def parse_movielens(threshold=4, **kwargs):
if os.path.isfile(BIN_DATA[ML_20M_ALT]):
LOG.info("Already processed, skipping.")
return
source_file = os.path.join(DOWNLOAD[ML_20M_ALT], "ratings.csv")
if not glob(source_file):
download_movielens()
LOG.info("Parsing movielens.")
df = pd.read_csv(source_file)
df.drop('timestamp', axis=1, inplace=True)
df["rating"] = make_feedback_implicit(df["rating"], threshold)
map_user_id = {u: i for i, u in enumerate(df.userId.unique())}
map_movie_id = {m: i for i, m in enumerate(df.movieId.unique())}
m_sp = sp.csr_matrix(
(df.rating,
([map_user_id[u] for u in df.userId],
[map_movie_id[m] for m in df.movieId])),
shape=(len(map_user_id), len(map_movie_id))
)
m_sp.eliminate_zeros()
save_as_npz(m_sp, BIN_DATA[ML_20M_ALT])
def download_movielens():
filepath = os.path.join(DATASETS_DIR, ML_20M_ALT + '.zip')
if not glob(filepath):
download_file(DOWNLOAD_URL[ML_20M], filepath)
LOG.info("Extracting")
extract_file(filepath, DATASETS_DIR)
def download_pinterest():
filepath = os.path.join(DATASETS_DIR, PINTEREST + '.zip')
if not glob(filepath):
download_file_from_google_drive(DOWNLOAD_GOOGLE_DRIVE_ID[PINTEREST],
filepath)
LOG.info("Extracting")
extract_file(filepath, DATASETS_DIR)
os.rename(os.path.join(DATASETS_DIR, 'pinterest_iccv'),
DOWNLOAD[PINTEREST])
def download_lastfm():
filepath = os.path.join(DATASETS_DIR, 'lastfm-dataset-360K.tar.gz')
if not glob(DOWNLOAD[LASTFM]):
download_file(DOWNLOAD_URL[LASTFM], filepath)
LOG.info("Extracting")
extract_file(filepath, DATASETS_DIR)
os.rename(os.path.join(DATASETS_DIR, 'lastfm-dataset-360K'),
os.path.join(DATASETS_DIR, 'lastfm'))
def download_netflix():
filepath = os.path.join(DATASETS_DIR, NETFLIX + '.tar.gz')
if not glob(filepath):
download_file(DOWNLOAD_URL[NETFLIX], filepath)
LOG.info("Extracting 1/2")
extract_file(filepath, tempfile.gettempdir())
LOG.info("Extracting 2/2")
extract_file(
os.path.join(tempfile.gettempdir(), 'download', 'training_set.tar'),
DATASETS_DIR)
os.rename(os.path.join(DATASETS_DIR, 'training_set'), DOWNLOAD[NETFLIX])
def train(
self,
n_epochs: int,
train_data: sparse.csr_matrix,
validation_data_input: sparse.csr_matrix,
validation_data_true: sparse.csr_matrix,
batch_size_train: int,
batch_size_validation: int,
metrics: dict, # Dict[str, matrix -> matrix -> float]
validation_step: 10,
):
"""
Train the model
:param n_epochs: number of epochs
:param train_data: train matrix of shape users count x items count
:param metrics: Dictionary of metric names to metric functions
:param validation_step: If it's set to n then validation is run once every n epochs
"""
self.metrics_history = defaultdict(lambda: [])
self.time_elapsed_training_history = []
self.time_elapsed_validation_history = []
self.session.run(self.iter.initializer)
for epoch in range(1, n_epochs + 1):
self.log_which_epoch(epoch, n_epochs)
init_time = time.time()
for _ in range(self.n_batch_per_train):
self.session.run(self.optimizer)
training_duration = time.time() - init_time
self.time_elapsed_training_history.append(training_duration)
LOG.info("Train time:\t{}".format(training_duration))
if epoch % validation_step == 0 or epoch == n_epochs:
init_time = time.time()
metrics_scores = self.test(validation_data_input,
validation_data_true, metrics)
for name, score in metrics_scores.items():
self.metrics_history[name].append(score)
validation_duration = time.time() - init_time
self.time_elapsed_validation_history.append(
validation_duration)
LOG.info("Valid time:\t{}".format(validation_duration))
self.log_metrics(epoch, metrics_scores, n_epochs)
self.log_training_time()
def log_training_time(self):
LOG.info("Total elapsed train time: {}".format(
np.sum(self.time_elapsed_training_history)))
LOG.info("Total elapsed valid time: {}".format(
np.sum(self.time_elapsed_validation_history)))
LOG.info("Epoch average train time: {}".format(
np.mean(self.time_elapsed_training_history)))
LOG.info("Epoch average valid time: {}".format(
np.mean(self.time_elapsed_validation_history)))
def parse_pinterest(**kwargs):
if os.path.isfile(BIN_DATA[PINTEREST]):
LOG.info("Already processed, skipping.")
