Clicko777/RVC_HFv2

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 kl_divergence(m_p, logs_p, m_q, logs_q):25    """KL(P||Q)"""26    kl = (logs_q - logs_p) - 0.527    kl += (28        0.5 * (torch.exp(2.0 * logs_p) + ((m_p - m_q) ** 2)) * torch.exp(-2.0 * logs_q)29    )30    return kl31 32 33def rand_gumbel(shape):34    """Sample from the Gumbel distribution, protect from overflows."""35    uniform_samples = torch.rand(shape) * 0.99998 + 0.0000136    return -torch.log(-torch.log(uniform_samples))37 38 39def rand_gumbel_like(x):40    g = rand_gumbel(x.size()).to(dtype=x.dtype, device=x.device)41    return g42 43 44def slice_segments(x, ids_str, segment_size=4):45    ret = torch.zeros_like(x[:, :, :segment_size])46    for i in range(x.size(0)):47        idx_str = ids_str[i]48        idx_end = idx_str + segment_size49        ret[i] = x[i, :, idx_str:idx_end]50    return ret51 52 53def slice_segments2(x, ids_str, segment_size=4):54    ret = torch.zeros_like(x[:, :segment_size])55    for i in range(x.size(0)):56        idx_str = ids_str[i]57        idx_end = idx_str + segment_size58        ret[i] = x[i, idx_str:idx_end]59    return ret60 61 62def rand_slice_segments(x, x_lengths=None, segment_size=4):63    b, d, t = x.size()64    if x_lengths is None:65        x_lengths = t66    ids_str_max = x_lengths - segment_size + 167    ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)68    ret = slice_segments(x, ids_str, segment_size)69    return ret, ids_str70 71 72def get_timing_signal_1d(length, channels, min_timescale=1.0, max_timescale=1.0e4):73    position = torch.arange(length, dtype=torch.float)74    num_timescales = channels // 275    log_timescale_increment = math.log(float(max_timescale) / float(min_timescale)) / (76        num_timescales - 177    )78    inv_timescales = min_timescale * torch.exp(79        torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment80    )81    scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)82    signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0)83    signal = F.pad(signal, [0, 0, 0, channels % 2])84    signal = signal.view(1, channels, length)85    return signal86 87 88def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4):89    b, channels, length = x.size()90    signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)91    return x + signal.to(dtype=x.dtype, device=x.device)92 93 94def cat_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4, axis=1):95    b, channels, length = x.size()96    signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)97    return torch.cat([x, signal.to(dtype=x.dtype, device=x.device)], axis)98 99 100def subsequent_mask(length):101    mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)102    return mask103 104 105@torch.jit.script106def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):107    n_channels_int = n_channels[0]108    in_act = input_a + input_b109    t_act = torch.tanh(in_act[:, :n_channels_int, :])110    s_act = torch.sigmoid(in_act[:, n_channels_int:, :])111    acts = t_act * s_act112    return acts113 114 115def convert_pad_shape(pad_shape):116    l = pad_shape[::-1]117    pad_shape = [item for sublist in l for item in sublist]118    return pad_shape119 120 121def shift_1d(x):122    x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]123    return x124 125 126def sequence_mask(length, max_length=None):127    if max_length is None:128        max_length = length.max()129    x = torch.arange(max_length, dtype=length.dtype, device=length.device)130    return x.unsqueeze(0) < length.unsqueeze(1)131 132 133def generate_path(duration, mask):134    """135    duration: [b, 1, t_x]136    mask: [b, 1, t_y, t_x]137    """138    device = duration.device139 140    b, _, t_y, t_x = mask.shape141    cum_duration = torch.cumsum(duration, -1)142 143    cum_duration_flat = cum_duration.view(b * t_x)144    path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype)145    path = path.view(b, t_x, t_y)146    path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]147    path = path.unsqueeze(1).transpose(2, 3) * mask148    return path149 150 151def clip_grad_value_(parameters, clip_value, norm_type=2):152    if isinstance(parameters, torch.Tensor):153        parameters = [parameters]154    parameters = list(filter(lambda p: p.grad is not None, parameters))155    norm_type = float(norm_type)156    if clip_value is not None:157        clip_value = float(clip_value)158 159    total_norm = 0160    for p in parameters:161        param_norm = p.grad.data.norm(norm_type)162        total_norm += param_norm.item() ** norm_type163        if clip_value is not None:164            p.grad.data.clamp_(min=-clip_value, max=clip_value)165    total_norm = total_norm ** (1.0 / norm_type)166    return total_norm167