def fit(self,
epochs=100,
learning_rate_vae=1e-2,
learning_rate_cvae=1e-3,
num_factors=50,
dimensions_vae=[200, 100],
epochs_vae=[50, 50],
batch_size=128,
lambda_u=0.1,
lambda_v=10,
lambda_r=1,
a=1,
b=0.01,
M=300,
temp_file_folder="result_experiments\CollaborativeVAE_RecommenderWrapper",
**earlystopping_kwargs):
if temp_file_folder is None:
print("{}: Using default Temp folder '{}'".format(
self.RECOMMENDER_NAME, self.DEFAULT_TEMP_FILE_FOLDER))
self.temp_file_folder = self.DEFAULT_TEMP_FILE_FOLDER
else:
print("{}: Using Temp folder '{}'".format(self.RECOMMENDER_NAME,
temp_file_folder))
self.temp_file_folder = temp_file_folder
if not os.path.isdir(self.temp_file_folder):
os.makedirs(self.temp_file_folder)
print("{}: Pretraining with VariationalAutoEncoder".format(
self.RECOMMENDER_NAME))
random_seed = 0
np.random.seed(random_seed)
tf.set_random_seed(random_seed)
activations = ['sigmoid', 'sigmoid']
init_logging(self.temp_file_folder + "_log.txt")
logging.info('loading data')
_, input_dim = self.ICM.shape
self.ICM = self.ICM.toarray()
idx = np.random.rand(self.ICM.shape[0]) < 0.8
train_X = self.ICM[idx, :]
test_X = self.ICM[~idx, :]
logging.info('initializing sdae model')
tf.reset_default_graph()
model = VariationalAutoEncoder(input_dim=input_dim,
dims=dimensions_vae,
z_dim=num_factors,
activations=activations,
epoch=epochs_vae,
noise='mask-0.3',
loss='cross-entropy',
lr=learning_rate_vae,
batch_size=batch_size,
print_step=1,
weight_path=self.temp_file_folder +
"_pretrain")
logging.info('fitting data starts...')
model.fit(train_X, test_X)
print("{}: Finalizing with VariationalAutoEncoder".format(
self.RECOMMENDER_NAME))
self._params = Params()
self._params.lambda_u = lambda_u
self._params.lambda_v = lambda_v
self._params.lambda_r = lambda_r
self._params.a = a
self._params.b = b
self._params.M = M
self._params.n_epochs = epochs
# These are the train instances as a list of lists
self._train_users = []
self.URM_train = sps.csr_matrix(self.URM_train)
for user_index in range(self.n_users):
start_pos = self.URM_train.indptr[user_index]
end_pos = self.URM_train.indptr[user_index + 1]
user_profile = self.URM_train.indices[start_pos:end_pos]
self._train_users.append(list(user_profile))
self._train_items = []
self.URM_train = sps.csc_matrix(self.URM_train)
for user_index in range(self.n_items):
start_pos = self.URM_train.indptr[user_index]
end_pos = self.URM_train.indptr[user_index + 1]
item_profile = self.URM_train.indices[start_pos:end_pos]
self._train_items.append(list(item_profile))
self.URM_train = sps.csr_matrix(self.URM_train)
tf.reset_default_graph()
self.model = CVAE(num_users=self.n_users,
num_items=self.n_items,
num_factors=num_factors,
params=self._params,
input_dim=input_dim,
dims=dimensions_vae,
n_z=num_factors,
activations=activations,
loss_type='cross-entropy',
lr=learning_rate_cvae,
random_seed=random_seed,
print_step=10,
verbose=False)
#weight_path = "model/pretrain"
self.model.load_model(weight_path=self.temp_file_folder + "_pretrain")
self._update_best_model()
self.model.m_theta[:] = self.model.transform(self.ICM)
self.model.m_V[:] = self.model.m_theta
self._train_with_early_stopping(epochs,
algorithm_name=self.RECOMMENDER_NAME,
**earlystopping_kwargs)
print("{}: Tranining complete".format(self.RECOMMENDER_NAME))
self.USER_factors = self.USER_factors_best
self.ITEM_factors = self.ITEM_factors_best
if self.temp_file_folder == self.DEFAULT_TEMP_FILE_FOLDER:
print("{}: cleaning temporary files".format(self.RECOMMENDER_NAME))
shutil.rmtree(self.DEFAULT_TEMP_FILE_FOLDER, ignore_errors=True)
class CollaborativeVAE_RecommenderWrapper(BaseMatrixFactorizationRecommender,
Incremental_Training_Early_Stopping):
RECOMMENDER_NAME = "CollaborativeVAE_RecommenderWrapper"
DEFAULT_TEMP_FILE_FOLDER = './result_experiments/__Temp_CollaborativeVAE_RecommenderWrapper/'
def __init__(self, URM_train, ICM):
super(CollaborativeVAE_RecommenderWrapper, self).__init__(URM_train)
self.ICM = check_matrix(ICM.copy(), 'csr', dtype=np.float32)
def fit(self,
epochs=100,
learning_rate_vae=1e-2,
learning_rate_cvae=1e-3,
num_factors=50,
dimensions_vae=[200, 100],
epochs_vae=[50, 50],
batch_size=128,
lambda_u=0.1,
lambda_v=10,
lambda_r=1,
a=1,
b=0.01,
M=300,
