def main():
batch = 8
epoch = 20
loss = MeanAbsoluteError()
learning_rate = PiecewiseConstantDecay(boundaries=[100000],
values=[1e-4, 5e-5])
res_blocks = [
15, 15, 15, 15, 15, 9, 9, 9, 9, 9, 5, 5, 5, 5, 5, 3, 3, 3, 3, 3
]
checkpoint_dir = './ckpt/edsr'
# 데이터 가져오기
ds_train = VCTK(subset='train').dataset()
ds_valid = VCTK(subset='valid').dataset()
# 모델 빌딩
edsr_model = edsr2(scale=4, res_blocks=res_blocks, res_block_scaling=0.7)
# 훈련
edsr_trainer = EDSRTrainer(model=edsr_model,
loss=loss,
learning_rate=learning_rate,
checkpoint_dir=checkpoint_dir)
edsr_trainer.train(train_dataset=ds_train,
valid_dataset=ds_valid,
batch=batch,
epoch=epoch)
edsr_model.save_weights(
f'./weights/EDSR_16000_{len(res_blocks)}res_{batch}batch_{epoch}epochs_tanh_entropy_glorot_uniform.h5'
)
import tensorflow as tf from data import VCTK data = VCTK(batch_size = 20) sess = tf.InteractiveSession() x = data.mfcc u = x.eval() print u
#
# hyper parameters
#
batch_size = 1 # batch size
num_blocks = 3 # dilated blocks
num_dim = 128 # latent dimension
#
# inputs
#
# VCTK corpus input tensor ( with QueueRunner )
data = VCTK(batch_size=batch_size,
data_path='/Volumes/Warehouse/VCTK-Corpus/',
mode='train')
# vocabulary size
voca_size = data.voca_size
# mfcc feature of audio
x = tf.placeholder(dtype=tf.sg_floatx, shape=(batch_size, None, 40))
# sequence length except zero-padding
seq_len = tf.not_equal(x.sg_sum(dims=2), 0.).sg_int().sg_sum(dims=1)
#
# encode graph ( atrous convolution )
#
tf.sg_verbosity(10) # # hyper parameters # batch_size = 16 # batch size num_blocks = 3 # dilated blocks num_dim = 128 # latent dimension # # inputs # # VCTK corpus input tensor ( with QueueRunner ) data = VCTK(batch_size=batch_size) # vocabulary size voca_size = data.voca_size # mfcc feature of audio x = data.mfcc # sequence length except zero-padding seq_len = tf.not_equal(x.sg_sum(dims=2), 0.).sg_int().sg_sum(dims=1) # target sentence label y = data.label # # encode graph ( atrous convolution )
def main(data_dir, p_name):
index = 1 # sample index
person = os.path.join(data_dir, p_name)
# # Get Audio files
filenames = [fname for fname in os.listdir(person)]
# # Play sample files
# # play_sound(get_wav_name(person, filenames[index]))
# # Read file
# rate, data = wavfile.read(get_wav_name(person, filenames[index]))
# print(data.min())
# text = os.path.join(person,filenames[index])
# file = tf.io.read_file(text)
# data, rate = tf.audio.decode_wav(file)
#
# print(type(data))
""" 다운샘플링한거 주파수영역에서 보기 """
# F = np.fft.fft(data.numpy().ravel())
# mag = np.abs(F)
# freq = rate.numpy()/len(mag)
# w = np.arange(0, len(mag)) * freq
# plt.plot(w,mag)
# plt.show()
# print(mag.min(), mag.max())
# #TODO: scale 조정(decode_wav로 했다면 필수)
# data = data.numpy()
# print(((2**15)*data).astype('int16').min(), rate)
# downsampled_data = signal.decimate(data.numpy().ravel(), 12)
# downsampled_data = signal.decimate(downsampled_data, 2)
# downsampled_data = downsampled_data * (2**15)
# print(len(downsampled_data) / 2000)
# print(downsampled_data.reshape((-1,)).shape)
# wavfile.write('test1.wav', 2000, downsampled_data.astype('int16'))
# a = [[1,2,3,4], [2,3,4], [3,5,6], [4,5,6], [5,1]]
# a = np.array(a)
# # print(type(a[0]))
# file = h5py.File('./data/test.h5', 'w')
# dt = h5py.vlen_dtype(np.dtype('int16'))
# dataset = file.create_dataset('test', shape=(5,), dtype=dt)
# for i, d in enumerate(a):
# dataset[i] = d
# file.close()
# TODO:hdf5파일 원소들의 길이가 가변인 파일 저장시키기
# file = h5py.File('./data/test.h5', 'r')
# data = file['test'][...]
# d_gen = data_generator(data)
# for file in d_gen:
# print(file.dtype)
ds_hr, ds_lr = VCTK().dataset()
i = 1
for hr, lr in zip(ds_hr.repeat(1), ds_lr.repeat(1)):
if i % 1000 == 0:
print(i)
i += 1
# print(hr.shape, lr.shape)
print(i)
tf.sg_verbosity(10) # # hyper parameters # batch_size = 1 # batch size num_blocks = 3 # dilated blocks num_dim = 128 # latent dimension # # inputs # # VCTK corpus input tensor ( with QueueRunner ) data = VCTK(vocabulary_loading=True) # vocabulary size voca_size = data.voca_size # mfcc feature of audio x = tf.placeholder(dtype=tf.sg_floatx, shape=(batch_size, None, 20)) # sequence length except zero-padding seq_len = tf.not_equal(x.sg_sum(dims=2), 0.).sg_int().sg_sum(dims=1) # # encode graph ( atrous convolution ) #
tf.sg_verbosity(10) # # hyper parameters # batch_size = 1 # batch size num_blocks = 3 # dilated blocks num_dim = 128 # latent dimension # # inputs # # VCTK corpus input tensor ( with QueueRunner ) data = VCTK() # vocabulary size voca_size = data.voca_size # mfcc feature of audio x = tf.placeholder(dtype=tf.sg_floatx, shape=(batch_size, None, 20)) # sequence length except zero-padding seq_len = tf.not_equal(x.sg_sum(dims=2), 0.).sg_int().sg_sum(dims=1) # # encode graph ( atrous convolution ) #