"]
if (funcname[1] or file_): # Esto es evitando funciones definidas
# en el CommandWindow.
if not file_:
file_ = p_conn.file_in_loadpath("'%s'" % self.__filename)
if self.__file == file_:
lines = dbstatus[0]["lines"]
data = Data()
data.set_data("file", self.__file)
data.set_data("function", funcname[1] if funcname[1] else funcname[0])
data.set_data("lines", lines)
self.emit_("breakpoints_update_request", data)
def _check_dbstatus(self, p_curdbstatus):
#FIXME: documentacion
#OJO: Puede que cambien los breakpoints y no se emita la sennal.
# Esto no es problema, porque no tiene solucion en Octave_3.2.3.
predbstatus = self.__predbstatus
if p_curdbstatus != predbstatus:
data = Data()
data.set_data("breakpoints", p_curdbstatus)
self.emit_("breakpoints_changed", data)
self.__predbstatus = p_curdbstatus
def get_year_relevance(movie_id:int, n:int=0):
release_years= Data.movie_meta()[Column.release_year.value]
movie_year=release_years.loc[movie_id]
release_years = release_years.subtract(movie_year)
release_years = release_years.abs()
release_years=1-release_years.div(release_years.max())
return release_years.drop(movie_id)
def build_index(cls):
if cls.ix is None:
cls.init()
# automatically calls iw.commit()
iw = cls.ix.writer()
for movie_id, movie in Data.movie_meta().iterrows():
# extract fields
fields: Dict = {
'movie_id': movie_id,
'title': movie[Column.title.value],
# 'tagline': movie[Column.tagline.value],
# 'summary': movie[Column.summary.value],
# 'keywords': movie[Column.keywords.value],
# 'popularity': movie[Column.num_ratings.value],
# 'genres': movie[Column.genres.value],
}
# filter empty values (inserting fails for np.nan values)
fields = {
key: val
for key, val in fields.items()
if val is not None and val is not np.nan and val != ''
}
# insert into index
iw.update_document(**fields)
iw.commit(optimize=True)
def get_imdb_id(movielens_id: int) -> int:
movies = Data.movie_meta()
if movielens_id not in movies.index:
raise MovieNotFoundException()
movie = movies.loc[movielens_id]
return movie[Column.imdb_id.value]
def get_movie_meta_for(movie_ids: List[int]) -> List[Dict]:
# if single movie, pack into list
if isinstance(movie_ids, int):
movie_ids = [movie_ids]
movie_ids = filter(lambda x: x is not None, movie_ids)
meta: pd.DataFrame = Data.movie_meta()
try:
# filter metadata
meta = meta.loc[movie_ids]
except KeyError as e:
raise MovieNotFoundException(e.args)
# fetch metadata for the movies, convert to dictionary
# orientation='records' results in [{'col1': 'val1', 'col2': 'val2'}, {'col1': 'val1', ..}]
meta_dict: List[Dict] = meta.to_dict(orient='records')
for item in meta_dict:
for col in [
Column.actors, Column.genres, Column.keywords, Column.directors
]:
if not pd.isnull(item[col.value]):
item[col.value] = eval(item[col.value])
add_poster_urls(meta_dict)
return meta_dict
def uuid_str():
'''
Return a [Base58 encoded][1] UUID that can be used to opaquely identify
applications, sessions, and users to external clients
[1]: https://en.wikipedia.org/wiki/Base58
'''
return Data(uuid4().bytes).stringWithEncoding(Base58)
def get_similarities_for(cls, movie_id: int, colname: str):
# get similarity matrix (calculate if necessary)
sim_matrix = cls.calculate_similarities(colname)
# get absolute index of movie
index = Data.movie_meta().index.get_loc(movie_id)
# get similarities for this movie
# use .toarray() to convert from sparse matrix
# use [0] to convert "matrix" with only one row to one-dimensional array
similarities = sim_matrix[index].toarray()[0]
# put into pandas Series
# use index=... to apply original index
series = pd.Series(index=Data.movie_meta().index, data=similarities)
return series.drop(movie_id)
