def fit(self,
user_ids_test,
item_ids_test,
ratings_test,
rating_data_train,
model_name,
dataset,
n_epoch=40,
max_checkout_without_progress=20):
self.batch_number = len(rating_data_train) // self.batch_size
self.close_session()
if self.transfer_learning:
latenFactorModelPath = os.path.join(
os.path.dirname(__file__), '../pretrained_models/' + dataset,
'latent-factor-model-' + dataset + '.ckpt')
restore_saver = tf.train.import_meta_graph(latenFactorModelPath +
".meta")
with tf.Session() as sess:
restore_saver.restore(sess, latenFactorModelPath)
self.w_user_mf_transfer_learning = tf.get_default_graph(
).get_tensor_by_name("embd_user:0").eval()
self.w_item_mf_transfer_learning = tf.get_default_graph(
).get_tensor_by_name("embd_item:0").eval()
Utility.reset_graph()
deepNeuralNetworkModelPath = os.path.join(
os.path.dirname(__file__), '../pretrained_models/' + dataset,
'deep-neural-network-model-' + dataset + '.ckpt')
restore_saver_dnn = tf.train.import_meta_graph(
deepNeuralNetworkModelPath + ".meta")
with tf.Session() as sess:
restore_saver_dnn.restore(sess, deepNeuralNetworkModelPath)
self.neural_network_kernels = []
self.w_user_mlp_transfer_learning = tf.get_default_graph(
).get_tensor_by_name("embd_user_mlp:0").eval()
self.w_item_mlp_transfer_learning = tf.get_default_graph(
).get_tensor_by_name("embd_item_mlp:0").eval()
for i in range(len(self.layers)):
self.neural_network_kernels.append(
tf.get_default_graph().get_tensor_by_name(
"hidden" + str(i + 1) + "/kernel:0").eval())
Utility.reset_graph()
self._graph = tf.Graph()
with self._graph.as_default():
self._build_graph()
checkout_without_progress = 0
min_test_error = np.infty
self._session = tf.Session(graph=self._graph)
with self._session.as_default() as sess:
self.init.run()
path = os.path.join(os.path.dirname(__file__),
'../tf_logs/' + dataset)
logdir = Utility.log_dir(path=path, prefix=model_name)
summary_writer = tf.summary.FileWriter(logdir=logdir,
graph=sess.graph)
print("Epoch\tloss\ttrain_error\tRMSE\tMAE\tAccuracy\ttime")
errors = deque(maxlen=self.batch_number)
for epoch in range(n_epoch):
start = time.time()
errors = []
losses = []
for iteration in range(self.batch_number):
user_ids, item_ids, ratings = next(self.iter_train)
_, loss_val, prediction_batch = sess.run(
[self.training_op, self.loss, self.prediction],
feed_dict={
self.user: user_ids,
self.item: item_ids,
self.rating: ratings
})
prediction_batch = np.clip(prediction_batch,
self.min_rating,
self.max_rating)
errors.append(np.power(prediction_batch - ratings, 2))
losses.append(loss_val)
train_err = np.sqrt(np.mean(errors))
loss_mean = np.mean(losses)
prediction_test_data = sess.run([self.prediction],
feed_dict={
self.user: user_ids_test,
self.item: item_ids_test
})
prediction_test_data = np.clip(prediction_test_data,
self.min_rating,
self.max_rating)
accuracy = np.sum(
np.round(np.asarray(prediction_test_data)) ==
ratings_test) / len(ratings_test)
mae = np.mean(np.abs(prediction_test_data - ratings_test))
test_err = np.sqrt(
np.mean(np.power(prediction_test_data - ratings_test, 2)))
end = time.time()
if (test_err < min_test_error):
min_test_error = test_err
saver_path = self.saver.save(
sess,
os.path.join(
os.path.dirname(__file__),
'../pretrained_models/' + dataset + '/' +
model_name))
checkout_without_progress = 0
final_prediction_test_data = sess.run([self.prediction],
feed_dict={
self.user:
user_ids_test,
self.item:
item_ids_test
})
self.final_prediction_test_data = np.clip(
final_prediction_test_data, self.min_rating,
self.max_rating)
else:
checkout_without_progress += 1
if checkout_without_progress >= max_checkout_without_progress:
print("Early stopping")
break
duration = end - start
print("{:3d}\t{}\t{}\t{}\t{}\t{}\t{}(s)".format(
epoch, loss_mean, train_err, test_err, mae, accuracy,
duration))
loss_summary = summary_pb2.Summary(value=[
summary_pb2.Summary.Value(tag="loss",
simple_value=loss_mean)
