xinference 0.16.3__py3-none-any.whl → 1.2.1__py3-none-any.whl

xinference/thirdparty/f5_tts/model/cfm.py

@@ -0,0 +1,285 @@
+"""
+ein notation:
+b - batch
+n - sequence
+nt - text sequence
+nw - raw wave length
+d - dimension
+"""
+
+from __future__ import annotations
+
+from random import random
+from typing import Callable
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+from torch.nn.utils.rnn import pad_sequence
+from torchdiffeq import odeint
+
+from f5_tts.model.modules import MelSpec
+from f5_tts.model.utils import (
+    default,
+    exists,
+    lens_to_mask,
+    list_str_to_idx,
+    list_str_to_tensor,
+    mask_from_frac_lengths,
+)
+
+
+class CFM(nn.Module):
+    def __init__(
+        self,
+        transformer: nn.Module,
+        sigma=0.0,
+        odeint_kwargs: dict = dict(
+            # atol = 1e-5,
+            # rtol = 1e-5,
+            method="euler"  # 'midpoint'
+        ),
+        audio_drop_prob=0.3,
+        cond_drop_prob=0.2,
+        num_channels=None,
+        mel_spec_module: nn.Module | None = None,
+        mel_spec_kwargs: dict = dict(),
+        frac_lengths_mask: tuple[float, float] = (0.7, 1.0),
+        vocab_char_map: dict[str:int] | None = None,
+    ):
+        super().__init__()
+
+        self.frac_lengths_mask = frac_lengths_mask
+
+        # mel spec
+        self.mel_spec = default(mel_spec_module, MelSpec(**mel_spec_kwargs))
+        num_channels = default(num_channels, self.mel_spec.n_mel_channels)
+        self.num_channels = num_channels
+
+        # classifier-free guidance
+        self.audio_drop_prob = audio_drop_prob
+        self.cond_drop_prob = cond_drop_prob
+
+        # transformer
+        self.transformer = transformer
+        dim = transformer.dim
+        self.dim = dim
+
+        # conditional flow related
+        self.sigma = sigma
+
+        # sampling related
+        self.odeint_kwargs = odeint_kwargs
+
+        # vocab map for tokenization
+        self.vocab_char_map = vocab_char_map
+
+    @property
+    def device(self):
+        return next(self.parameters()).device
+
+    @torch.no_grad()
+    def sample(
+        self,
+        cond: float["b n d"] | float["b nw"],  # noqa: F722
+        text: int["b nt"] | list[str],  # noqa: F722
+        duration: int | int["b"],  # noqa: F821
+        *,
+        lens: int["b"] | None = None,  # noqa: F821
+        steps=32,
+        cfg_strength=1.0,
+        sway_sampling_coef=None,
+        seed: int | None = None,
+        max_duration=4096,
+        vocoder: Callable[[float["b d n"]], float["b nw"]] | None = None,  # noqa: F722
+        no_ref_audio=False,
+        duplicate_test=False,
+        t_inter=0.1,
+        edit_mask=None,
+    ):
+        self.eval()
+        # raw wave
+
+        if cond.ndim == 2:
+            cond = self.mel_spec(cond)
+            cond = cond.permute(0, 2, 1)
+            assert cond.shape[-1] == self.num_channels
+
+        cond = cond.to(next(self.parameters()).dtype)
+
+        batch, cond_seq_len, device = *cond.shape[:2], cond.device
+        if not exists(lens):
+            lens = torch.full((batch,), cond_seq_len, device=device, dtype=torch.long)
+
+        # text
+
+        if isinstance(text, list):
+            if exists(self.vocab_char_map):
+                text = list_str_to_idx(text, self.vocab_char_map).to(device)
+            else:
+                text = list_str_to_tensor(text).to(device)
+            assert text.shape[0] == batch
+
+        if exists(text):
+            text_lens = (text != -1).sum(dim=-1)
+            lens = torch.maximum(text_lens, lens)  # make sure lengths are at least those of the text characters
+
+        # duration
+
+        cond_mask = lens_to_mask(lens)
+        if edit_mask is not None:
+            cond_mask = cond_mask & edit_mask
+
+        if isinstance(duration, int):
