text_to_speech/modules/tts/diffspeech/shallow_diffusion_tts.py

import math
import random
from functools import partial
from inspect import isfunction
import numpy as np
import torch
import torch.nn.functional as F
from torch import nn
from tqdm import tqdm

from text_to_speech.modules.tts.fs2_orig import FastSpeech2Orig
from text_to_speech.modules.tts.diffspeech.net import DiffNet
from text_to_speech.modules.tts.commons.align_ops import expand_states


def exists(x):
    return x is not None


def default(val, d):
    if exists(val):
        return val
    return d() if isfunction(d) else d


# gaussian diffusion trainer class

def extract(a, t, x_shape):
    b, *_ = t.shape
    out = a.gather(-1, t)
    return out.reshape(b, *((1,) * (len(x_shape) - 1)))


def noise_like(shape, device, repeat=False):
    repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1)))
    noise = lambda: torch.randn(shape, device=device)
    return repeat_noise() if repeat else noise()


def linear_beta_schedule(timesteps, max_beta=0.01):
    """
    linear schedule
    """
    betas = np.linspace(1e-4, max_beta, timesteps)
    return betas


def cosine_beta_schedule(timesteps, s=0.008):
    """
    cosine schedule
    as proposed in https://openreview.net/forum?id=-NEXDKk8gZ
    """
    steps = timesteps + 1
    x = np.linspace(0, steps, steps)
    alphas_cumprod = np.cos(((x / steps) + s) / (1 + s) * np.pi * 0.5) ** 2
    alphas_cumprod = alphas_cumprod / alphas_cumprod[0]
    betas = 1 - (alphas_cumprod[1:] / alphas_cumprod[:-1])
    return np.clip(betas, a_min=0, a_max=0.999)


beta_schedule = {
    "cosine": cosine_beta_schedule,
    "linear": linear_beta_schedule,
}


DIFF_DECODERS = {
    'wavenet': lambda hp: DiffNet(hp),
}


class AuxModel(FastSpeech2Orig):
    def forward(self, txt_tokens, mel2ph=None, spk_embed=None, spk_id=None,
                f0=None, uv=None, energy=None, infer=False, **kwargs):
        ret = {}
        encoder_out = self.encoder(txt_tokens)  # [B, T, C]
        src_nonpadding = (txt_tokens > 0).float()[:, :, None]
        style_embed = self.forward_style_embed(spk_embed, spk_id)

        # add dur
        dur_inp = (encoder_out + style_embed) * src_nonpadding
        mel2ph = self.forward_dur(dur_inp, mel2ph, txt_tokens, ret)
        tgt_nonpadding = (mel2ph > 0).float()[:, :, None]
        decoder_inp = decoder_inp_ = expand_states(encoder_out, mel2ph)

        # add pitch and energy embed
        if self.hparams['use_pitch_embed']:
            pitch_inp = (decoder_inp_ + style_embed) * tgt_nonpadding
            decoder_inp = decoder_inp + self.forward_pitch(pitch_inp, f0, uv, mel2ph, ret, encoder_out)

        # add pitch and energy embed
        if self.hparams['use_energy_embed']:
            energy_inp = (decoder_inp_ + style_embed) * tgt_nonpadding
            decoder_inp = decoder_inp + self.forward_energy(energy_inp, energy, ret)

        # decoder input
        ret['decoder_inp'] = decoder_inp = (decoder_inp + style_embed) * tgt_nonpadding
        if self.hparams['dec_inp_add_noise']:
            B, T, _ = decoder_inp.shape
            z = kwargs.get('adv_z', torch.randn([B, T, self.z_channels])).to(decoder_inp.device)
            ret['adv_z'] = z
            decoder_inp = torch.cat([decoder_inp, z], -1)
            decoder_inp = self.dec_inp_noise_proj(decoder_inp) * tgt_nonpadding
        if kwargs['skip_decoder']:
            return ret
        ret['mel_out'] = self.forward_decoder(decoder_inp, tgt_nonpadding, ret, infer=infer, **kwargs)
        return ret


class GaussianDiffusion(nn.Module):
    def __init__(self, dict_size, hparams, out_dims=None):
        super().__init__()
        self.hparams = hparams
        out_dims = hparams['audio_num_mel_bins']
        denoise_fn = DIFF_DECODERS[hparams['diff_decoder_type']](hparams)
        timesteps = hparams['timesteps']
        K_step = hparams['K_step']
        loss_type = hparams['diff_loss_type']
        spec_min = hparams['spec_min']
        spec_max = hparams['spec_max']

        self.denoise_fn = denoise_fn
        self.fs2 = AuxModel(dict_size, hparams)
        self.mel_bins = out_dims

        if hparams['schedule_type'] == 'linear':
            betas = linear_beta_schedule(timesteps, hparams['max_beta'])
        else:
            betas = cosine_beta_schedule(timesteps)

        alphas = 1. - betas
        alphas_cumprod = np.cumprod(alphas, axis=0)
        alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])

        timesteps, = betas.shape
        self.num_timesteps = int(timesteps)
        self.K_step = K_step
        self.loss_type = loss_type

        to_torch = partial(torch.tensor, dtype=torch.float32)

        self.register_buffer('betas', to_torch(betas))
        self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
        self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))

        # calculations for diffusion q(x_t | x_{t-1}) and others
        self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
        self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
        self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
        self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
        self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))

