kevinwang676/FreeVC-OpenAI-TTS

1import math2import numpy as np3import torch4from torch import nn5from torch.nn import functional as F6 7 8def init_weights(m, mean=0.0, std=0.01):9  classname = m.__class__.__name__10  if classname.find("Conv") != -1:11    m.weight.data.normal_(mean, std)12 13 14def get_padding(kernel_size, dilation=1):15  return int((kernel_size*dilation - dilation)/2)16 17 18def convert_pad_shape(pad_shape):19  l = pad_shape[::-1]20  pad_shape = [item for sublist in l for item in sublist]21  return pad_shape22 23 24def intersperse(lst, item):25  result = [item] * (len(lst) * 2 + 1)26  result[1::2] = lst27  return result28 29 30def kl_divergence(m_p, logs_p, m_q, logs_q):31  """KL(P||Q)"""32  kl = (logs_q - logs_p) - 0.533  kl += 0.5 * (torch.exp(2. * logs_p) + ((m_p - m_q)**2)) * torch.exp(-2. * logs_q)34  return kl35 36 37def rand_gumbel(shape):38  """Sample from the Gumbel distribution, protect from overflows."""39  uniform_samples = torch.rand(shape) * 0.99998 + 0.0000140  return -torch.log(-torch.log(uniform_samples))41 42 43def rand_gumbel_like(x):44  g = rand_gumbel(x.size()).to(dtype=x.dtype, device=x.device)45  return g46 47 48def slice_segments(x, ids_str, segment_size=4):49  ret = torch.zeros_like(x[:, :, :segment_size])50  for i in range(x.size(0)):51    idx_str = ids_str[i]52    idx_end = idx_str + segment_size53    ret[i] = x[i, :, idx_str:idx_end]54  return ret55 56 57def rand_slice_segments(x, x_lengths=None, segment_size=4):58  b, d, t = x.size()59  if x_lengths is None:60    x_lengths = t61  ids_str_max = x_lengths - segment_size + 162  ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)63  ret = slice_segments(x, ids_str, segment_size)64  return ret, ids_str65 66 67def rand_spec_segments(x, x_lengths=None, segment_size=4):68  b, d, t = x.size()69  if x_lengths is None:70    x_lengths = t71  ids_str_max = x_lengths - segment_size72  ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)73  ret = slice_segments(x, ids_str, segment_size)74  return ret, ids_str75 76 77def get_timing_signal_1d(78    length, channels, min_timescale=1.0, max_timescale=1.0e4):79  position = torch.arange(length, dtype=torch.float)80  num_timescales = channels // 281  log_timescale_increment = (82      math.log(float(max_timescale) / float(min_timescale)) /83      (num_timescales - 1))84  inv_timescales = min_timescale * torch.exp(85      torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment)86  scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)87  signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0)88  signal = F.pad(signal, [0, 0, 0, channels % 2])89  signal = signal.view(1, channels, length)90  return signal91 92 93def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4):94  b, channels, length = x.size()95  signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)96  return x + signal.to(dtype=x.dtype, device=x.device)97 98 99def cat_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4, axis=1):100  b, channels, length = x.size()101  signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)102  return torch.cat([x, signal.to(dtype=x.dtype, device=x.device)], axis)103 104 105def subsequent_mask(length):106  mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)107  return mask108 109 110@torch.jit.script111def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):112  n_channels_int = n_channels[0]113  in_act = input_a + input_b114  t_act = torch.tanh(in_act[:, :n_channels_int, :])115  s_act = torch.sigmoid(in_act[:, n_channels_int:, :])116  acts = t_act * s_act117  return acts118 119 120def convert_pad_shape(pad_shape):121  l = pad_shape[::-1]122  pad_shape = [item for sublist in l for item in sublist]123  return pad_shape124 125 126def shift_1d(x):127  x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]128  return x129 130 131def sequence_mask(length, max_length=None):132  if max_length is None:133    max_length = length.max()134  x = torch.arange(max_length, dtype=length.dtype, device=length.device)135  return x.unsqueeze(0) < length.unsqueeze(1)136 137 138def generate_path(duration, mask):139  """140  duration: [b, 1, t_x]141  mask: [b, 1, t_y, t_x]142  """143  device = duration.device144  145  b, _, t_y, t_x = mask.shape146  cum_duration = torch.cumsum(duration, -1)147  148  cum_duration_flat = cum_duration.view(b * t_x)149  path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype)150  path = path.view(b, t_x, t_y)151  path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]152  path = path.unsqueeze(1).transpose(2,3) * mask153  return path154 155 156def clip_grad_value_(parameters, clip_value, norm_type=2):157  if isinstance(parameters, torch.Tensor):158    parameters = [parameters]159  parameters = list(filter(lambda p: p.grad is not None, parameters))160  norm_type = float(norm_type)161  if clip_value is not None:162    clip_value = float(clip_value)163 164  total_norm = 0165  for p in parameters:166    param_norm = p.grad.data.norm(norm_type)167    total_norm += param_norm.item() ** norm_type168    if clip_value is not None:169      p.grad.data.clamp_(min=-clip_value, max=clip_value)170  total_norm = total_norm ** (1. / norm_type)171  return total_norm172