+            duration = torch.full((batch,), duration, device=device, dtype=torch.long)
+
+        duration = torch.maximum(lens + 1, duration)  # just add one token so something is generated
+        duration = duration.clamp(max=max_duration)
+        max_duration = duration.amax()
+
+        # duplicate test corner for inner time step oberservation
+        if duplicate_test:
+            test_cond = F.pad(cond, (0, 0, cond_seq_len, max_duration - 2 * cond_seq_len), value=0.0)
+
+        cond = F.pad(cond, (0, 0, 0, max_duration - cond_seq_len), value=0.0)
+        cond_mask = F.pad(cond_mask, (0, max_duration - cond_mask.shape[-1]), value=False)
+        cond_mask = cond_mask.unsqueeze(-1)
+        step_cond = torch.where(
+            cond_mask, cond, torch.zeros_like(cond)
+        )  # allow direct control (cut cond audio) with lens passed in
+
+        if batch > 1:
+            mask = lens_to_mask(duration)
+        else:  # save memory and speed up, as single inference need no mask currently
+            mask = None
+
+        # test for no ref audio
+        if no_ref_audio:
+            cond = torch.zeros_like(cond)
+
+        # neural ode
+
+        def fn(t, x):
+            # at each step, conditioning is fixed
+            # step_cond = torch.where(cond_mask, cond, torch.zeros_like(cond))
+
+            # predict flow
+            pred = self.transformer(
+                x=x, cond=step_cond, text=text, time=t, mask=mask, drop_audio_cond=False, drop_text=False
+            )
+            if cfg_strength < 1e-5:
+                return pred
+
+            null_pred = self.transformer(
+                x=x, cond=step_cond, text=text, time=t, mask=mask, drop_audio_cond=True, drop_text=True
+            )
+            return pred + (pred - null_pred) * cfg_strength
+
+        # noise input
+        # to make sure batch inference result is same with different batch size, and for sure single inference
+        # still some difference maybe due to convolutional layers
+        y0 = []
+        for dur in duration:
+            if exists(seed):
+                torch.manual_seed(seed)
+            y0.append(torch.randn(dur, self.num_channels, device=self.device, dtype=step_cond.dtype))
+        y0 = pad_sequence(y0, padding_value=0, batch_first=True)
+
+        t_start = 0
+
+        # duplicate test corner for inner time step oberservation
+        if duplicate_test:
+            t_start = t_inter
+            y0 = (1 - t_start) * y0 + t_start * test_cond
+            steps = int(steps * (1 - t_start))
+
+        t = torch.linspace(t_start, 1, steps + 1, device=self.device, dtype=step_cond.dtype)
+        if sway_sampling_coef is not None:
+            t = t + sway_sampling_coef * (torch.cos(torch.pi / 2 * t) - 1 + t)
+
+        trajectory = odeint(fn, y0, t, **self.odeint_kwargs)
+
+        sampled = trajectory[-1]
+        out = sampled
+        out = torch.where(cond_mask, cond, out)
+
+        if exists(vocoder):
+            out = out.permute(0, 2, 1)
+            out = vocoder(out)
+
+        return out, trajectory
+
+    def forward(
+        self,
+        inp: float["b n d"] | float["b nw"],  # mel or raw wave  # noqa: F722
+        text: int["b nt"] | list[str],  # noqa: F722
+        *,
+        lens: int["b"] | None = None,  # noqa: F821
+        noise_scheduler: str | None = None,
+    ):
+        # handle raw wave
+        if inp.ndim == 2:
+            inp = self.mel_spec(inp)
+            inp = inp.permute(0, 2, 1)
+            assert inp.shape[-1] == self.num_channels
+
+        batch, seq_len, dtype, device, _σ1 = *inp.shape[:2], inp.dtype, self.device, self.sigma
+
+        # handle text as string
+        if isinstance(text, list):
+            if exists(self.vocab_char_map):
+                text = list_str_to_idx(text, self.vocab_char_map).to(device)