def __call__(self, movie_id: int, n: int = 5):
meta = Data.movie_meta()
collection = meta[get_collection_mask(movie_id, meta)].index.values
results: pd.Series = self.method(movie_id, n + 10)
results = results.drop(collection, errors='ignore')
return results
def get_normalized_popularity():
# used for popularity bias
popularity = Data.movie_meta()[Column.num_ratings.value]
# apply root reduce linearity
# (if movie A has double the ratings of movie B, its popularity should only be slightly higher)
popularity **= (1 / 10)
# normalize
popularity /= popularity.max()
return popularity
def avg_rating_for_user(user_id: int) -> float:
""" Calculates the average score for ratings from a specified user. """
ratings: pd.Series = Data.ratings_as_series()
# check if user_id exists and raise exception if it does not
if user_id not in ratings:
raise UserNotFoundException()
# select ratings from specified user_id
user_ratings: pd.Series = ratings.loc[user_id]
# calculate average using integrated function
return user_ratings.mean()
def calculate_similarities(cls,
colname: str,
overwrite_existing: bool = False):
if colname not in cls.similarity_matrices or overwrite_existing:
# calculate tf_idf for column
tfidf_matrix = cls.tf_idf.fit_transform(
Data.movie_meta()[colname].fillna(''))
# calculate similarities between movies
# use dense_output=False (results in sparse matrix) to reduce memory usage
cls.similarity_matrices[colname] = linear_kernel(
tfidf_matrix, tfidf_matrix, dense_output=False)
return cls.similarity_matrices[colname]
def get_movielens_id(tmdb_id: int = None, imdb_id: int = None) -> int:
movies: pd.DataFrame = Data.movie_meta()
if tmdb_id is not None:
movie = movies.query(f'{Column.tmdb_id.value} == {tmdb_id}')
elif imdb_id is not None:
movie = movies.query(f'{Column.imdb_id.value} == {imdb_id}')
else:
return None
if movie.empty:
return None
return movie.index[0]
def tmdb_reference(movie_id: int, n: int = 5):
movie = Data.movie_meta().loc[movie_id]
# get list from string representation
similar_tmdb = eval(movie[Column.tmdb_similar.value])
# get movielens id from tmdb_id
similar = map(lambda tmdb_id: get_movielens_id(tmdb_id=tmdb_id),
similar_tmdb)
# return with artificial decreasing score
return pd.Series({
item: -index
for index, item in enumerate(similar) if item is not None
})
def recommend_movies(movie_id: int, n: int = 5, filter_below_avg_ratings: bool = False, popularity_bias: bool = False) \
-> List[int]:
ratings = Data.ratings()
# first get the ratings for the base movie
ratings_of_base_movie = ratings.query('movie_id == %s' % movie_id)
# check if there are reviews for this movie
if ratings_of_base_movie.empty:
raise MissingDataException('no ratings for movie_id %s' % movie_id)
if filter_below_avg_ratings:
# of those, select the above average ratings
avg_rating = ratings_of_base_movie['rating'].mean()
# query is actually faster than the python subscription syntax ( users[users['rating'] >= avg] )
ratings_of_base_movie = ratings_of_base_movie.query('rating >= %f' %
avg_rating)
# to get ratings from all the users that have rated/liked the base movie,
# perform a (left outer) join on all the ratings on user_id
relevant_movies = ratings_of_base_movie.join(ratings,
on='user_id',
lsuffix='_L')
# remove the columns that were duplicated as result of the join
relevant_movies = relevant_movies[['movie_id', 'rating']]
# remove the base movie from the results
relevant_movies = relevant_movies.query('movie_id != %s' % movie_id)
if relevant_movies.empty:
raise MissingDataException(
'no other ratings from users that rated movie_id %s' % movie_id)
# group by movie
relevant_movie_groups = relevant_movies.groupby('movie_id')
