cocktailpeanut/DiffRhythm

1"""2ein notation:3b - batch4n - sequence5nt - text sequence6nw - raw wave length7d - dimension8"""9 10from __future__ import annotations11from typing import Callable12from random import random13 14import torch15from torch import nn16import torch17import torch.nn.functional as F18from torch.nn.utils.rnn import pad_sequence19 20from torchdiffeq import odeint21 22from diffrhythm.model.modules import MelSpec23from diffrhythm.model.utils import (24    default,25    exists,26    list_str_to_idx,27    list_str_to_tensor,28    lens_to_mask,29    mask_from_frac_lengths,30)31 32def custom_mask_from_start_end_indices(seq_len: int["b"], start: int["b"], end: int["b"], device, max_seq_len):  # noqa: F722 F82133    max_seq_len = max_seq_len34    seq = torch.arange(max_seq_len, device=device).long()35    start_mask = seq[None, :] >= start[:, None]36    end_mask = seq[None, :] < end[:, None]37    return start_mask & end_mask38 39class CFM(nn.Module):40    def __init__(41        self,42        transformer: nn.Module,43        sigma=0.0,44        odeint_kwargs: dict = dict(45            # atol = 1e-5,46            # rtol = 1e-5,47            method="euler" # 'midpoint'48            # method="adaptive_heun"  # dopri549        ),50        odeint_options: dict = dict(51            min_step=0.0552        ),53        audio_drop_prob=0.3,54        cond_drop_prob=0.2,55        style_drop_prob=0.1,56        lrc_drop_prob=0.1,57        num_channels=None,58        frac_lengths_mask: tuple[float, float] = (0.7, 1.0),59        vocab_char_map: dict[str:int] | None = None,60        use_style_prompt: bool = False61    ):62        super().__init__()63 64        self.frac_lengths_mask = frac_lengths_mask65 66        self.num_channels = num_channels67 68        # classifier-free guidance69        self.audio_drop_prob = audio_drop_prob70        self.cond_drop_prob = cond_drop_prob71        self.style_drop_prob = style_drop_prob72        self.lrc_drop_prob = lrc_drop_prob73 74        print(f"audio drop prob -> {self.audio_drop_prob}; style_drop_prob -> {self.style_drop_prob}; lrc_drop_prob: {self.lrc_drop_prob}")75 76        # transformer77        self.transformer = transformer78        dim = transformer.dim79        self.dim = dim80 81        # conditional flow related82        self.sigma = sigma83 84        # sampling related85        self.odeint_kwargs = odeint_kwargs86        # print(f"ODE SOLVER: {self.odeint_kwargs['method']}")87        88        self.odeint_options = odeint_options89 90        # vocab map for tokenization91        self.vocab_char_map = vocab_char_map92 93        self.use_style_prompt = use_style_prompt94 95    @property96    def device(self):97        return next(self.parameters()).device98 99    @torch.no_grad()100    def sample(101        self,102        cond: float["b n d"] | float["b nw"],  # noqa: F722103        text: int["b nt"] | list[str],  # noqa: F722104        duration: int | int["b"],  # noqa: F821105        *,106        style_prompt = None,107        style_prompt_lens = None,108        negative_style_prompt = None,109        lens: int["b"] | None = None,  # noqa: F821110        steps=32,111        cfg_strength=4.0,112        sway_sampling_coef=None,113        seed: int | None = None,114        max_duration=4096,115        #max_duration=6144,116        vocoder: Callable[[float["b d n"]], float["b nw"]] | None = None,  # noqa: F722117        no_ref_audio=False,118        duplicate_test=False,119        t_inter=0.1,120        edit_mask=None,121        start_time=None,122        latent_pred_start_frame=0,123        latent_pred_end_frame=2048,124        vocal_flag=False,125        odeint_method="euler"126    ):127        self.eval()128 129        if next(self.parameters()).dtype == torch.float16:130            cond = cond.half()131 132        # raw wave133        134        if cond.shape[1] > duration:135            cond = cond[:, :duration, :]136 137        if cond.ndim == 2:138            cond = self.mel_spec(cond)139            cond = cond.permute(0, 2, 1)140            assert cond.shape[-1] == self.num_channels141 142        batch, cond_seq_len, device = *cond.shape[:2], cond.device143        if not exists(lens):144            lens = torch.full((batch,), cond_seq_len, device=device, dtype=torch.long)145 146        # text147 148        if isinstance(text, list):149            if exists(self.vocab_char_map):150                text = list_str_to_idx(text, self.vocab_char_map).to(device)151            else:152                text = list_str_to_tensor(text).to(device)153            assert text.shape[0] == batch154 155        if exists(text):156            text_lens = (text != -1).sum(dim=-1)157            #lens = torch.maximum(text_lens, lens)  # make sure lengths are at least those of the text characters158 159        # duration160        # import pdb; pdb.set_trace()161        cond_mask = lens_to_mask(lens)162        if edit_mask is not None:163            cond_mask = cond_mask & edit_mask164 165        latent_pred_start_frame = torch.tensor([latent_pred_start_frame]).to(cond.device)166        latent_pred_end_frame = duration167        latent_pred_end_frame = torch.tensor([latent_pred_end_frame]).to(cond.device)168        fixed_span_mask = custom_mask_from_start_end_indices(cond_seq_len, latent_pred_start_frame, latent_pred_end_frame, device=cond.device, max_seq_len=duration)169 170        fixed_span_mask = fixed_span_mask.unsqueeze(-1)171        step_cond = torch.where(fixed_span_mask, torch.zeros_like(cond), cond)172 173        if isinstance(duration, int):174            duration = torch.full((batch,), duration, device=device, dtype=torch.long)175 176        # duration = torch.maximum(lens + 1, duration)  # just add one token so something is generated177        duration = duration.clamp(max=max_duration)178        max_duration = duration.amax()179 180        # duplicate test corner for inner time step oberservation181        if duplicate_test:182            test_cond = F.pad(cond, (0, 0, cond_seq_len, max_duration - 2 * cond_seq_len), value=0.0)183 184        # cond = F.pad(cond, (0, 0, 0, max_duration - cond_seq_len), value=0.0) # [b, t, d]185        # cond_mask = F.pad(cond_mask, (0, max_duration - cond_mask.shape[-1]), value=False) # [b, max_duration]186        # cond_mask = cond_mask.unsqueeze(-1) #[b, t, d]187        # step_cond = torch.where(188        #     cond_mask, cond, torch.zeros_like(cond)189        # )  # allow direct control (cut cond audio) with lens passed in190 191        if batch > 1:192            mask = lens_to_mask(duration)193        else:  # save memory and speed up, as single inference need no mask currently194            mask = None195 196        # test for no ref audio197        if no_ref_audio:198            cond = torch.zeros_like(cond)199        200        start_time_embed, positive_text_embed, positive_text_residuals = self.transformer.forward_timestep_invariant(text, step_cond.shape[1], drop_text=False, start_time=start_time)201        _, negative_text_embed, negative_text_residuals = self.transformer.forward_timestep_invariant(text, step_cond.shape[1], drop_text=True, start_time=start_time)202 203        if vocal_flag:204            style_prompt = negative_style_prompt205            negative_style_prompt = torch.zeros_like(style_prompt)206 207 208        text_embed = torch.cat([positive_text_embed, negative_text_embed], 0)209        text_residuals = [torch.cat([a, b], 0) for a, b in zip(positive_text_residuals, negative_text_residuals)]210        step_cond = torch.cat([step_cond, step_cond], 0)211        style_prompt = torch.cat([style_prompt, negative_style_prompt], 0)212        start_time_embed = torch.cat([start_time_embed, start_time_embed], 0)213            214 215        def fn(t, x):216            x = torch.cat([x, x], 0)217            pred = self.transformer(218                x=x, text_embed=text_embed, text_residuals=text_residuals, cond=step_cond, time=t, 219                drop_audio_cond=True, drop_prompt=False, style_prompt=style_prompt, start_time=start_time_embed220            )221 222            positive_pred, negative_pred = pred.chunk(2, 0)223            cfg_pred = positive_pred + (positive_pred - negative_pred) * cfg_strength224 225            return cfg_pred226 