model/model_utils.py

from __future__ import annotations

import os
import re
import math
import random
import string
from tqdm import tqdm
from collections import defaultdict

import matplotlib
matplotlib.use("Agg")
import matplotlib.pylab as plt

import torch
import torch.nn.functional as F
from torch.nn.utils.rnn import pad_sequence
import torchaudio

import einx
from einops import rearrange, reduce

import jieba
from pypinyin import lazy_pinyin, Style

from model.ecapa_tdnn import ECAPA_TDNN_SMALL
from model.modules import MelSpec


# seed everything

def seed_everything(seed = 0):
    random.seed(seed)
    os.environ['PYTHONHASHSEED'] = str(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False

# helpers

def exists(v):
    return v is not None

def default(v, d):
    return v if exists(v) else d

# tensor helpers

def lens_to_mask(
    t: int['b'],
    length: int | None = None
) -> bool['b n']:

    if not exists(length):
        length = t.amax()

    seq = torch.arange(length, device = t.device)
    return einx.less('n, b -> b n', seq, t)

def mask_from_start_end_indices(
    seq_len: int['b'],
    start: int['b'],
    end: int['b']
):
    max_seq_len = seq_len.max().item()  
    seq = torch.arange(max_seq_len, device = start.device).long()
    return einx.greater_equal('n, b -> b n', seq, start) & einx.less('n, b -> b n', seq, end)

def mask_from_frac_lengths(
    seq_len: int['b'],
    frac_lengths: float['b']
):
    lengths = (frac_lengths * seq_len).long()
    max_start = seq_len - lengths

    rand = torch.rand_like(frac_lengths)
    start = (max_start * rand).long().clamp(min = 0)
    end = start + lengths

    return mask_from_start_end_indices(seq_len, start, end)

def maybe_masked_mean(
    t: float['b n d'],
    mask: bool['b n'] = None
) -> float['b d']:

    if not exists(mask):
        return t.mean(dim = 1)

    t = einx.where('b n, b n d, -> b n d', mask, t, 0.)
    num = reduce(t, 'b n d -> b d', 'sum')
    den = reduce(mask.float(), 'b n -> b', 'sum')

    return einx.divide('b d, b -> b d', num, den.clamp(min = 1.))


# simple utf-8 tokenizer, since paper went character based
def list_str_to_tensor(
    text: list[str],
    padding_value = -1
) -> int['b nt']:
    list_tensors = [torch.tensor([*bytes(t, 'UTF-8')]) for t in text]  # ByT5 style
    text = pad_sequence(list_tensors, padding_value = padding_value, batch_first = True)
    return text

# char tokenizer, based on custom dataset's extracted .txt file
def list_str_to_idx(
    text: list[str] | list[list[str]],
    vocab_char_map: dict[str, int],  # {char: idx}
    padding_value = -1
) -> int['b nt']:
    list_idx_tensors = [torch.tensor([vocab_char_map.get(c, 0) for c in t]) for t in text]  # pinyin or char style
    text = pad_sequence(list_idx_tensors, padding_value = padding_value, batch_first = True)
    return text


# Get tokenizer

def get_tokenizer(dataset_name, tokenizer: str = "pinyin"):
    ''' 
    tokenizer   - "pinyin" do g2p for only chinese characters, need .txt vocab_file
                - "char" for char-wise tokenizer, need .txt vocab_file
                - "byte" for utf-8 tokenizer
                - "custom" if you're directly passing in a path to the vocab.txt you want to use
    vocab_size  - if use "pinyin", all available pinyin types, common alphabets (also those with accent) and symbols
                - if use "char", derived from unfiltered character & symbol counts of custom dataset
                - if use "byte", set to 256 (unicode byte range) 
    ''' 
    if tokenizer in ["pinyin", "char"]:
        with open (f"data/{dataset_name}_{tokenizer}/vocab.txt", "r", encoding="utf-8") as f:
            vocab_char_map = {}
            for i, char in enumerate(f):
                vocab_char_map[char[:-1]] = i
        vocab_size = len(vocab_char_map)
        assert vocab_char_map[" "] == 0, "make sure space is of idx 0 in vocab.txt, cuz 0 is used for unknown char"

    elif tokenizer == "byte":
        vocab_char_map = None
        vocab_size = 256
    elif tokenizer == "custom":
        with open (dataset_name, "r", encoding="utf-8") as f:
            vocab_char_map = {}
            for i, char in enumerate(f):
                vocab_char_map[char[:-1]] = i
        vocab_size = len(vocab_char_map)