])
summary_writer.add_summary(loss_summary, epoch)
train_error_summary = summary_pb2.Summary(value=[
summary_pb2.Summary.Value(tag="training_error",
simple_value=train_err)
])
summary_writer.add_summary(train_error_summary, epoch)
test_error_summary = summary_pb2.Summary(value=[
summary_pb2.Summary.Value(tag="RMSE",
simple_value=test_err)
])
summary_writer.add_summary(test_error_summary, epoch)
accuracy_summary = summary_pb2.Summary(value=[
summary_pb2.Summary.Value(tag="accuracy",
simple_value=accuracy)
])
summary_writer.add_summary(accuracy_summary, epoch)
mae_summary = summary_pb2.Summary(value=[
summary_pb2.Summary.Value(tag="MAE", simple_value=mae)
])
summary_writer.add_summary(mae_summary, epoch)
def fit(self,
user_ids_test,
item_ids_test,
ratings_test,
rating_data_train,
model_name,
dataset,
n_epoch=40,
max_checkout_without_progress=20):
self.batch_number = len(rating_data_train) // self.batch_size
self.close_session()
self._graph = tf.Graph()
with self._graph.as_default():
self._build_graph()
checkout_without_progress = 0
min_test_error = np.infty
latent_factor_rmse = []
latent_factor_mae = []
self._session = tf.Session(graph=self._graph)
with self._session.as_default() as sess:
self.init.run()
path = os.path.join(os.path.dirname(__file__),
'../tf_logs/' + dataset)
logdir = Utility.log_dir(path=path, prefix=model_name)
summary_writer = tf.summary.FileWriter(logdir=logdir,
graph=sess.graph)
print("Epoch\ttrain_error\tRMSE\tMAE\tAccuracy\ttime")
errors = deque(maxlen=self.batch_number)
for epoch in range(n_epoch):
start = time.time()
for iteration in range(self.batch_number):
user_ids, item_ids, ratings = next(self.iter_train)
_, prediction_batch = sess.run(
[self.training_op, self.prediction],
feed_dict={
self.user: user_ids,
self.item: item_ids,
self.rating: ratings
})
prediction_batch = np.clip(prediction_batch,
self.min_rating,
self.max_rating)
errors.append(np.power(prediction_batch - ratings, 2))
train_err = np.sqrt(np.mean(errors))
prediction_test_data = sess.run([self.prediction],
feed_dict={
self.user: user_ids_test,
self.item: item_ids_test
})
prediction_test_data = np.clip(prediction_test_data,
self.min_rating,
self.max_rating)
mf_accuracy = np.sum(
np.round(np.asarray(prediction_test_data)) ==
ratings_test) / len(ratings_test)
mae = np.mean(np.abs(prediction_test_data - ratings_test))
test_err = np.sqrt(
np.mean(np.power(prediction_test_data - ratings_test, 2)))
end = time.time()
if (test_err < min_test_error):
min_test_error = test_err
saver_path = self.saver.save(
sess,
os.path.join(
os.path.dirname(__file__),
'../pretrained_models/' + dataset + '/' +
model_name + ".ckpt"))
checkout_without_progress = 0
final_prediction_test_data = sess.run([self.prediction],
feed_dict={
self.user:
user_ids_test,
self.item:
item_ids_test
})
self.final_prediction_test_data = np.clip(
final_prediction_test_data, self.min_rating,
self.max_rating)
else:
checkout_without_progress += 1
if checkout_without_progress >= max_checkout_without_progress:
print("Early stopping")
break
duration = end - start
print("{:3d}\t{}\t{}\t{}\t{}\t{}(s)".format(
epoch, train_err, test_err, mae, mf_accuracy, duration))
train_error_summary = summary_pb2.Summary(value=[
summary_pb2.Summary.Value(tag="training_error",
simple_value=train_err)
])
summary_writer.add_summary(train_error_summary, epoch)
test_error_summary = summary_pb2.Summary(value=[
summary_pb2.Summary.Value(tag="RMSE",
simple_value=test_err)
])
summary_writer.add_summary(test_error_summary, epoch)
latent_factor_rmse.append(test_err)
latent_factor_mae.append(mae)
mf_accuracy_summary = summary_pb2.Summary(value=[
summary_pb2.Summary.Value(tag="accuracy",
simple_value=mf_accuracy)
])
summary_writer.add_summary(mf_accuracy_summary, epoch)
mae_summary = summary_pb2.Summary(value=[
summary_pb2.Summary.Value(tag="MAE", simple_value=mae)
])
summary_writer.add_summary(mae_summary, epoch)