Doiyan/vits-models

1import math2import torch3from torch.nn import functional as F4import torch.jit5 6 7def script_method(fn, _rcb=None):8  return fn9 10 11def script(obj, optimize=True, _frames_up=0, _rcb=None):12  return obj13 14 15torch.jit.script_method = script_method16torch.jit.script = script17 18 19def init_weights(m, mean=0.0, std=0.01):20  classname = m.__class__.__name__21  if classname.find("Conv") != -1:22    m.weight.data.normal_(mean, std)23 24 25def get_padding(kernel_size, dilation=1):26  return int((kernel_size*dilation - dilation)/2)27 28 29def convert_pad_shape(pad_shape):30  l = pad_shape[::-1]31  pad_shape = [item for sublist in l for item in sublist]32  return pad_shape33 34 35def intersperse(lst, item):36  result = [item] * (len(lst) * 2 + 1)37  result[1::2] = lst38  return result39 40 41def kl_divergence(m_p, logs_p, m_q, logs_q):42  """KL(P||Q)"""43  kl = (logs_q - logs_p) - 0.544  kl += 0.5 * (torch.exp(2. * logs_p) + ((m_p - m_q)**2)) * torch.exp(-2. * logs_q)45  return kl46 47 48def rand_gumbel(shape):49  """Sample from the Gumbel distribution, protect from overflows."""50  uniform_samples = torch.rand(shape) * 0.99998 + 0.0000151  return -torch.log(-torch.log(uniform_samples))52 53 54def rand_gumbel_like(x):55  g = rand_gumbel(x.size()).to(dtype=x.dtype, device=x.device)56  return g57 58 59def slice_segments(x, ids_str, segment_size=4):60  ret = torch.zeros_like(x[:, :, :segment_size])61  for i in range(x.size(0)):62    idx_str = ids_str[i]63    idx_end = idx_str + segment_size64    ret[i] = x[i, :, idx_str:idx_end]65  return ret66 67 68def rand_slice_segments(x, x_lengths=None, segment_size=4):69  b, d, t = x.size()70  if x_lengths is None:71    x_lengths = t72  ids_str_max = x_lengths - segment_size + 173  ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)74  ret = slice_segments(x, ids_str, segment_size)75  return ret, ids_str76 77 78def get_timing_signal_1d(79    length, channels, min_timescale=1.0, max_timescale=1.0e4):80  position = torch.arange(length, dtype=torch.float)81  num_timescales = channels // 282  log_timescale_increment = (83      math.log(float(max_timescale) / float(min_timescale)) /84      (num_timescales - 1))85  inv_timescales = min_timescale * torch.exp(86      torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment)87  scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)88  signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0)89  signal = F.pad(signal, [0, 0, 0, channels % 2])90  signal = signal.view(1, channels, length)91  return signal92 93 94def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4):95  b, channels, length = x.size()96  signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)97  return x + signal.to(dtype=x.dtype, device=x.device)98 99 100def cat_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4, axis=1):101  b, channels, length = x.size()102  signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)103  return torch.cat([x, signal.to(dtype=x.dtype, device=x.device)], axis)104 105 106def subsequent_mask(length):107  mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)108  return mask109 110 111@torch.jit.script112def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):113  n_channels_int = n_channels[0]114  in_act = input_a + input_b115  t_act = torch.tanh(in_act[:, :n_channels_int, :])116  s_act = torch.sigmoid(in_act[:, n_channels_int:, :])117  acts = t_act * s_act118  return acts119 120 121def convert_pad_shape(pad_shape):122  l = pad_shape[::-1]123  pad_shape = [item for sublist in l for item in sublist]124  return pad_shape125 126 127def shift_1d(x):128  x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]129  return x130 131 132def sequence_mask(length, max_length=None):133  if max_length is None:134    max_length = length.max()135  x = torch.arange(max_length, dtype=length.dtype, device=length.device)136  return x.unsqueeze(0) < length.unsqueeze(1)137 138 139def generate_path(duration, mask):140  """141  duration: [b, 1, t_x]142  mask: [b, 1, t_y, t_x]143  """144  device = duration.device145  146  b, _, t_y, t_x = mask.shape147  cum_duration = torch.cumsum(duration, -1)148  149  cum_duration_flat = cum_duration.view(b * t_x)150  path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype)151  path = path.view(b, t_x, t_y)152  path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]153  path = path.unsqueeze(1).transpose(2,3) * mask154  return path155 156 157def clip_grad_value_(parameters, clip_value, norm_type=2):158  if isinstance(parameters, torch.Tensor):159    parameters = [parameters]160  parameters = list(filter(lambda p: p.grad is not None, parameters))161  norm_type = float(norm_type)162  if clip_value is not None:163    clip_value = float(clip_value)164 165  total_norm = 0166  for p in parameters:167    param_norm = p.grad.data.norm(norm_type)168    total_norm += param_norm.item() ** norm_type169    if clip_value is not None:170      p.grad.data.clamp_(min=-clip_value, max=clip_value)171  total_norm = total_norm ** (1. / norm_type)172  return total_norm173