Clicko777/RVC_HFv2

1import torch2from torch.nn import functional as F3 4import numpy as np5 6 7DEFAULT_MIN_BIN_WIDTH = 1e-38DEFAULT_MIN_BIN_HEIGHT = 1e-39DEFAULT_MIN_DERIVATIVE = 1e-310 11 12def piecewise_rational_quadratic_transform(13    inputs,14    unnormalized_widths,15    unnormalized_heights,16    unnormalized_derivatives,17    inverse=False,18    tails=None,19    tail_bound=1.0,20    min_bin_width=DEFAULT_MIN_BIN_WIDTH,21    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,22    min_derivative=DEFAULT_MIN_DERIVATIVE,23):24    if tails is None:25        spline_fn = rational_quadratic_spline26        spline_kwargs = {}27    else:28        spline_fn = unconstrained_rational_quadratic_spline29        spline_kwargs = {"tails": tails, "tail_bound": tail_bound}30 31    outputs, logabsdet = spline_fn(32        inputs=inputs,33        unnormalized_widths=unnormalized_widths,34        unnormalized_heights=unnormalized_heights,35        unnormalized_derivatives=unnormalized_derivatives,36        inverse=inverse,37        min_bin_width=min_bin_width,38        min_bin_height=min_bin_height,39        min_derivative=min_derivative,40        **spline_kwargs41    )42    return outputs, logabsdet43 44 45def searchsorted(bin_locations, inputs, eps=1e-6):46    bin_locations[..., -1] += eps47    return torch.sum(inputs[..., None] >= bin_locations, dim=-1) - 148 49 50def unconstrained_rational_quadratic_spline(51    inputs,52    unnormalized_widths,53    unnormalized_heights,54    unnormalized_derivatives,55    inverse=False,56    tails="linear",57    tail_bound=1.0,58    min_bin_width=DEFAULT_MIN_BIN_WIDTH,59    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,60    min_derivative=DEFAULT_MIN_DERIVATIVE,61):62    inside_interval_mask = (inputs >= -tail_bound) & (inputs <= tail_bound)63    outside_interval_mask = ~inside_interval_mask64 65    outputs = torch.zeros_like(inputs)66    logabsdet = torch.zeros_like(inputs)67 68    if tails == "linear":69        unnormalized_derivatives = F.pad(unnormalized_derivatives, pad=(1, 1))70        constant = np.log(np.exp(1 - min_derivative) - 1)71        unnormalized_derivatives[..., 0] = constant72        unnormalized_derivatives[..., -1] = constant73 74        outputs[outside_interval_mask] = inputs[outside_interval_mask]75        logabsdet[outside_interval_mask] = 076    else:77        raise RuntimeError("{} tails are not implemented.".format(tails))78 79    (80        outputs[inside_interval_mask],81        logabsdet[inside_interval_mask],82    ) = rational_quadratic_spline(83        inputs=inputs[inside_interval_mask],84        unnormalized_widths=unnormalized_widths[inside_interval_mask, :],85        unnormalized_heights=unnormalized_heights[inside_interval_mask, :],86        unnormalized_derivatives=unnormalized_derivatives[inside_interval_mask, :],87        inverse=inverse,88        left=-tail_bound,89        right=tail_bound,90        bottom=-tail_bound,91        top=tail_bound,92        min_bin_width=min_bin_width,93        min_bin_height=min_bin_height,94        min_derivative=min_derivative,95    )96 97    return outputs, logabsdet98 99 100def rational_quadratic_spline(101    inputs,102    unnormalized_widths,103    unnormalized_heights,104    unnormalized_derivatives,105    inverse=False,106    left=0.0,107    right=1.0,108    bottom=0.0,109    top=1.0,110    min_bin_width=DEFAULT_MIN_BIN_WIDTH,111    min_bin_height=DEFAULT_MIN_BIN_HEIGHT,112    min_derivative=DEFAULT_MIN_DERIVATIVE,113):114    if torch.min(inputs) < left or torch.max(inputs) > right:115        raise ValueError("Input to a transform is not within its domain")116 117    num_bins = unnormalized_widths.shape[-1]118 119    if min_bin_width * num_bins > 1.0:120        raise ValueError("Minimal bin width too large for the number of