+            else:
+                text = list_str_to_tensor(text).to(device)
+            assert text.shape[0] == batch
+
+        # lens and mask
+        if not exists(lens):
+            lens = torch.full((batch,), seq_len, device=device)
+
+        mask = lens_to_mask(lens, length=seq_len)  # useless here, as collate_fn will pad to max length in batch
+
+        # get a random span to mask out for training conditionally
+        frac_lengths = torch.zeros((batch,), device=self.device).float().uniform_(*self.frac_lengths_mask)
+        rand_span_mask = mask_from_frac_lengths(lens, frac_lengths)
+
+        if exists(mask):
+            rand_span_mask &= mask
+
+        # mel is x1
+        x1 = inp
+
+        # x0 is gaussian noise
+        x0 = torch.randn_like(x1)
+
+        # time step
+        time = torch.rand((batch,), dtype=dtype, device=self.device)
+        # TODO. noise_scheduler
+
+        # sample xt (φ_t(x) in the paper)
+        t = time.unsqueeze(-1).unsqueeze(-1)
+        φ = (1 - t) * x0 + t * x1
+        flow = x1 - x0
+
+        # only predict what is within the random mask span for infilling
+        cond = torch.where(rand_span_mask[..., None], torch.zeros_like(x1), x1)
+
+        # transformer and cfg training with a drop rate
+        drop_audio_cond = random() < self.audio_drop_prob  # p_drop in voicebox paper
+        if random() < self.cond_drop_prob:  # p_uncond in voicebox paper
+            drop_audio_cond = True
+            drop_text = True
+        else:
+            drop_text = False
+
+        # if want rigourously mask out padding, record in collate_fn in dataset.py, and pass in here
+        # adding mask will use more memory, thus also need to adjust batchsampler with scaled down threshold for long sequences
+        pred = self.transformer(
+            x=φ, cond=cond, text=text, time=time, drop_audio_cond=drop_audio_cond, drop_text=drop_text
+        )
+
+        # flow matching loss
+        loss = F.mse_loss(pred, flow, reduction="none")
+        loss = loss[rand_span_mask]
+
+        return loss.mean(), cond, pred

xinference/thirdparty/f5_tts/model/dataset.py

@@ -0,0 +1,319 @@
+import json
+import random
+from importlib.resources import files
+
+import torch
+import torch.nn.functional as F
+import torchaudio
+from datasets import Dataset as Dataset_
+from datasets import load_from_disk
+from torch import nn
+from torch.utils.data import Dataset, Sampler
+from tqdm import tqdm
+
+from f5_tts.model.modules import MelSpec
+from f5_tts.model.utils import default
+
+
+class HFDataset(Dataset):
+    def __init__(
+        self,
+        hf_dataset: Dataset,
+        target_sample_rate=24_000,
+        n_mel_channels=100,
+        hop_length=256,
+        n_fft=1024,
+        win_length=1024,
+        mel_spec_type="vocos",
+    ):
+        self.data = hf_dataset
+        self.target_sample_rate = target_sample_rate
+        self.hop_length = hop_length
+
+        self.mel_spectrogram = MelSpec(
+            n_fft=n_fft,
+            hop_length=hop_length,
+            win_length=win_length,
+            n_mel_channels=n_mel_channels,
+            target_sample_rate=target_sample_rate,
+            mel_spec_type=mel_spec_type,
+        )
+
+    def get_frame_len(self, index):
+        row = self.data[index]
+        audio = row["audio"]["array"]
+        sample_rate = row["audio"]["sampling_rate"]
+        return audio.shape[-1] / sample_rate * self.target_sample_rate / self.hop_length
+
+    def __len__(self):
+        return len(self.data)
+
+    def __getitem__(self, index):
+        row = self.data[index]
+        audio = row["audio"]["array"]
+
+        # logger.info(f"Audio shape: {audio.shape}")
+
+        sample_rate = row["audio"]["sampling_rate"]
+        duration = audio.shape[-1] / sample_rate
