def _data_generator(self, mode=DataGeneratorMode.training):
def gen_training_data():
(x_train, y_train), (_, _), _ = load(self.data_mode)
z = list(zip(x_train, y_train))
pos = int(len(x_train) * self.validation_split)
z_tr = z[pos:]
z_val = z[:pos]
self.validation_x, self.validation_y = zip(*z_val)
self.train_x, self.train_y = zip(*z_tr)
return x_train, y_train
inputs, outputs = [], []
if mode == DataGeneratorMode.training or mode == DataGeneratorMode.validation:
if len(self.validation_x) == 0:
inputs, outputs = gen_training_data()
else:
inputs, outputs = self.validation_x, self.validation_y
else:
(_, _), (x_test, y_test), _ = load(self.data_mode)
inputs, outputs = x_test, y_test
index = 0
last_batch_x, last_batch_y = [], []
while True:
def get_batch(i):
try:
vectorized_inp = []
vectorized_out = []
for b in range(self.batch_size):
playlist = inputs[i + b]
vectorized_inp.append(
[self.vocab_index[str(song)] for song in playlist])
vectorized_out.append(
self.embeddings_matrix[self.vocab_index[str(
outputs[i + b])]])
x = np.asarray(
pad_sequences(vectorized_inp, maxlen=self.max_length))
y = np.asarray(vectorized_out)
return x, y, i
except KeyError as key_error:
# logging.warning('Can\'t find key', key_error)
return get_batch(i + self.batch_size)
try:
last_batch_x, last_batch_y, index = get_batch(index)
index += self.batch_size
except IndexError:
index = 0
except Exception as exp:
logging.error('Data generator error', exp)
index = 0
yield last_batch_x, last_batch_y
def gen_training_data():
(x_train, y_train), (_, _), _ = load(self.data_mode)
z = list(zip(x_train, y_train))
pos = int(len(x_train) * self.validation_split)
z_tr = z[pos:]
z_val = z[:pos]
self.validation_x, self.validation_y = zip(*z_val)
self.train_x, self.train_y = zip(*z_tr)
return x_train, y_train
def __init__(self,
mode=DatasetMode.small,
model_name=ModelName.simple_gru):
self.optimizer = Adam(lr=0.001,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-08,
decay=0.0)
self.activation = 'softmax'
self.loss = 'categorical_crossentropy'
self.metrics = top_k_accuracy_func_list([50, 100, 200, 300, 400, 500])
self.epochs = 15
self.batch_size = 512
self.validation_split = 0.1
self.data_mode = mode
'''
Index of songs in x_train(or test) starts from 1 because of zero padding.
Index of songs in y_train(or test) starts from zero like song hash.
For instance:
In dataset, index of songA is 21.
songA's index is 22 in x_train(or test)
songA's index is 21 in y_train(or test).
The goal is the neural network having the ability to ignore zero-paddings
'''
(x_train, y_train), (x_test, y_test), songs = load(mode)
self.max_length = max([len(playlist) for playlist in x_train])
self.song_hash = songs
self.train_len = len(x_train)
self.test_len = len(x_test)
self.validation_x, self.validation_y = [], []
self.train_x, self.train_y = [], []
if model_name is ModelName.simple_gru:
self.weights_loc = TRAINED_MODELS_BASE_PATH + 'simple_gru_weights.best.hdf5'
self.tb_logs_path = LOGS_BASE_PATH + 'simple_gru_logs'
self.model = self.create_simple_gru_model()
elif model_name is ModelName.bi_directional_lstm:
self.weights_loc = TRAINED_MODELS_BASE_PATH + 'bi_lstm_weights.best.hdf5'
self.tb_logs_path = LOGS_BASE_PATH + 'bi_lstm_logs'
self.model = self.create_bi_lstm_model()
elif model_name is ModelName.attention_bilstm:
self.weights_loc = TRAINED_MODELS_BASE_PATH + 'ablstm_model_weights.best.hdf5'
self.tb_logs_path = LOGS_BASE_PATH + 'ablstm_model_logs'
self.model = self.create_ablstm_model()
else:
raise Exception("Unknown Model Name; ", model_name)
self.callbacks = self._get_callback_functions()
def __init__(self,
mode=DatasetMode.small,
model_name=ModelName.simple_gru):
self.optimizer = RMSprop(lr=0.001, rho=0.9, epsilon=None, decay=0.0)
self.activation = 'tanh'
self.loss = 'cosine'
self.metrics = [
'mae', 'mse', 'acc', 'mape', 'cosine', 'categorical_crossentropy'
]
self.epochs = 50
self.batch_size = 512
self.validation_split = 0.1
self.data_mode = mode
v_index, embedding_matrix = SongEmbeddings().get_embeddings_matrix()
self.embeddings_matrix = embedding_matrix
self.vocab_index = v_index
self.validation_x, self.validation_y = [], []
self.train_x, self.train_y = [], []
'''
Index of songs in x_train(or test) starts from 1 because of zero padding.
