def test(self):
"""Translate speech using StarGAN ."""
# Load the trained generator.
self.restore_model(self.test_iters)
norm = Normalizer()
# Set data loader.
d, speaker = TestSet(self.test_dir).test_data(self.src_speaker)
targets = self.trg_speaker
for target in targets:
print(target)
assert target in speakers
label_t = self.spk_enc.transform([target])[0]
if label_t == [0]: label_t = [1, 0]
elif label_t == [1]: label_t = [0, 1]
label_t = np.asarray([label_t])
with torch.no_grad():
for filename, content in d.items():
f0 = content['f0']
ap = content['ap']
sp_norm_pad = self.pad_coded_sp(content['coded_sp_norm'])
convert_result = []
for start_idx in range(0,
sp_norm_pad.shape[1] - FRAMES + 1,
FRAMES):
one_seg = sp_norm_pad[:, start_idx:start_idx + FRAMES]
one_seg = torch.FloatTensor(one_seg).to(self.device)
one_seg = one_seg.view(1, 1, one_seg.size(0),
one_seg.size(1))
l = torch.FloatTensor(label_t)
one_seg = one_seg.to(self.device)
l = l.to(self.device)
one_set_return = self.G(one_seg, l).data.cpu().numpy()
one_set_return = np.squeeze(one_set_return)
one_set_return = norm.backward_process(
one_set_return, target)
convert_result.append(one_set_return)
convert_con = np.concatenate(convert_result, axis=1)
convert_con = convert_con[:, 0:content['coded_sp_norm'].
shape[1]]
contigu = np.ascontiguousarray(convert_con.T,
dtype=np.float64)
decoded_sp = decode_spectral_envelope(contigu,
SAMPLE_RATE,
fft_size=FFTSIZE)
f0_converted = norm.pitch_conversion(f0, speaker, target)
wav = synthesize(f0_converted, decoded_sp, ap, SAMPLE_RATE)
name = f'{speaker}-{target}_iter{self.test_iters}_{filename}'
path = os.path.join(self.result_dir, name)
print(f'[save]:{path}')
librosa.output.write_wav(path, wav, SAMPLE_RATE)
def train(self):
# 衰减的学习率缓存
g_lr = self.g_lr
d_lr = self.d_lr
c_lr = self.c_lr
# 开始训练步骤数为0
start_iters = 0
# 如果存在就跳过
if self.resume_iters:
pass
# 调用定义的个性化标准化方法
norm = Normalizer()
# iter用来生成迭代器，这里用来迭代加载数据集
data_iter = iter(self.data_loader)
print('开始训练......')
# 记录当前时间，now函数取当前时间
start_time = datetime.now()
# 利用总迭代次数来进行遍历
for i in range(start_iters, self.num_iters):
# =================================================================================== #
# 1.预处理输入数据 #
# =================================================================================== #
# 获取真实的图像和对应标签标签
try:
# next方法为迭代下一个迭代器
# 利用自定义的加载器获取真实x值，发音者标签在组中索引与源标签
x_real, speaker_idx_org, label_org = next(data_iter)
except:
# 如果迭代器有问题就再转换为迭代器一次然后迭代
data_iter = iter(self.data_loader)
x_real, speaker_idx_org, label_org = next(data_iter)
# 随机生成目标域标签
# torch.randperm返回一个从0到参数-1范围的随机数组
# 因为标签二值化了，所以这里的标签是10组成的，所以一共有label_org.size(0)个标签
# 获得的是随机索引
rand_idx = torch.randperm(label_org.size(0))
# 根据随机数作为源标签的索引作为目标标签数
label_trg = label_org[rand_idx]
# 同理得到随机目标发音者
speaker_idx_trg = speaker_idx_org[rand_idx]
# to表示使用cpu或者gpu运行
x_real = x_real.to(self.device) # 输入数据
label_org = label_org.to(self.device) # 源域one-hot格式标签
label_trg = label_trg.to(self.device) # 目标域ont-hot格式标签
speaker_idx_org = speaker_idx_org.to(self.device) # 源域标签
speaker_idx_trg = speaker_idx_trg.to(self.device) # 目标域标签