+
+        if duration > 30 or duration < 0.3:
+            return self.__getitem__((index + 1) % len(self.data))
+
+        audio_tensor = torch.from_numpy(audio).float()
+
+        if sample_rate != self.target_sample_rate:
+            resampler = torchaudio.transforms.Resample(sample_rate, self.target_sample_rate)
+            audio_tensor = resampler(audio_tensor)
+
+        audio_tensor = audio_tensor.unsqueeze(0)  # 't -> 1 t')
+
+        mel_spec = self.mel_spectrogram(audio_tensor)
+
+        mel_spec = mel_spec.squeeze(0)  # '1 d t -> d t'
+
+        text = row["text"]
+
+        return dict(
+            mel_spec=mel_spec,
+            text=text,
+        )
+
+
+class CustomDataset(Dataset):
+    def __init__(
+        self,
+        custom_dataset: Dataset,
+        durations=None,
+        target_sample_rate=24_000,
+        hop_length=256,
+        n_mel_channels=100,
+        n_fft=1024,
+        win_length=1024,
+        mel_spec_type="vocos",
+        preprocessed_mel=False,
+        mel_spec_module: nn.Module | None = None,
+    ):
+        self.data = custom_dataset
+        self.durations = durations
+        self.target_sample_rate = target_sample_rate
+        self.hop_length = hop_length
+        self.n_fft = n_fft
+        self.win_length = win_length
+        self.mel_spec_type = mel_spec_type
+        self.preprocessed_mel = preprocessed_mel
+
+        if not preprocessed_mel:
+            self.mel_spectrogram = default(
+                mel_spec_module,
+                MelSpec(
+                    n_fft=n_fft,
+                    hop_length=hop_length,
+                    win_length=win_length,
+                    n_mel_channels=n_mel_channels,
+                    target_sample_rate=target_sample_rate,
+                    mel_spec_type=mel_spec_type,
+                ),
+            )
+
+    def get_frame_len(self, index):
+        if (
+            self.durations is not None
+        ):  # Please make sure the separately provided durations are correct, otherwise 99.99% OOM
+            return self.durations[index] * self.target_sample_rate / self.hop_length
+        return self.data[index]["duration"] * self.target_sample_rate / self.hop_length
+
+    def __len__(self):
+        return len(self.data)
+
+    def __getitem__(self, index):
+        while True:
+            row = self.data[index]
+            audio_path = row["audio_path"]
+            text = row["text"]
+            duration = row["duration"]
+
+            # filter by given length
+            if 0.3 <= duration <= 30:
+                break  # valid
+
+            index = (index + 1) % len(self.data)
+
+        if self.preprocessed_mel:
+            mel_spec = torch.tensor(row["mel_spec"])
+        else:
+            audio, source_sample_rate = torchaudio.load(audio_path)
+
+            # make sure mono input
+            if audio.shape[0] > 1:
+                audio = torch.mean(audio, dim=0, keepdim=True)
+
+            # resample if necessary
+            if source_sample_rate != self.target_sample_rate:
+                resampler = torchaudio.transforms.Resample(source_sample_rate, self.target_sample_rate)
+                audio = resampler(audio)
+
+            # to mel spectrogram
+            mel_spec = self.mel_spectrogram(audio)
+            mel_spec = mel_spec.squeeze(0)  # '1 d t -> d t'
+
+        return {
+            "mel_spec": mel_spec,
+            "text": text,
+        }
+
+
+# Dynamic Batch Sampler
+class DynamicBatchSampler(Sampler[list[int]]):
+    """Extension of Sampler that will do the following:
+    1.  Change the batch size (essentially number of sequences)
+        in a batch to ensure that the total number of frames are less
+        than a certain threshold.
+    2.  Make sure the padding efficiency in the batch is high.
