def train(self,
no_epochs=20,
batches=1024,
lr=0.001,
no_factors=10,
no_negatives=10,
gen_mode='point',
val_split=0.1):
print('Generating training instances', 'of type', gen_mode)
x, y = generator(self.observed_relevance,
self.categories,
self.no_categories,
self.category_per_item,
self.categories_per_user,
no_negatives=no_negatives,
gen_mode=gen_mode)
print('Performing training -', 'Epochs', no_epochs, 'Batch Size',
batches, 'Learning Rate', lr, 'Factors', no_factors, 'Negatives',
no_negatives, 'Mode', gen_mode)
self.model = self.__get_model(no_factors,
np.array([64, 32, 16, 8], np.int32))
self.model.compile(optimizer=tf.keras.optimizers.Adam(lr=lr),
loss='binary_crossentropy')
user_input, item_i_input, _ = x
labels = y
callbacks = [
tf.keras.callbacks.EarlyStopping(monitor='val_loss',
patience=2,
verbose=1)
]
self.model.fit([np.array(user_input),
np.array(item_i_input)],
np.array(labels),
validation_split=val_split,
batch_size=batches,
epochs=no_epochs,
verbose=1,
shuffle=True,
callbacks=callbacks)
def train(self,
no_epochs=100,
batches=1024,
lr=0.001,
no_factors=10,
no_negatives=10,
gen_mode='pair',
val_split=0.01,
val_interval=4):
print('Generating training instances', 'of type', gen_mode)
x, y = generator(self.observed_relevance,
self.categories,
self.no_categories,
self.category_per_item,
self.categories_per_user,
no_negatives=no_negatives,
gen_mode=gen_mode)
print('Performing training -', 'Epochs', no_epochs, 'Batch Size',
batches, 'Learning Rate', lr, 'Factors', no_factors, 'Negatives',
no_negatives, 'Mode', gen_mode)
self.model = self.__get_model()
self.model.compile(optimizer=tf.keras.optimizers.Adam(lr=lr),
loss=get_bpr_loss)
user_input, item_i_input, item_j_input = x
labels = y
train_instance_indexes = np.random.choice(
list(range(len(user_input))),
int(len(user_input) * (1 - val_split)),
replace=False)
val_instance_indexes = np.array(
list(set(range(len(user_input))) - set(train_instance_indexes)))
user_input_train = user_input[train_instance_indexes]
item_i_input_train = item_i_input[train_instance_indexes]
item_j_input_train = item_j_input[train_instance_indexes]
labels_train = labels[train_instance_indexes]
user_input_val = user_input[val_instance_indexes]
item_i_input_val = item_i_input[val_instance_indexes]
item_j_input_val = item_j_input[val_instance_indexes]
best_auc_score = 0
for epoch in range(no_epochs):
self.model.fit(
[user_input_train, item_i_input_train, item_j_input_train],
labels_train,
initial_epoch=epoch,
epochs=epoch + 1,
batch_size=batches,
verbose=1,
shuffle=True)
if (epoch % val_interval) == 0:
user_matrix = self.model.get_layer('UserEmb').get_weights()[0]
item_matrix = self.model.get_layer('ItemEmb').get_weights()[0]
auc_scores = []
for t, (u, i, j) in enumerate(
zip(user_input_val, item_i_input_val,
item_j_input_val)):
auc_scores.append(
1 if np.dot(user_matrix[u], item_matrix[i]) > np.
dot(user_matrix[u], item_matrix[j]) else 0)
if (t % 1000) == 0:
print('\rValidation accuracy:',
auc_scores.count(1) / len(auc_scores),
'(Sample',
t,
'of',
str(len(val_instance_indexes)) + ')',
end='')
print()
if (auc_scores.count(1) / len(auc_scores)) < best_auc_score:
break
else:
best_auc_score = (auc_scores.count(1) / len(auc_scores))