# calculate mean rating and number of ratings for each movie
# (select rating to remove first level of column index. before: (rating: (mean, count)), after: (mean, count) )
measures: pd.DataFrame = relevant_movie_groups.agg(['mean',
'count'])['rating']
if popularity_bias:
# give more weight to the number of ratings (~popularity)
# by raising the avg ratings to some power (to preserve some notion of good vs. bad ratings)
# and multiplying the count back in
results = measures.eval('(mean ** 3) * count')
else:
results = measures['mean']
return results
def _recommend_movies(movie_id: int, n: int, method: Method) -> List[Dict]:
if movie_id not in Data.movie_meta().index:
raise MovieNotFoundException
# start with the movie itself
movies: List[int] = [movie_id]
# calculate similarities
scores: Series = method(movie_id)
# and filter out any movies that were recommended recently
scores = History.filter(scores)
if method == Method.reference or method == Method.sequels:
n = 20
# movies = [base_movie, ...recommendations]
movies.extend(scores.nlargest(n).index)
# add recommendations for movies
History.append(movies)
return get_movie_meta_for(movies)
def search(cls, query_text: str, n: int, add_posters: bool = True):
# this method applies a popularity bias to search results
# as they need to be resorted, more search terms should be provided than necessary,
# to be able to recover popular results that have rather low scores
results = cls._search(query_text, n + 25)
# encapsulate in pandas.Series for further operations
scores = pd.Series(results, name='score')
# perform a (right outer) join to connect the search results to the metadata
df = Data.movie_meta().join(scores, how='right')
# calculate the weighted score by raising it to some power
# in order for the popularity to not overpower the score completely
# and multiply with the number of ratings (the popularity)
df.eval(f'weighted = score**16 * {Column.num_ratings.value}',
inplace=True)
# extract the n best results and export as list
movie_ids = list(df.nlargest(n, 'weighted').index)
# fetch metadata
meta = get_movie_meta_for(movie_ids)
return meta
def get_collection(movie_id: int,
df: pd.DataFrame = None,
include_base_movie: bool = True,
start_from_base_movie: bool = False,
wrap_to_start: bool = False) -> pd.DataFrame:
"""
Get movies from a collection.
:param movie_id: a movie that is in a collection
:param df: the pandas DataFrame to search
:param include_base_movie: whether to include movie_id itself in the result
:param start_from_base_movie: whether to split the result and start at movie_id
:param wrap_to_start: if start_from_base_movie: at the end of the collection, wrap over to the start and include the prequels
:return: a DataFrame containing the movies in the collection
"""
if df is None:
df = Data.movie_meta()
# select movies that are in collection
m = df[get_collection_mask(movie_id, df)]
# sort by release year
m = m.sort_values(by=Column.release_date.value)
if not include_base_movie:
m = m.drop(movie_id)
if start_from_base_movie:
# split dataframe at base_movie
sequels = m.loc[movie_id:]
prequels = m.loc[:movie_id - 1]
if wrap_to_start:
# reverse order and join again
m = pd.concat([sequels, prequels])
else:
# just return the movies starting with the base movie
m = sequels
return m
def __init__(self, p_mwindow):
"""
p_mwindow: un 'MainWindow'.
Retorna: una 'Connection'.
Crea una nueva 'Connection'.
"""
threading.Thread.__init__(self)
gobject.GObject.__init__(self)
self.__mwindow = p_mwindow
self.__continue = True
self.__term = Terminal(self)
# Aqui se almacenan los comandos
# que envia el usuario a octave.
self.__tail = []
