def test_personal_matrix_factorization_trains_multiple_users_fails(self):
num_users = 2
num_items = 8
num_latent_factors = 5
personal_model = True
add_biases = False
l2_regularization = 0.0
with self.assertRaises(ValueError):
models.get_matrix_factorization_model(
num_users,
num_items,
num_latent_factors,
personal_model=personal_model,
add_biases=add_biases,
l2_regularization=l2_regularization)
def test_personal_matrix_factorization_trains_keras_accuracy_threshold(self):
num_users = 1
num_items = 8
num_latent_factors = 5
personal_model = True
add_biases = False
l2_regularization = 0.0
matrix_factorization_model = models.get_matrix_factorization_model(
num_users,
num_items,
num_latent_factors,
personal_model=personal_model,
add_biases=add_biases,
l2_regularization=l2_regularization)
keras_model = models.build_keras_model(matrix_factorization_model)
# Ensure number of parameters of model is as expected as a quick check.
expected_num_params = (
num_latent_factors # User embeddings.
+ num_items * num_latent_factors) # Item embeddings.
num_params = keras_model.count_params()
self.assertEqual(expected_num_params, num_params)
num_trainable_params = count_trainable_params(keras_model)
self.assertEqual(expected_num_params, num_trainable_params)
history = keras_model.fit(
self.train_items, self.train_preferences, batch_size=1, epochs=1)
# Ensure the model has a valid loss after one epoch (not NaN).
self.assertIn(_KERAS_LOSS, history.history)
losses = history.history[_KERAS_LOSS]
self.assertLen(losses, 1)
self.assertFalse(np.isnan(losses[0]))
# Ensure the model has valid number of examples after one epoch (not NaN).
self.assertIn(_KERAS_NUM_EXAMPLES, history.history)
num_examples = history.history[_KERAS_NUM_EXAMPLES]
self.assertLen(num_examples, 1)
self.assertEqual(num_examples[0], 5)
# Ensure the model has valid number of batches after one epoch (not NaN).
self.assertIn(_KERAS_NUM_BATCHES, history.history)
num_batches = history.history[_KERAS_NUM_BATCHES]
self.assertLen(num_batches, 1)
self.assertEqual(num_batches[0], 5)
# Ensure the model has a valid reconstruction accuracy after one epoch (not
# NaN).
self.assertIn(_KERAS_ACCURACY, history.history)
accuracies = history.history[_KERAS_ACCURACY]
self.assertLen(accuracies, 1)
self.assertFalse(np.isnan(accuracies[0]))
def test_matrix_factorization_trains_keras(self):
num_users = 10
num_items = 8
num_latent_factors = 5
personal_model = False
add_biases = False
l2_regularization = 0.0
learning_rate = 0.5
matrix_factorization_model = models.get_matrix_factorization_model(
num_users,
num_items,
num_latent_factors,
personal_model=personal_model,
add_biases=add_biases,
l2_regularization=l2_regularization)
keras_model = models.build_keras_model(
matrix_factorization_model, tf.keras.optimizers.SGD(learning_rate))
# Ensure number of parameters of model is as expected as a quick check.
expected_num_params = (
num_users * num_latent_factors # User embeddings.
+ num_items * num_latent_factors) # Item embeddings.
num_params = keras_model.count_params()
self.assertEqual(expected_num_params, num_params)
num_trainable_params = count_trainable_params(keras_model)
self.assertEqual(expected_num_params, num_trainable_params)
history = keras_model.fit([self.train_users, self.train_items],
self.train_preferences,
batch_size=1,
epochs=1)
# Ensure the model has a valid loss after one epoch (not NaN).
self.assertIn(_KERAS_LOSS, history.history)
losses = history.history[_KERAS_LOSS]
self.assertLen(losses, 1)
self.assertFalse(np.isnan(losses[0]))
# Ensure the model has valid number of examples after one epoch (not NaN).
self.assertIn(_KERAS_NUM_EXAMPLES, history.history)
num_examples = history.history[_KERAS_NUM_EXAMPLES]
self.assertLen(num_examples, 1)
self.assertEqual(num_examples[0], 5)
# Ensure the model has valid number of batches after one epoch (not NaN).
self.assertIn(_KERAS_NUM_BATCHES, history.history)
num_batches = history.history[_KERAS_NUM_BATCHES]
self.assertLen(num_batches, 1)
self.assertEqual(num_batches[0], 5)