# =================================================================================== #
# 2.训练判别器 #
# =================================================================================== #
# 用真实音频数据计算损失
# nn.CrossEntropyLoss()为交叉熵损失函数，但是不是普通的形式，而是主要是将softmax-log-NLLLoss合并到一块得到的结果。
CELoss = nn.CrossEntropyLoss()
# 调用分类器计算真实数据
cls_real = self.C(x_real)
# 计算对应的域分类损失，即用交叉熵实现
cls_loss_real = CELoss(input=cls_real, target=speaker_idx_org)
# 重置缓冲区，具体实现在下面
self.reset_grad()
# tensor.backward为自动求导函数
cls_loss_real.backward()
# optimizer.step这个方法会更新模型所有的参数以提升学习率，一般在backward函数后根据其计算的梯度来更新参数
self.c_optimizer.step()
# 记录中
loss = {}
# 从真实域分类损失张量中获取元素值
# item()得到一个元素张量里面的元素值
loss['C/C_loss'] = cls_loss_real.item()
# 基于源数据的D判断结果
out_r = self.D(x_real, label_org)
# 用假音频帧计算损失
# 根据真实样本与目标标签生成生成样本
x_fake = self.G(x_real, label_trg)
# detach截断反向传播的梯度流，从而让梯度不影响判别器D
# 基于生成样本的D判断结果
out_f = self.D(x_fake.detach(), label_trg)
# torch.nn.Function.binary_cross_entropy_with_logits度量目标逻辑和输出逻辑之间的二进制交叉熵的函数
# 接受任意形状的输入，target要求与输入形状一致。切记：target的值必须在[0,N-1]之间，其中N为类别数，否则会出现莫名其妙的错误，比如loss为负数。
# 计算其实就是交叉熵，不过输入不要求在0，1之间，该函数会自动添加sigmoid运算
# 返回一个填充了标量值1的张量，其大小与输入相同。torch.ones_like(input)
# 相当于torch.ones(input.size(), dtype=input.dtype, layout=input.layout, device=input.device)
# binary_cross_entropy_with_logits和binary_cross_entropy的区别
# 有一个（类）损失函数名字中带了with_logits. 而这里的logits指的是,该损失函数已经内部自带了计算logit的操作，
# 无需在传入给这个loss函数之前手动使用sigmoid/softmax将之前网络的输入映射到[0,1]之间
d_loss_t = F.binary_cross_entropy_with_logits(input=out_f,target=torch.zeros_like(out_f, dtype=torch.float)) + \
F.binary_cross_entropy_with_logits(input=out_r, target=torch.ones_like(out_r, dtype=torch.float))
# 生成样本的分类结果
out_cls = self.C(x_fake)
# 交叉熵计算生成样本的域分类损失
d_loss_cls = CELoss(input=out_cls, target=speaker_idx_trg)
# 计算梯度惩罚的损失
alpha = torch.rand(x_real.size(0), 1, 1, 1).to(self.device)
# 计算x_hat
# requires_grad_设置积分方法，将requires_grad是否积分的属性设置为真
# 取一个随机数混合真实样本和生成样本得到一个x尖
x_hat = (alpha * x_real.data +
(1 - alpha) * x_fake.data).requires_grad_(True)
# 计算混合样本和目标标签的判别结果
out_src = self.D(x_hat, label_trg)
# 调用自定义方法得到处理导数后的数据
d_loss_gp = self.gradient_penalty(out_src, x_hat)
# 计算判别器的总体损失
d_loss = d_loss_t + self.lambda_cls * d_loss_cls + 5 * d_loss_gp
# 调用自定义方法重置梯度变化缓冲区
self.reset_grad()
# 对D的损失求导
d_loss.backward()
# 更新模型判别器D参数
self.d_optimizer.step()
# loss['D/d_loss_t'] = d_loss_t.item()
# loss['D/loss_cls'] = d_loss_cls.item()
# loss['D/D_gp'] = d_loss_gp.item()
# 获取判别器损失
loss['D/D_loss'] = d_loss.item()
# =================================================================================== #
# 3.训练生成器 #
# =================================================================================== #
# 进行模运算，判读更新时间
if (i + 1) % self.n_critic == 0:
# 源至目标域
# 利用真实样本和目标标签生成生成样本
x_fake = self.G(x_real, label_trg)
# 判别生成样本与目标标签
g_out_src = self.D(x_fake, label_trg)
# 将生成与目标标签的损失与相同大小纯1张量计算交叉熵得到生成G损失
g_loss_fake = F.binary_cross_entropy_with_logits(
input=g_out_src,
target=torch.ones_like(g_out_src, dtype=torch.float))
# 得到真实样本通过域分类器得到的类别
out_cls = self.C(x_real)
# 计算C计算类别与输入的类别的交叉熵损失即G的分类损失
g_loss_cls = CELoss(input=out_cls, target=speaker_idx_org)
# 目标至源域
# 通过G将生成样本转换为源标签
x_reconst = self.G(x_fake, label_org)
# 得到循环一致性损失，即通过G转回来的损失，按道理这两个是同样的
# l1_loss为L1损失函数，即平均绝对误差
g_loss_rec = F.l1_loss(x_reconst, x_real)