        # calculations for posterior q(x_{t-1} | x_t, x_0)
        posterior_variance = betas * (1. - alphas_cumprod_prev) / (1. - alphas_cumprod)
        # above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
        self.register_buffer('posterior_variance', to_torch(posterior_variance))
        # below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
        self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20))))
        self.register_buffer('posterior_mean_coef1', to_torch(
            betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod)))
        self.register_buffer('posterior_mean_coef2', to_torch(
            (1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod)))

        self.register_buffer('spec_min', torch.FloatTensor(spec_min)[None, None, :hparams['keep_bins']])
        self.register_buffer('spec_max', torch.FloatTensor(spec_max)[None, None, :hparams['keep_bins']])

    def q_mean_variance(self, x_start, t):
        mean = extract(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start
        variance = extract(1. - self.alphas_cumprod, t, x_start.shape)
        log_variance = extract(self.log_one_minus_alphas_cumprod, t, x_start.shape)
        return mean, variance, log_variance

    def predict_start_from_noise(self, x_t, t, noise):
        return (
                extract(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
                extract(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
        )

    def q_posterior(self, x_start, x_t, t):
        posterior_mean = (
                extract(self.posterior_mean_coef1, t, x_t.shape) * x_start +
                extract(self.posterior_mean_coef2, t, x_t.shape) * x_t
        )
        posterior_variance = extract(self.posterior_variance, t, x_t.shape)
        posterior_log_variance_clipped = extract(self.posterior_log_variance_clipped, t, x_t.shape)
        return posterior_mean, posterior_variance, posterior_log_variance_clipped

    def p_mean_variance(self, x, t, cond, clip_denoised: bool):
        noise_pred = self.denoise_fn(x, t, cond=cond)
        x_recon = self.predict_start_from_noise(x, t=t, noise=noise_pred)

        if clip_denoised:
            x_recon.clamp_(-1., 1.)

        model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
        return model_mean, posterior_variance, posterior_log_variance

    @torch.no_grad()
    def p_sample(self, x, t, cond, clip_denoised=True, repeat_noise=False):
        b, *_, device = *x.shape, x.device
        model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, cond=cond, clip_denoised=clip_denoised)
        noise = noise_like(x.shape, device, repeat_noise)
        # no noise when t == 0
        nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
        return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise

    def q_sample(self, x_start, t, noise=None):
        noise = default(noise, lambda: torch.randn_like(x_start))
        return (
                extract(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
                extract(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise
        )

    def p_losses(self, x_start, t, cond, noise=None, nonpadding=None):
        noise = default(noise, lambda: torch.randn_like(x_start))

        x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
        x_recon = self.denoise_fn(x_noisy, t, cond)

        if self.loss_type == 'l1':
            if nonpadding is not None:
                loss = ((noise - x_recon).abs() * nonpadding.unsqueeze(1)).mean()
            else:
                # print('are you sure w/o nonpadding?')
                loss = (noise - x_recon).abs().mean()

        elif self.loss_type == 'l2':
            loss = F.mse_loss(noise, x_recon)
        else:
            raise NotImplementedError()

        return loss

    def forward(self, txt_tokens, mel2ph=None, spk_embed=None, spk_id=None,
                ref_mels=None, f0=None, uv=None, energy=None, infer=False, **kwargs):
        b, *_, device = *txt_tokens.shape, txt_tokens.device
        ret = self.fs2(txt_tokens, mel2ph=mel2ph, spk_embed=spk_embed, spk_id=spk_id,
                                f0=f0, uv=uv, energy=energy, infer=infer, skip_decoder=(not infer), **kwargs)
        cond = ret['decoder_inp'].transpose(1, 2)

        if not infer:
            t = torch.randint(0, self.K_step, (b,), device=device).long()
            x = ref_mels
            x = self.norm_spec(x)
            x = x.transpose(1, 2)[:, None, :, :]  # [B, 1, M, T]
            ret['diff_loss'] = self.p_losses(x, t, cond)
            # nonpadding = (mel2ph != 0).float()
            # ret['diff_loss'] = self.p_losses(x, t, cond, nonpadding=nonpadding)
            ret['mel_out'] = None
        else:
            ret['fs2_mel'] = ret['mel_out']
            fs2_mels = ret['mel_out']
            t = self.K_step
            fs2_mels = self.norm_spec(fs2_mels)
            fs2_mels = fs2_mels.transpose(1, 2)[:, None, :, :]

            x = self.q_sample(x_start=fs2_mels, t=torch.tensor([t - 1], device=device).long())
            if self.hparams.get('gaussian_start') is not None and self.hparams['gaussian_start']:
                print('===> gaussian start.')
                shape = (cond.shape[0], 1, self.mel_bins, cond.shape[2])
                x = torch.randn(shape, device=device)
            for i in tqdm(reversed(range(0, t)), desc='sample time step', total=t):
                x = self.p_sample(x, torch.full((b,), i, device=device, dtype=torch.long), cond)
            x = x[:, 0].transpose(1, 2)
            ret['mel_out'] = self.denorm_spec(x)

        return ret

    def norm_spec(self, x):
        return (x - self.spec_min) / (self.spec_max - self.spec_min) * 2 - 1

    def denorm_spec(self, x):
        return (x + 1) / 2 * (self.spec_max - self.spec_min) + self.spec_min

    def cwt2f0_norm(self, cwt_spec, mean, std, mel2ph):
        return self.fs2.cwt2f0_norm(cwt_spec, mean, std, mel2ph)

    def out2mel(self, x):
        return x