network_data.encoding_features = 40
network_data.learning_rate = 0.001
network_data.adam_epsilon = 0.01
network_data.optimizer = tf.train.AdamOptimizer(
learning_rate=network_data.learning_rate,
epsilon=network_data.adam_epsilon)
###########################################################################################################
encoder_decoder = RecurrentEncoderDecoder(network_data)
encoder_decoder.create_graph()
test_database = Database.fromFile(project_data.TEST_ENCODER_DATABASE_FILE,
project_data)
test_feats, _ = test_database.to_set()
test_feats = test_feats[0:1]
encoder_decoder.train(input_seq=test_feats,
output_seq=test_feats,
restore_run=True,
save_partial=True,
save_freq=10,
use_tensorboard=True,
tensorboard_freq=5,
training_epochs=30,
batch_size=5)
out = encoder_decoder.predict(test_feats[0])
f1 = plt.figure(0)
feature_config.mfcc_window = np.hanning
label_type = "classic" # "classic", "las", "optim"
use_embedding = False
word_level = False
vocab_file = project_data.VOCAB_FILE
if label_type == "classic":
label_class = ClassicLabel
elif label_type == "las":
label_class = LASLabel
else:
label_class = OptimalLabel
for wav_dir in wav_dirs:
database = Database(project_data)
# Get the names of each wav file in the directory
wav_names = os.listdir(wav_dir)
# wav_names = wav_names[0:100]
for wav_index in range(len(wav_names)):
if wav_dir == project_data.WAV_TRAIN_DIR:
label_dir = project_data.TRANSCRIPTION_TRAIN_DIR
else:
label_dir = project_data.TRANSCRIPTION_TEST_DIR
# Get filenames
wav_filename = wav_dir + '/' + wav_names[wav_index]
label_filename = label_dir + '/' + wav_names[wav_index].split(
".")[0] + '.TXT'
network_data.rnn_regularizer = 0
network_data.use_dropout = True
network_data.decoder_function = tf.nn.ctc_greedy_decoder
network_data.learning_rate = 0.001
network_data.adam_epsilon = 0.0001
network_data.optimizer = tf.train.AdamOptimizer(
learning_rate=network_data.learning_rate,
epsilon=network_data.adam_epsilon)
###########################################################################################################
network = ZorzNet(network_data)
network.create_graph()
train_database = Database.fromFile(project_data.TRAIN_DATABASE_FILE,
project_data)
test_database = Database.fromFile(project_data.TEST_DATABASE_FILE,
project_data)
train_feats, train_labels = train_database.to_set()
test_feats, test_labels = test_database.to_set()
train_feats = train_feats[2:3]
train_labels = train_labels[2:3]
network.train(train_features=train_feats,
train_labels=train_labels,
restore_run=False,
save_partial=True,
save_freq=10,
use_tensorboard=True,
# Load project data
project_data = ProjectData()
transcription_file_list = ['promptsl40.train', 'promptsl40.test']
is_train_flag = False
for file in transcription_file_list:
if 'train' in file:
print('Loading TRAIN database')
is_train_flag = True
else:
print('Loading TEST database')
is_train_flag = False
database = Database(project_data)
# Create a list of all labels
transcription_list = []
with open(project_data.TRANSCRIPTION_DIR+'/'+file, 'r', encoding='latin1') as f:
for line in f.readlines():
aux = line.split(' ', 1)
name = aux[0].rstrip()
# Remove accents
transcription = remove_accents(aux[1].rstrip())
dict = {'name': name, 'transcription': transcription}
transcription_list.append(dict)
# Get the names of each wav file in the directory
if is_train_flag:
wav_dir = project_data.WAV_TRAIN_DIR
feature_config.preemph = 0.98
feature_config.num_filters = 40
feature_config.num_ceps = 26
feature_config.mfcc_window = np.hamming
label_type = "las" # "classic", "las", "optim"
if label_type == "classic":
label_class = ClassicLabel
elif label_type == "las":
label_class = LASLabel
else:
label_class = OptimalLabel
for wav_dir in wav_dirs:
database = Database(project_data)
# Get the names of each wav file in the directory
wav_names = os.listdir(wav_dir)
for wav_index in range(len(wav_names)):
if wav_dir == project_data.WAV_TRAIN_DIR:
label_dir = project_data.TRANSCRIPTION_TRAIN_DIR
else:
label_dir = project_data.TRANSCRIPTION_TEST_DIR
# Get filenames
wav_filename = wav_dir + '/' + wav_names[wav_index]
label_filename = label_dir + '/' + wav_names[wav_index].split(".")[0] + '.TXT'
audio_feature = AudioFeature.fromFile(wav_filename, feature_config)
from ConfigSpace.hyperparameters import CategoricalHyperparameter, UniformFloatHyperparameter, \
UniformIntegerHyperparameter
from smac.stats.stats import Stats
from smac.utils.io.traj_logging import TrajLogger
from tensorflow.contrib.layers import l2_regularizer
from src.neural_network.data_conversion import indexToStr
from src.neural_network.ZorzNet.ZorzNetData import ZorzNetData
from src.neural_network.ZorzNet.ZorzNet import ZorzNet
from src.utils.Database import Database
from src.utils.ProjectData import ProjectData
from src.utils.smac_utils import wait_for_user_input_non_block, remove_if_exist
project_data = ProjectData()
train_database = Database.fromFile(project_data.TRAIN_DATABASE_FILE, project_data)
val_database = Database.fromFile(project_data.VAL_DATABASE_FILE, project_data)
# TODO Add a different method for this
train_feats, train_labels, _, _, _, _ = train_database.split_sets(1.0, 0.0, 0.0)
val_feats, val_labels, _, _, _, _ = val_database.split_sets(1.0, 0.0, 0.0)
# -----------------------------------------------------------------------------------------
optimization_epochs = 100
validation_epochs = 200
# space_optimization_evals = 1
batch_size = 50
run_folder = 'run'
run_count_step = 1