227        # noise input228        # to make sure batch inference result is same with different batch size, and for sure single inference229        # still some difference maybe due to convolutional layers230        y0 = []231        for dur in duration:232            if exists(seed):233                torch.manual_seed(seed)234            y0.append(torch.randn(dur, self.num_channels, device=self.device, dtype=step_cond.dtype))235        y0 = pad_sequence(y0, padding_value=0, batch_first=True)236 237        t_start = 0238 239        # duplicate test corner for inner time step oberservation240        if duplicate_test:241            t_start = t_inter242            y0 = (1 - t_start) * y0 + t_start * test_cond243            steps = int(steps * (1 - t_start))244 245        t = torch.linspace(t_start, 1, steps, device=self.device, dtype=step_cond.dtype)246        if sway_sampling_coef is not None:247            t = t + sway_sampling_coef * (torch.cos(torch.pi / 2 * t) - 1 + t)248 249        trajectory = odeint(fn, y0, t, **self.odeint_kwargs)250 251        sampled = trajectory[-1]252        out = sampled253        # out = torch.where(cond_mask, cond, out)254        out = torch.where(fixed_span_mask, out, cond)255 256        if exists(vocoder):257            out = out.permute(0, 2, 1)258            out = vocoder(out)259 260        return out, trajectory261 262    def forward(263        self,264        inp: float["b n d"] | float["b nw"],  # mel or raw wave  # noqa: F722265        text: int["b nt"] | list[str],  # noqa: F722266        style_prompt = None,267        style_prompt_lens = None,268        lens: int["b"] | None = None,  # noqa: F821269        noise_scheduler: str | None = None,270        grad_ckpt = False,271        start_time = None,272    ):273 274        batch, seq_len, dtype, device, _σ1 = *inp.shape[:2], inp.dtype, self.device, self.sigma275 276        # lens and mask277        if not exists(lens):278            lens = torch.full((batch,), seq_len, device=device)279 280        mask = lens_to_mask(lens, length=seq_len)  # useless here, as collate_fn will pad to max length in batch281 282        # get a random span to mask out for training conditionally283        frac_lengths = torch.zeros((batch,), device=self.device).float().uniform_(*self.frac_lengths_mask)284        rand_span_mask = mask_from_frac_lengths(lens, frac_lengths)285 286        if exists(mask):287            rand_span_mask = mask288            # rand_span_mask &= mask289 290        # mel is x1291        x1 = inp292 293        # x0 is gaussian noise294        x0 = torch.randn_like(x1)295 296        # time step297        # time = torch.rand((batch,), dtype=dtype, device=self.device)298        time = torch.normal(mean=0, std=1, size=(batch,), device=self.device)299        time = torch.nn.functional.sigmoid(time)300        # TODO. noise_scheduler301 302        # sample xt (φ_t(x) in the paper)303        t = time.unsqueeze(-1).unsqueeze(-1)304        φ = (1 - t) * x0 + t * x1305        flow = x1 - x0306 307        # only predict what is within the random mask span for infilling308        cond = torch.where(rand_span_mask[..., None], torch.zeros_like(x1), x1)309 310        # transformer and cfg training with a drop rate311        drop_audio_cond = random() < self.audio_drop_prob  # p_drop in voicebox paper312        drop_text = random() < self.lrc_drop_prob313        drop_prompt = random() < self.style_drop_prob314 315        # if want rigourously mask out padding, record in collate_fn in dataset.py, and pass in here316        # adding mask will use more memory, thus also need to adjust batchsampler with scaled down threshold for long sequences317        pred = self.transformer(318            x=φ, cond=cond, text=text, time=time, drop_audio_cond=drop_audio_cond, drop_text=drop_text, drop_prompt=drop_prompt,319            style_prompt=style_prompt, style_prompt_lens=style_prompt_lens, grad_ckpt=grad_ckpt, start_time=start_time320        )321 322        # flow matching loss323        loss = F.mse_loss(pred, flow, reduction="none")324        loss = loss[rand_span_mask]325 326        return loss.mean(), cond, pred327