Index of songs in y_train(or test) starts from zero like song hash.
For instance:
In dataset, index of songA is 21.
songA's index is 22 in x_train(or test)
songA's index is 21 in y_train(or test).
The goal is the neural network having the ability to ignore zero-paddings
'''
(x_train, y_train), (x_test, _), songs = load(mode)
self.max_length = max([len(playlist) for playlist in x_train])
self.song_hash = songs
self.train_len = len(x_train)
self.test_len = len(x_test)
if model_name is ModelName.simple_gru:
self.weights_loc = TRAINED_MODELS_BASE_PATH + 'e_simple_gru_weights.best.hdf5'
self.tb_logs_path = LOGS_BASE_PATH + 'e_simple_gru_logs'
self.model = self._create_simple_gru_model()
elif model_name is ModelName.bi_directional_lstm:
self.weights_loc = TRAINED_MODELS_BASE_PATH + 'e_bi_lstm_weights.best.hdf5'
self.tb_logs_path = LOGS_BASE_PATH + 'e_bi_lstm_logs'
self.model = self._create_bi_lstm_model()
elif model_name is ModelName.attention_bilstm:
self.weights_loc = TRAINED_MODELS_BASE_PATH + 'e_ablstm_model_weights.best.hdf5'
self.tb_logs_path = LOGS_BASE_PATH + 'e_ablstm_model_logs'
self.model = self._create_ablstm_model()
else:
raise Exception("Unknown Model Name; ", model_name)
self.callbacks = self._get_callback_functions()
def _data_generator(self, mode=DataGeneratorMode.training):
def gen_training_data():
(x_train, y_train), (_, _), _ = load(self.data_mode)
z = list(zip(x_train, y_train))
pos = int(len(x_train) * self.validation_split)
z_tr = z[pos:]
z_val = z[:pos]
self.validation_x, self.validation_y = zip(*z_val)
self.train_x, self.train_y = zip(*z_tr)
return x_train, y_train
inputs, outputs = [], []
if mode == DataGeneratorMode.training or mode == DataGeneratorMode.validation:
if len(self.validation_x) == 0:
inputs, outputs = gen_training_data()
else:
inputs, outputs = self.validation_x, self.validation_y
else:
(_, _), (x_test, y_test), _ = load(self.data_mode)
inputs, outputs = x_test, y_test
index = 0
last_batch_x, last_batch_y = [], []
while True:
try:
inp = inputs[index:index + self.batch_size]
out = outputs[index:index + self.batch_size]
last_batch_x = np.asarray(sequence.pad_sequences(
inp, maxlen=self.max_length),
dtype="int64")
last_batch_y = to_categorical(out,
len(self.song_hash) +
1) # Zero is included
index += self.batch_size
except IndexError:
index = 0
except Exception as exp:
logging.error('Data generator error', exp)
index = 0
yield last_batch_x, last_batch_y