# Aqui se almacenan los comandos que se
# ejecutaran cuando el octave este listo.
self.__tail_prio = TailWithPriority()
self.__dbstatus_code = self.__get_dbstatus_code()
self.__dbstack_code = self.__get_dbstack_code()
self.__prestate = None
self.__predbstack = None
self.__predbwhere = None
self.__predbstatus = None
self.__predir = None
#self.__prevars = None# para emitir un "vars_changed"
flag = Data()
key = "key"
self.connect("changed", self.__on_changed, flag, key)
self.connect("stack_changed", self.__on_stack_changed, flag, key)
self.connect("where_changed", self.__on_where_changed, flag, key)
def actors_as_lists():
# since eval (convert string representation to object) is costly time-wise, cache results
return Data.movie_meta()[Column.actors.value].map(eval)
def execute(self, p_conn):
if "is patch for Octave-3.2.3":
dbwhere = p_conn.dbwhere() if p_conn.dbstack()["frames"] else {}
#dbwhere = p_conn.dbwhere() # Puede hacer un beep!
file_ = self.__file
positions = []
current = False
if dbwhere and dbwhere["file"] == file_:
pos = (dbwhere["line"], dbwhere["column"])
if None not in pos:
positions.append(pos)
current = True
data = Data()
data.set_data("positions", positions)
data.set_data("current", current)
self.emit_("stack_update_request", data)
return
file_ = self.__file
omit = self.__omit
omit = str(omit) if omit else ""
dbstack = p_conn.dbstack(omit)
positions = []
frames = dbstack["frames"]
current = bool(frames and frames[0]["file"] == file_)
for frame in frames:
if frame["file"] == file_:
positions.append((frame["line"], frame["column"]))
data = Data()
data.set_data("positions", positions)
data.set_data("current", current)
self.emit_("stack_update_request", data)
logging.basicConfig(
level=logging.INFO,
format=
'%(asctime)s %(filename)s[line:%(lineno)d] %(levelname)s %(message)s',
datefmt='%a, %d %b %Y %H:%M:%S',
)
if len(sys.argv) >= 2:
dataIdx = '{:0>2}'.format(sys.argv[1])
else:
dataIdx = 'CoNLL'
trainPath = 'data/normal/en_train_{}.txt'.format(dataIdx)
testPath = 'data/normal/en_test_CoNLL.txt'
data = Data(inputPathList=[trainPath], testPath=testPath)
return_data = data.loadCoNLL(trainPath, loadFeatures=True)
split_data = train_test_split(*return_data, test_size=0.1, random_state=0)
X_train = split_data[:-2:2]
X_val = split_data[1:-2:2]
y_train, y_val = split_data[-2:]
modelWrapper = BiLSTMCRF(data)
model = modelWrapper.buildModel(feature2idx=data.feature2idx)
history = metricHistory(X_val, y_val, saveDir=dataIdx)
history.set_model(model)
model.fit(X_train,
y_train,
epochs=50,
from util.plot import Plot, multiplot
from util.stats import cdf_fit_func
import numpy as np
from util.data import Data
# d = Data.load("knn_results_[yelp-mnist].pkl")
d = Data.load("prediction_results.pkl")
names = d.names.copy()
names = [names[0], names[6], names[1]] + names[2:6] + names[7:]
d.reorder([names[0], names[6], names[1]])
d.sort()
d["Errors"] = ([float(v) if (v <= 5) else 5.0 for v in e] for e in d["Errors"])
d["Mean Error"] = (sum(e) / len(e) for e in d["Errors"])
d["Mean Squared Error"] = (sum(v**2 for v in e) / len(e) for e in d["Errors"])
d["Error Variance"] = (float(np.var(e)) for e in d["Errors"])
d._max_display = 1000
print(d)
all_data = d
# Get the unique dimensions and algorithms.
dims = sorted(set(all_data["Dimension"]))
algs = sorted(set(all_data["Algorithm"]))
data_sets = sorted(set(all_data["Data"]))
for ds in data_sets:
for alg in algs:
d = all_data[all_data["Data"] == ds]
d = d[d["Algorithm"] == alg]
min_index = int(np.argmin(d["Mean Error"]))
m = d[min_index, "Method"]
def directors_as_lists():
return Data.movie_meta()[Column.directors.value].map(eval)
def get_genre_as_lists():
return Data.movie_meta()[Column.genres.value].map(eval)