return
data_file = 'subset_iccv_board_pins.bson'
source_file = os.path.join(DOWNLOAD[PINTEREST], data_file)
if not glob(source_file):
raise Exception("Cannot find pinterest dataset")
LOG.info("Parsing pinterest")
with open(source_file, 'rb') as f:
bsob = bson.decode_all(f.read())
map_id_pin = dict()
map_pin_id = dict()
map_board_id = dict()
map_id_board = dict()
pins = 0
board_pin_pairs = []
for i, board in enumerate(bsob):
map_id_board[i] = board
map_board_id[board['board_id']] = i
for pin in board['pins']:
if (pin not in map_pin_id):
map_pin_id[pin] = pins
map_id_pin[pins] = pin
pins += 1
board_pin_pairs.append((map_board_id[board['board_id']], map_pin_id[pin]))
boards = [board for (board, pin) in board_pin_pairs]
pins = [pin for (board, pin) in board_pin_pairs]
m_sp = sp.csr_matrix(([1] * len(boards), (boards, pins)), shape=(len(map_board_id), len(map_pin_id)))
save_as_npz(m_sp, BIN_DATA[PINTEREST])
def parse_lastfm(**kwargs):
if os.path.isfile(BIN_DATA[LASTFM]):
LOG.info("Already processed, skipping.")
return
data_file = 'usersha1-artmbid-artname-plays.tsv'
source_file = os.path.join(DOWNLOAD[LASTFM], data_file)
if not glob(source_file):
download_lastfm()
LOG.info("Parsing lastfm")
df = pd.read_csv(source_file, delimiter='\t', names=["User", "Artist id", "Artist name", "Plays"], dtype=str)
artist_column = list(zip([str(i) for i in df['Artist id']], [str(i) for i in df['Artist name']]))
user_column = df['User']
map_artist_id = {artist: i for i, artist in enumerate(sorted(set(artist_column)))}
map_user_id = {user: i for i, user in enumerate(sorted(set(user_column)))}
user_ids = [map_user_id[user] for user in user_column]
artist_ids = [map_artist_id[artist] for artist in artist_column]
m_sp = sp.csr_matrix(([1] * df.shape[0], (user_ids, artist_ids)), shape=(len(map_user_id), len(map_artist_id)))
save_as_npz(m_sp, BIN_DATA[LASTFM])
def load_dataset(dataset: str, *args, **kwargs):
"""
Generic data loader.
:param dataset: name of dataset to be loaded
:return: 5 csr_matrices {train, valid_in, valid_out, test_in, test_out}
"""
assert dataset in DATASETS, "Wrong dataset name"
if dataset == ML_20M:
out = load_and_parse_ML_20M(DOWNLOAD[ML_20M], *args, **kwargs)
LOG.info("Done")
return out
handler_map_parse = {
NETFLIX: parse_netflix,
ML_20M_ALT: parse_movielens,
LASTFM: parse_lastfm,
PINTEREST: parse_pinterest
}
handler_map_parse[dataset]()
out = load_data(BIN_DATA[dataset], *args, **kwargs)
LOG.info("Done")
return out
def parse_netflix(threshold=3, **kwargs):
if os.path.isfile(BIN_DATA[NETFLIX]):
LOG.info("Already processed, skipping.")
return
files = glob(os.path.join(DOWNLOAD[NETFLIX], '*'))
if not files:
download_netflix()
LOG.info("Parsing netflix")
users = get_users(files)
map_user_id = {u: i for i, u in enumerate(users)}
csr_rows = []
csr_columns = []
csr_data = []
LOG.info("Parsing netflix, step 2/2")
for movie_id, file_path in tqdm(enumerate(files)):
df = pd.read_csv(file_path, names=['User', 'Rating', 'Date'])
df.drop(0, inplace=True)
df['Rating'] = make_feedback_implicit(df['Rating'], threshold)
rows = [map_user_id[user] for user in df['User']]
columns = [movie_id] * len(rows)
data = list(df['Rating'])
assert len(rows) == len(columns) and len(columns) == len(data)
csr_rows += rows
csr_columns += columns
csr_data += data
m_sp = sp.csr_matrix((csr_data, (csr_rows, csr_columns)), shape=(len(users), len(files)))
m_sp.eliminate_zeros()
save_as_npz(m_sp, BIN_DATA[NETFLIX])
def log_which_epoch(self, epoch, n_epochs):
LOG.info("Epoch: {}".format(epoch))
def log_metrics(self, epoch, metrics_scores, n_epochs):
for name, score in metrics_scores.items():
LOG.info("Mean {}:\t{}".format(name, score))