def create_h5(self, split_num, iter_num, input_sequence, DEL_TRAIN_WAV):
cpu_cores = int(split_num / iter_num)
tmp1 = []
tmp2 = []
tmp3 = []
# noisy_dir = join(noisy_dir, 'train')
training_data_list = search_wav(self.noisy_dir)
print('Total training files: ', len(training_data_list))
for file in training_data_list:
try:
snr, noise_name, clean_name1, clean_neme2 = file.split(
'/')[-1].split('_')
clean_file = join(
self.noisy_dir,
'_'.join(['0ms', 'n0', clean_name1, clean_neme2]))
noisy_file = file
except:
snr, noise_name, clean_name = file.split('/')[-1].split('_')
clean_file = join(self.noisy_dir,
'_'.join(['0ms', 'n0', clean_name]))
noisy_file = file
tmp1.append(clean_file)
tmp2.append(noisy_file)
training_num = 30000
t1, t2 = shuffle(np.array(tmp1), np.array(tmp2))
t1 = t1[:training_num]
t2 = t2[:training_num]
clean_split_list = split_list(t1, wanted_parts=split_num)
noisy_split_list = split_list(t2, wanted_parts=split_num)
start = 0
end = cpu_cores
for num in range(iter_num):
print(start, end)
pool = Pool(cpu_cores)
func = partial(_create_split_h5, clean_split_list,
noisy_split_list, self.save_h5_dir,
self.save_h5_name, input_sequence)
pool.map(func, range(start, end))
pool.close()
pool.join()
start = end
end += cpu_cores
if DEL_TRAIN_WAV:
shutil.rmtree(self.noisy_dir)
def test(self, testing_data_dir, result_dir, test_saver, n_cores, num_test=False):
print('Start Testing')
tmp_list = search_wav(testing_data_dir)
if num_test:
test_list = np.random.choice(tmp_list, num_test)
else:
test_list = tmp_list
print('All testing data number:', len(test_list))
REG_dir = join(result_dir, 'REG')
Noisy_write_dir = join(result_dir, 'Source')
Clean_write_dir = join(result_dir, 'Target')
if not os.path.exists(result_dir):
os.makedirs(result_dir)
os.makedirs(REG_dir)
os.makedirs(Noisy_write_dir)
os.makedirs(Clean_write_dir)
with tf.Session(config=self.config) as sess:
self.saver.restore(sess=sess, save_path=test_saver)
for file in tqdm(test_list):
hop_length = 256
file_name = file.split('/')[-1]
try:
snr, noise_name, clean_name1, clean_neme2 = file.split(
'/')[-1].split('_')
clean_file = join(testing_data_dir, '_'.join(
['0dB', 'n0', clean_name1, clean_neme2]))
noisy_file = file
except:
snr, noise_name, clean_name = file.split(
'/')[-1].split('_')
noisy_file = join(testing_data_dir, file_name)
REG_file = join(REG_dir, file_name)
Noisy_file = join(Noisy_write_dir, file_name)
Clean_file = join(Clean_write_dir, file_name)
X_in_seq = wav2spec(noisy_file, sr=16000,
forward_backward=True, SEQUENCE=False, norm=True, hop_length=hop_length)
re_reg = sess.run([self.reg_layer],
feed_dict={self.x_noisy: X_in_seq})[:][0]
spec2wav(noisy_file, 16000, REG_file,
re_reg, hop_length=hop_length)
copy_file(noisy_file, Noisy_file)
copy_file(clean_file, Clean_file)
def main():
print('haha')
clean_dir = FLAGS.clean_dir
noise_dir = FLAGS.noise_dir
noisy_dir = FLAGS.noisy_dir
enhanced_dir = FLAGS.enhanced_dir
tb_dir = FLAGS.tb_dir
saver_dir = FLAGS.saver_dir
TRAIN = FLAGS.TRAIN
TEST = FLAGS.TEST
ncores = FLAGS.n_cores
epochs = FLAGS.epochs
batch_size = FLAGS.batch_size
lr = FLAGS.learning_rate
train_task = 'same_noise' # set task name for noting your dataset
# ===========================================================
# =========== Synthesize Noisy Data ============
# ===========================================================
clean_file_list = search_wav(clean_dir)
clean_train_list, clean_test_list = train_test_split(
clean_file_list, test_size=0.2)
noise_file_list = search_wav(noise_dir)
noise_train_list, noise_test_list = train_test_split(
noise_file_list, test_size=0.2)
print('--- Synthesize Training Noisy Data ---')
train_noisy_dir = join(noisy_dir, 'train')
sr_clean = 16000
sr_noise = 44100
snr_list = ['20dB', '0dB', '-20dB']
data_num = 100 # set data_num to make training data numbers for different snr
syn_train = Synth(clean_train_list, noise_train_list[
0:10], sr_clean, sr_noise)
syn_train.gen_noisy(snr_list, train_noisy_dir,
data_num=data_num, ADD_CLEAN=True, cpu_cores=ncores)
print('--- Synthesize Testing Noisy Data ---')
test_noisy_dir = join(noisy_dir, 'test')
sr_clean = 16000
sr_noise = 44100
data_num = 10 # set data_num to make testing data numbers for different snr
snr_list = ['15dB', '5dB', '-5dB']
syn_train = Synth(clean_test_list, noise_train_list[
0:10], sr_clean, sr_noise)
syn_train.gen_noisy(snr_list, test_noisy_dir,
data_num=data_num, ADD_CLEAN=True, cpu_cores=ncores)
# ===========================================================
# =========== Create Training Matrix ============
# ===========================================================
print('--- Generate Training Matrix ---')
train_task = 'same_noise' # set task name for noting your dataset
training_files_dir = FLAGS.training_files_dir
train_noisy_dir = join(noisy_dir, 'train')
DEL_TRAIN_WAV = True
gen_mat = GenMatrix(training_files_dir, train_task, train_noisy_dir)
split_num = 50 # number of spliting files
iter_num = 2 # set iter number to use multi-processing, cpu_cores = split_num/iter_num
input_sequence = True # set input data is sequence or not
gen_mat.create_h5(split_num=split_num, iter_num=iter_num,
input_sequence=input_sequence,
DEL_TRAIN_WAV=DEL_TRAIN_WAV)
# ===========================================================
# =========== Main Model ============
# ===========================================================
print('--- Build Model ---')
note = 'DDAE'
date = '0604'
split_num = 50
training_files_dir = FLAGS.training_files_dir
model = REG(tb_dir, saver_dir, train_task, date, gpu_num='3', note=note)
model.build(init_learning_rate=1e-3, reuse=False)
print('--- Train Model ---')
model.train(training_files_dir, split_num, epochs, batch_size)
print('--- Test Model ---')
testing_data_dir = join(noisy_dir, 'test')
result_dir = '../data/enhanced/{}_{}/'.format(note, date)
num_test = 30 # Set this number to decide how many testing data you wanna use. (None => All)
cpu_cores = 30
test_saver = '{}_{}/{}/best_saver_{}'.format(
saver_dir, note, date, train_task)
model.test(testing_data_dir, result_dir,
test_saver, cpu_cores, num_test)