temp_file_folder="result_experiments\CollaborativeVAE_RecommenderWrapper",
**earlystopping_kwargs):
if temp_file_folder is None:
print("{}: Using default Temp folder '{}'".format(
self.RECOMMENDER_NAME, self.DEFAULT_TEMP_FILE_FOLDER))
self.temp_file_folder = self.DEFAULT_TEMP_FILE_FOLDER
else:
print("{}: Using Temp folder '{}'".format(self.RECOMMENDER_NAME,
temp_file_folder))
self.temp_file_folder = temp_file_folder
if not os.path.isdir(self.temp_file_folder):
os.makedirs(self.temp_file_folder)
print("{}: Pretraining with VariationalAutoEncoder".format(
self.RECOMMENDER_NAME))
random_seed = 0
np.random.seed(random_seed)
tf.set_random_seed(random_seed)
activations = ['sigmoid', 'sigmoid']
init_logging(self.temp_file_folder + "_log.txt")
logging.info('loading data')
_, input_dim = self.ICM.shape
self.ICM = self.ICM.toarray()
idx = np.random.rand(self.ICM.shape[0]) < 0.8
train_X = self.ICM[idx, :]
test_X = self.ICM[~idx, :]
logging.info('initializing sdae model')
tf.reset_default_graph()
model = VariationalAutoEncoder(input_dim=input_dim,
dims=dimensions_vae,
z_dim=num_factors,
activations=activations,
epoch=epochs_vae,
noise='mask-0.3',
loss='cross-entropy',
lr=learning_rate_vae,
batch_size=batch_size,
print_step=1,
weight_path=self.temp_file_folder +
"_pretrain")
logging.info('fitting data starts...')
model.fit(train_X, test_X)
print("{}: Finalizing with VariationalAutoEncoder".format(
self.RECOMMENDER_NAME))
self._params = Params()
self._params.lambda_u = lambda_u
self._params.lambda_v = lambda_v
self._params.lambda_r = lambda_r
self._params.a = a
self._params.b = b
self._params.M = M
self._params.n_epochs = epochs
# These are the train instances as a list of lists
self._train_users = []
self.URM_train = sps.csr_matrix(self.URM_train)
for user_index in range(self.n_users):
start_pos = self.URM_train.indptr[user_index]
end_pos = self.URM_train.indptr[user_index + 1]
user_profile = self.URM_train.indices[start_pos:end_pos]
self._train_users.append(list(user_profile))
self._train_items = []
self.URM_train = sps.csc_matrix(self.URM_train)
for user_index in range(self.n_items):
start_pos = self.URM_train.indptr[user_index]
end_pos = self.URM_train.indptr[user_index + 1]
item_profile = self.URM_train.indices[start_pos:end_pos]
self._train_items.append(list(item_profile))
self.URM_train = sps.csr_matrix(self.URM_train)
tf.reset_default_graph()
self.model = CVAE(num_users=self.n_users,
num_items=self.n_items,
num_factors=num_factors,
params=self._params,
input_dim=input_dim,
dims=dimensions_vae,
n_z=num_factors,
activations=activations,
loss_type='cross-entropy',
lr=learning_rate_cvae,
random_seed=random_seed,
print_step=10,
verbose=False)
#weight_path = "model/pretrain"
self.model.load_model(weight_path=self.temp_file_folder + "_pretrain")
self._update_best_model()
self.model.m_theta[:] = self.model.transform(self.ICM)
self.model.m_V[:] = self.model.m_theta
self._train_with_early_stopping(epochs,
algorithm_name=self.RECOMMENDER_NAME,
**earlystopping_kwargs)
print("{}: Tranining complete".format(self.RECOMMENDER_NAME))
self.USER_factors = self.USER_factors_best
self.ITEM_factors = self.ITEM_factors_best
if self.temp_file_folder == self.DEFAULT_TEMP_FILE_FOLDER:
print("{}: cleaning temporary files".format(self.RECOMMENDER_NAME))
shutil.rmtree(self.DEFAULT_TEMP_FILE_FOLDER, ignore_errors=True)
def _prepare_model_for_validation(self):
pass
def _update_best_model(self):
self.USER_factors_best = self.model.m_U.copy()
self.ITEM_factors_best = self.model.m_V.copy()
def _run_epoch(self, currentEpoch):
n = self.ICM.shape[0]
# for epoch in range(self._params.n_epochs):
num_iter = int(n / self._params.batch_size)
# gen_loss = self.cdl_estimate(data_x, params.cdl_max_iter)
gen_loss = self.model.cdl_estimate(self.ICM, num_iter)
self.model.m_theta[:] = self.model.transform(self.ICM)
likelihood = self.model.pmf_estimate(self._train_users,
self._train_items, None, None,
self._params)
loss = -likelihood + 0.5 * gen_loss * n * self._params.lambda_r
self.USER_factors = self.model.m_U.copy()
self.ITEM_factors = self.model.m_V.copy()
logging.info(
"[#epoch=%06d], loss=%.5f, neg_likelihood=%.5f, gen_loss=%.5f" %
(currentEpoch, loss, -likelihood, gen_loss))