# 源到源域(身份一致性损失).
# 通过真实样本与源标签生成，按道理也是生成x_real
x_fake_iden = self.G(x_real, label_org)
# 利用L1损失函数计算
id_loss = F.l1_loss(x_fake_iden, x_real)
# 后退和优化
# 得到生成器的总体损失函数
g_loss = g_loss_fake + self.lambda_cycle * g_loss_rec +\
self.lambda_cls * g_loss_cls + self.lambda_identity * id_loss
# 重置梯度变化缓冲区
self.reset_grad()
# 对G损失求导
g_loss.backward()
# 更新生成器参数
self.g_optimizer.step()
# 记录对应的损失
loss['G/loss_fake'] = g_loss_fake.item()
loss['G/loss_rec'] = g_loss_rec.item()
loss['G/loss_cls'] = g_loss_cls.item()
loss['G/loss_id'] = id_loss.item()
loss['G/g_loss'] = g_loss.item()
# =================================================================================== #
# 4.其他 #
# =================================================================================== #
# 打印训练相关信息
if (i + 1) % self.log_step == 0:
# 得到训练时间
et = datetime.now() - start_time
# 截取后面的时间段
et = str(et)[:-7]
# 耗时与迭代次数
log = "耗时：[{}], 迭代次数：[{}/{}]".format(et, i + 1, self.num_iters)
# 打印对应损失值
for tag, value in loss.items():
log += ", {}: {:.4f}".format(tag, value)
print(log)
# 如果调用tensorboard来记录训练过程
if self.use_tensorboard:
for tag, value in loss.items():
# 添加到log中
self.logger.scalar_summary(tag, value, i + 1)
# 翻译固定数据进行调试
if (i + 1) % self.sample_step == 0:
# torch.no_grad是一个上下文管理器，被该语句包括起来的部分将不会track 梯度
# 所有依赖他的tensor会全部变成True，反向传播时就不会自动求导了，反向传播就不会保存梯度，因此大大节约了显存或者说内存。
with torch.no_grad():
# 调用自定义方法，定义一个路由，并随机选取一个发音者作为测试数据
d, speaker = TestSet(self.test_dir).test_data()
# random.choice返回参数的随机项
# 随机在speakers中选择一个不是目标的发音者
target = random.choice(
[x for x in speakers if x != speaker])
# 将二值化的标签组取出第一个作为目标
# LabelBinary.transfrom方法将复杂类标签转换为二进制标签
label_t = self.spk_enc.transform([target])[0]
# np.asarray将python原生列表或元组形式的现有数据来创建numpy数组
label_t = np.asarray([label_t])
# 取出字典中的文件名与内容
for filename, content in d.items():
f0 = content['f0']
ap = content['ap']
# 调用自定义方法处理对应的数据
sp_norm_pad = self.pad_coded_sp(
content['coded_sp_norm'])
convert_result = []
for start_idx in range(
0, sp_norm_pad.shape[1] - FRAMES + 1, FRAMES):
one_seg = sp_norm_pad[:,
start_idx:start_idx + FRAMES]
one_seg = torch.FloatTensor(one_seg).to(
self.device)
one_seg = one_seg.view(1, 1, one_seg.size(0),
one_seg.size(1))
l = torch.FloatTensor(label_t)
one_seg = one_seg.to(self.device)
l = l.to(self.device)
one_set_return = self.G(one_seg,
l).data.cpu().numpy()
one_set_return = np.squeeze(one_set_return)
one_set_return = norm.backward_process(
one_set_return, target)
convert_result.append(one_set_return)
convert_con = np.concatenate(convert_result, axis=1)
convert_con = convert_con[:,
0:content['coded_sp_norm'].