+    """
+
+    def __init__(
+        self, sampler: Sampler[int], frames_threshold: int, max_samples=0, random_seed=None, drop_last: bool = False
+    ):
+        self.sampler = sampler
+        self.frames_threshold = frames_threshold
+        self.max_samples = max_samples
+
+        indices, batches = [], []
+        data_source = self.sampler.data_source
+
+        for idx in tqdm(
+            self.sampler, desc="Sorting with sampler... if slow, check whether dataset is provided with duration"
+        ):
+            indices.append((idx, data_source.get_frame_len(idx)))
+        indices.sort(key=lambda elem: elem[1])
+
+        batch = []
+        batch_frames = 0
+        for idx, frame_len in tqdm(
+            indices, desc=f"Creating dynamic batches with {frames_threshold} audio frames per gpu"
+        ):
+            if batch_frames + frame_len <= self.frames_threshold and (max_samples == 0 or len(batch) < max_samples):
+                batch.append(idx)
+                batch_frames += frame_len
+            else:
+                if len(batch) > 0:
+                    batches.append(batch)
+                if frame_len <= self.frames_threshold:
+                    batch = [idx]
+                    batch_frames = frame_len
+                else:
+                    batch = []
+                    batch_frames = 0
+
+        if not drop_last and len(batch) > 0:
+            batches.append(batch)
+
+        del indices
+
+        # if want to have different batches between epochs, may just set a seed and log it in ckpt
+        # cuz during multi-gpu training, although the batch on per gpu not change between epochs, the formed general minibatch is different
+        # e.g. for epoch n, use (random_seed + n)
+        random.seed(random_seed)
+        random.shuffle(batches)
+
+        self.batches = batches
+
+    def __iter__(self):
+        return iter(self.batches)
+
+    def __len__(self):
+        return len(self.batches)
+
+
+# Load dataset
+
+
+def load_dataset(
+    dataset_name: str,
+    tokenizer: str = "pinyin",
+    dataset_type: str = "CustomDataset",
+    audio_type: str = "raw",
+    mel_spec_module: nn.Module | None = None,
+    mel_spec_kwargs: dict = dict(),
+) -> CustomDataset | HFDataset:
+    """
+    dataset_type    - "CustomDataset" if you want to use tokenizer name and default data path to load for train_dataset
+                    - "CustomDatasetPath" if you just want to pass the full path to a preprocessed dataset without relying on tokenizer
+    """
+
+    print("Loading dataset ...")
+
+    if dataset_type == "CustomDataset":
+        rel_data_path = str(files("f5_tts").joinpath(f"../../data/{dataset_name}_{tokenizer}"))
+        if audio_type == "raw":
+            try:
+                train_dataset = load_from_disk(f"{rel_data_path}/raw")
+            except:  # noqa: E722
+                train_dataset = Dataset_.from_file(f"{rel_data_path}/raw.arrow")
+            preprocessed_mel = False
+        elif audio_type == "mel":
+            train_dataset = Dataset_.from_file(f"{rel_data_path}/mel.arrow")
+            preprocessed_mel = True
+        with open(f"{rel_data_path}/duration.json", "r", encoding="utf-8") as f:
+            data_dict = json.load(f)
+        durations = data_dict["duration"]
+        train_dataset = CustomDataset(
+            train_dataset,
+            durations=durations,
+            preprocessed_mel=preprocessed_mel,
+            mel_spec_module=mel_spec_module,
+            **mel_spec_kwargs,
+        )
+
+    elif dataset_type == "CustomDatasetPath":
+        try:
+            train_dataset = load_from_disk(f"{dataset_name}/raw")
+        except:  # noqa: E722
+            train_dataset = Dataset_.from_file(f"{dataset_name}/raw.arrow")
+
+        with open(f"{dataset_name}/duration.json", "r", encoding="utf-8") as f:
+            data_dict = json.load(f)
+        durations = data_dict["duration"]
+        train_dataset = CustomDataset(
+            train_dataset, durations=durations, preprocessed_mel=preprocessed_mel, **mel_spec_kwargs
+        )
+
+    elif dataset_type == "HFDataset":
+        print(
+            "Should manually modify the path of huggingface dataset to your need.\n"
+            + "May also the corresponding script cuz different dataset may have different format."
+        )
+        pre, post = dataset_name.split("_")
+        train_dataset = HFDataset(
+            load_dataset(f"{pre}/{pre}", split=f"train.{post}", cache_dir=str(files("f5_tts").joinpath("../../data"))),
+        )
+
+    return train_dataset
+
+
+# collation
+
+
+def collate_fn(batch):
+    mel_specs = [item["mel_spec"].squeeze(0) for item in batch]
+    mel_lengths = torch.LongTensor([spec.shape[-1] for spec in mel_specs])
+    max_mel_length = mel_lengths.amax()
+
+    padded_mel_specs = []
+    for spec in mel_specs:  # TODO. maybe records mask for attention here
+        padding = (0, max_mel_length - spec.size(-1))
+        padded_spec = F.pad(spec, padding, value=0)
+        padded_mel_specs.append(padded_spec)
+
+    mel_specs = torch.stack(padded_mel_specs)
+
+    text = [item["text"] for item in batch]
+    text_lengths = torch.LongTensor([len(item) for item in text])
+
+    return dict(
+        mel=mel_specs,
+        mel_lengths=mel_lengths,
+        text=text,
+        text_lengths=text_lengths,
+    )