bins")121    if min_bin_height * num_bins > 1.0:122        raise ValueError("Minimal bin height too large for the number of bins")123 124    widths = F.softmax(unnormalized_widths, dim=-1)125    widths = min_bin_width + (1 - min_bin_width * num_bins) * widths126    cumwidths = torch.cumsum(widths, dim=-1)127    cumwidths = F.pad(cumwidths, pad=(1, 0), mode="constant", value=0.0)128    cumwidths = (right - left) * cumwidths + left129    cumwidths[..., 0] = left130    cumwidths[..., -1] = right131    widths = cumwidths[..., 1:] - cumwidths[..., :-1]132 133    derivatives = min_derivative + F.softplus(unnormalized_derivatives)134 135    heights = F.softmax(unnormalized_heights, dim=-1)136    heights = min_bin_height + (1 - min_bin_height * num_bins) * heights137    cumheights = torch.cumsum(heights, dim=-1)138    cumheights = F.pad(cumheights, pad=(1, 0), mode="constant", value=0.0)139    cumheights = (top - bottom) * cumheights + bottom140    cumheights[..., 0] = bottom141    cumheights[..., -1] = top142    heights = cumheights[..., 1:] - cumheights[..., :-1]143 144    if inverse:145        bin_idx = searchsorted(cumheights, inputs)[..., None]146    else:147        bin_idx = searchsorted(cumwidths, inputs)[..., None]148 149    input_cumwidths = cumwidths.gather(-1, bin_idx)[..., 0]150    input_bin_widths = widths.gather(-1, bin_idx)[..., 0]151 152    input_cumheights = cumheights.gather(-1, bin_idx)[..., 0]153    delta = heights / widths154    input_delta = delta.gather(-1, bin_idx)[..., 0]155 156    input_derivatives = derivatives.gather(-1, bin_idx)[..., 0]157    input_derivatives_plus_one = derivatives[..., 1:].gather(-1, bin_idx)[..., 0]158 159    input_heights = heights.gather(-1, bin_idx)[..., 0]160 161    if inverse:162        a = (inputs - input_cumheights) * (163            input_derivatives + input_derivatives_plus_one - 2 * input_delta164        ) + input_heights * (input_delta - input_derivatives)165        b = input_heights * input_derivatives - (inputs - input_cumheights) * (166            input_derivatives + input_derivatives_plus_one - 2 * input_delta167        )168        c = -input_delta * (inputs - input_cumheights)169 170        discriminant = b.pow(2) - 4 * a * c171        assert (discriminant >= 0).all()172 173        root = (2 * c) / (-b - torch.sqrt(discriminant))174        outputs = root * input_bin_widths + input_cumwidths175 176        theta_one_minus_theta = root * (1 - root)177        denominator = input_delta + (178            (input_derivatives + input_derivatives_plus_one - 2 * input_delta)179            * theta_one_minus_theta180        )181        derivative_numerator = input_delta.pow(2) * (182            input_derivatives_plus_one * root.pow(2)183            + 2 * input_delta * theta_one_minus_theta184            + input_derivatives * (1 - root).pow(2)185        )186        logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)187 188        return outputs, -logabsdet189    else:190        theta = (inputs - input_cumwidths) / input_bin_widths191        theta_one_minus_theta = theta * (1 - theta)192 193        numerator = input_heights * (194            input_delta * theta.pow(2) + input_derivatives * theta_one_minus_theta195        )196        denominator = input_delta + (197            (input_derivatives + input_derivatives_plus_one - 2 * input_delta)198            * theta_one_minus_theta199        )200        outputs = input_cumheights + numerator / denominator201 202        derivative_numerator = input_delta.pow(2) * (203            input_derivatives_plus_one * theta.pow(2)204            + 2 * input_delta * theta_one_minus_theta205            + input_derivatives * (1 - theta).pow(2)206        )207        logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)208 209        return outputs, logabsdet210