shape[1]]
contigu = np.ascontiguousarray(convert_con.T,
dtype=np.float64)
decoded_sp = decode_spectral_envelope(contigu,
SAMPLE_RATE,
fft_size=FFTSIZE)
f0_converted = norm.pitch_conversion(
f0, speaker, target)
wav = synthesize(f0_converted, decoded_sp, ap,
SAMPLE_RATE)
name = f'{speaker}-{target}_iter{i+1}_{filename}'
path = os.path.join(self.sample_dir, name)
print(f'[save]:{path}')
librosa.output.write_wav(path, wav, SAMPLE_RATE)
# 保存模型检查点
if (i + 1) % self.model_save_step == 0:
G_path = os.path.join(self.model_save_dir,
'{}-G.ckpt'.format(i + 1))
D_path = os.path.join(self.model_save_dir,
'{}-D.ckpt'.format(i + 1))
C_path = os.path.join(self.model_save_dir,
'{}-C.ckpt'.format(i + 1))
torch.save(self.G.state_dict(), G_path)
torch.save(self.D.state_dict(), D_path)
torch.save(self.C.state_dict(), C_path)
print('Saved model checkpoints into {}...'.format(
self.model_save_dir))
# 衰减学习率
if (i + 1) % self.lr_update_step == 0 and (i + 1) > (
self.num_iters - self.num_iters_decay):
g_lr -= (self.g_lr / float(self.num_iters_decay))
d_lr -= (self.d_lr / float(self.num_iters_decay))
c_lr -= (self.c_lr / float(self.num_iters_decay))
self.update_lr(g_lr, d_lr, c_lr)
print('Decayed learning rates, g_lr: {}, d_lr: {}.'.format(
g_lr, d_lr))
def train(self):
# Learning rate cache for decaying.
g_lr = self.g_lr
d_lr = self.d_lr
c_lr = self.c_lr
start_iters = 0
if self.resume_iters:
pass
norm = Normalizer()
data_iter = iter(self.data_loader)
print('Start training......')
start_time = datetime.now()
for i in range(start_iters, self.num_iters):
# =================================================================================== #
# 1. Preprocess input data #
# =================================================================================== #
# Fetch real images and labels.
try:
x_real, speaker_idx_org, label_org = next(data_iter)
except:
data_iter = iter(self.data_loader)
x_real, speaker_idx_org, label_org = next(data_iter)
# Generate target domain labels randomly.
rand_idx = torch.randperm(label_org.size(0))
label_trg = label_org[rand_idx]
speaker_idx_trg = speaker_idx_org[rand_idx]
x_real = x_real.to(self.device) # Input images.
label_org = label_org.to(
self.device) # Original domain one-hot labels.
label_trg = label_trg.to(
self.device) # Target domain one-hot labels.
speaker_idx_org = speaker_idx_org.to(
self.device) # Original domain labels
speaker_idx_trg = speaker_idx_trg.to(
self.device) #Target domain labels
# =================================================================================== #
# 2. Train the discriminator #
# =================================================================================== #
# Compute loss with real audio frame.
CELoss = nn.CrossEntropyLoss()
cls_real = self.C(x_real)
cls_loss_real = CELoss(input=cls_real, target=speaker_idx_org)
self.reset_grad()
cls_loss_real.backward()
self.c_optimizer.step()
# Logging.
loss = {}
loss['C/C_loss'] = cls_loss_real.item()
out_r = self.D(x_real, label_org)
# Compute loss with fake audio frame.
x_fake = self.G(x_real, label_trg)
out_f = self.D(x_fake.detach(), label_trg)
d_loss_t = F.binary_cross_entropy_with_logits(input=out_f,target=torch.zeros_like(out_f, dtype=torch.float)) + \
F.binary_cross_entropy_with_logits(input=out_r, target=torch.ones_like(out_r, dtype=torch.float))
out_cls = self.C(x_fake)
d_loss_cls = CELoss(input=out_cls, target=speaker_idx_trg)
# Compute loss for gradient penalty.
alpha = torch.rand(x_real.size(0), 1, 1, 1).to(self.device)
x_hat = (alpha * x_real.data +
(1 - alpha) * x_fake.data).requires_grad_(True)
out_src = self.D(x_hat, label_trg)
d_loss_gp = self.gradient_penalty(out_src, x_hat)
d_loss = d_loss_t + self.lambda_cls * d_loss_cls + 5 * d_loss_gp
self.reset_grad()
d_loss.backward()
self.d_optimizer.step()
# loss['D/d_loss_t'] = d_loss_t.item()
# loss['D/loss_cls'] = d_loss_cls.item()
# loss['D/D_gp'] = d_loss_gp.item()
loss['D/D_loss'] = d_loss.item()
# =================================================================================== #
# 3. Train the generator #
# =================================================================================== #
if (i + 1) % self.n_critic == 0:
# Original-to-target domain.
x_fake = self.G(x_real, label_trg)
g_out_src = self.D(x_fake, label_trg)
g_loss_fake = F.binary_cross_entropy_with_logits(
input=g_out_src,
target=torch.ones_like(g_out_src, dtype=torch.float))
out_cls = self.C(x_fake)
g_loss_cls = CELoss(input=out_cls, target=speaker_idx_trg)
# Target-to-original domain.
x_reconst = self.G(x_fake, label_org)
g_loss_rec = F.l1_loss(x_reconst, x_real)
# Original-to-Original domain(identity).
x_fake_iden = self.G(x_real, label_org)
id_loss = F.l1_loss(x_fake_iden, x_real)
# Backward and optimize.
g_loss = g_loss_fake + self.lambda_cycle * g_loss_rec +\
self.lambda_cls * g_loss_cls + self.lambda_identity * id_loss
self.reset_grad()
g_loss.backward()
self.g_optimizer.step()
# Logging.
loss['G/loss_fake'] = g_loss_fake.item()
loss['G/loss_rec'] = g_loss_rec.item()
loss['G/loss_cls'] = g_loss_cls.item()
loss['G/loss_id'] = id_loss.item()
loss['G/g_loss'] = g_loss.item()
# =================================================================================== #
# 4. Miscellaneous #
# =================================================================================== #
# Print out training information.
if (i + 1) % self.log_step == 0:
et = datetime.now() - start_time
et = str(et)[:-7]
log = "Elapsed [{}], Iteration [{}/{}]".format(
et, i + 1, self.num_iters)
for tag, value in loss.items():
log += ", {}: {:.4f}".format(tag, value)
print(log)
if self.use_tensorboard:
for tag, value in loss.items():
self.logger.scalar_summary(tag, value, i + 1)
# Translate fixed images for debugging.
if (i + 1) % self.sample_step == 0:
with torch.no_grad():
d, speaker = TestSet(self.test_dir).test_data()
target = random.choice(
[x for x in speakers if x != speaker])
label_t = self.spk_enc.transform([target])[0]
label_t = np.asarray([label_t])
for filename, content in d.items():
f0 = content['f0']
ap = content['ap']
sp_norm_pad = self.pad_coded_sp(
content['coded_sp_norm'])
convert_result = []
for start_idx in range(
0, sp_norm_pad.shape[1] - FRAMES + 1, FRAMES):
one_seg = sp_norm_pad[:,
start_idx:start_idx + FRAMES]
one_seg = torch.FloatTensor(one_seg).to(
self.device)
one_seg = one_seg.view(1, 1, one_seg.size(0),
one_seg.size(1))
l = torch.FloatTensor(label_t)
one_seg = one_seg.to(self.device)
l = l.to(self.device)
one_set_return = self.G(one_seg,
l).data.cpu().numpy()
one_set_return = np.squeeze(one_set_return)
one_set_return = norm.backward_process(
one_set_return, target)
convert_result.append(one_set_return)
convert_con = np.concatenate(convert_result, axis=1)
convert_con = convert_con[:,
0:content['coded_sp_norm'].
shape[1]]
contigu = np.ascontiguousarray(convert_con.T,
dtype=np.float64)
decoded_sp = decode_spectral_envelope(contigu,
SAMPLE_RATE,
fft_size=FFTSIZE)
f0_converted = norm.pitch_conversion(
f0, speaker, target)
wav = synthesize(f0_converted, decoded_sp, ap,
SAMPLE_RATE)
name = f'{speaker}-{target}_iter{i+1}_{filename}'
path = os.path.join(self.sample_dir, name)
print(f'[save]:{path}')
librosa.output.write_wav(path, wav, SAMPLE_RATE)
# Save model checkpoints.
if (i + 1) % self.model_save_step == 0:
G_path = os.path.join(self.model_save_dir,
'{}-G.ckpt'.format(i + 1))
D_path = os.path.join(self.model_save_dir,
'{}-D.ckpt'.format(i + 1))
C_path = os.path.join(self.model_save_dir,
'{}-C.ckpt'.format(i + 1))
torch.save(self.G.state_dict(), G_path)
torch.save(self.D.state_dict(), D_path)
torch.save(self.C.state_dict(), C_path)
print('Saved model checkpoints into {}...'.format(
self.model_save_dir))
# Decay learning rates.
if (i + 1) % self.lr_update_step == 0 and (i + 1) > (
self.num_iters - self.num_iters_decay):
g_lr -= (self.g_lr / float(self.num_iters_decay))
d_lr -= (self.d_lr / float(self.num_iters_decay))
c_lr -= (self.c_lr / float(self.num_iters_decay))
self.update_lr(g_lr, d_lr, c_lr)
print('Decayed learning rates, g_lr: {}, d_lr: {}.'.format(
g_lr, d_lr))