    return vocab_char_map, vocab_size


# convert char to pinyin

def convert_char_to_pinyin(text_list, polyphone = True):
    final_text_list = []
    god_knows_why_en_testset_contains_zh_quote = str.maketrans({'“': '"', '”': '"', '‘': "'", '’': "'"})  # in case librispeech (orig no-pc) test-clean
    custom_trans = str.maketrans({';': ','})  # add custom trans here, to address oov
    for text in text_list:
        char_list = []
        text = text.translate(god_knows_why_en_testset_contains_zh_quote)
        text = text.translate(custom_trans)
        for seg in jieba.cut(text):
            seg_byte_len = len(bytes(seg, 'UTF-8'))
            if seg_byte_len == len(seg):  # if pure alphabets and symbols
                if char_list and seg_byte_len > 1 and char_list[-1] not in " :'\"":
                    char_list.append(" ")
                char_list.extend(seg)
            elif polyphone and seg_byte_len == 3 * len(seg):  # if pure chinese characters
                seg = lazy_pinyin(seg, style=Style.TONE3, tone_sandhi=True)
                for c in seg:
                    if c not in "。，、；：？！《》【】—…":
                        char_list.append(" ")
                    char_list.append(c)
            else:  # if mixed chinese characters, alphabets and symbols
                for c in seg:
                    if ord(c) < 256:
                        char_list.extend(c)
                    else:
                        if c not in "。，、；：？！《》【】—…":
                            char_list.append(" ")
                            char_list.extend(lazy_pinyin(c, style=Style.TONE3, tone_sandhi=True))
                        else:  # if is zh punc
                            char_list.append(c)
        final_text_list.append(char_list)

    return final_text_list


# save spectrogram
def save_spectrogram(spectrogram, path):
    plt.figure(figsize=(12, 4))
    plt.imshow(spectrogram, origin='lower', aspect='auto')
    plt.colorbar()
    plt.savefig(path)
    plt.close()


# seedtts testset metainfo: utt, prompt_text, prompt_wav, gt_text, gt_wav
def get_seedtts_testset_metainfo(metalst):
    f = open(metalst); lines = f.readlines(); f.close()
    metainfo = []
    for line in lines:
        if len(line.strip().split('|')) == 5:
            utt, prompt_text, prompt_wav, gt_text, gt_wav = line.strip().split('|')
        elif len(line.strip().split('|')) == 4:
            utt, prompt_text, prompt_wav, gt_text = line.strip().split('|')
            gt_wav = os.path.join(os.path.dirname(metalst), "wavs", utt + ".wav")
        if not os.path.isabs(prompt_wav):
            prompt_wav = os.path.join(os.path.dirname(metalst), prompt_wav)
        metainfo.append((utt, prompt_text, prompt_wav, gt_text, gt_wav))
    return metainfo


# librispeech test-clean metainfo: gen_utt, ref_txt, ref_wav, gen_txt, gen_wav
def get_librispeech_test_clean_metainfo(metalst, librispeech_test_clean_path):
    f = open(metalst); lines = f.readlines(); f.close()
    metainfo = []
    for line in lines:
        ref_utt, ref_dur, ref_txt, gen_utt, gen_dur, gen_txt = line.strip().split('\t')

        # ref_txt = ref_txt[0] + ref_txt[1:].lower() + '.'  # if use librispeech test-clean (no-pc)
        ref_spk_id, ref_chaptr_id, _ =  ref_utt.split('-')
        ref_wav = os.path.join(librispeech_test_clean_path, ref_spk_id, ref_chaptr_id, ref_utt + '.flac')

        # gen_txt = gen_txt[0] + gen_txt[1:].lower() + '.'  # if use librispeech test-clean (no-pc)
        gen_spk_id, gen_chaptr_id, _ =  gen_utt.split('-')
        gen_wav = os.path.join(librispeech_test_clean_path, gen_spk_id, gen_chaptr_id, gen_utt + '.flac')

        metainfo.append((gen_utt, ref_txt, ref_wav, " " + gen_txt, gen_wav))

    return metainfo


# padded to max length mel batch
def padded_mel_batch(ref_mels):
    max_mel_length = torch.LongTensor([mel.shape[-1] for mel in ref_mels]).amax()
    padded_ref_mels = []
    for mel in ref_mels:
        padded_ref_mel = F.pad(mel, (0, max_mel_length - mel.shape[-1]), value = 0)
        padded_ref_mels.append(padded_ref_mel)
    padded_ref_mels = torch.stack(padded_ref_mels)
    padded_ref_mels = rearrange(padded_ref_mels, 'b d n -> b n d')
    return padded_ref_mels


# get prompts from metainfo containing: utt, prompt_text, prompt_wav, gt_text, gt_wav

def get_inference_prompt(
    metainfo, 
    speed = 1., tokenizer = "pinyin", polyphone = True, 
    target_sample_rate = 24000, n_mel_channels = 100, hop_length = 256, target_rms = 0.1,
    use_truth_duration = False,
    infer_batch_size = 1, num_buckets = 200, min_secs = 3, max_secs = 40,
):
    prompts_all = []

    min_tokens = min_secs * target_sample_rate // hop_length
    max_tokens = max_secs * target_sample_rate // hop_length

    batch_accum = [0] * num_buckets
    utts, ref_rms_list, ref_mels, ref_mel_lens, total_mel_lens, final_text_list = \
        ([[] for _ in range(num_buckets)] for _ in range(6))

    mel_spectrogram = MelSpec(target_sample_rate=target_sample_rate, n_mel_channels=n_mel_channels, hop_length=hop_length)

    for utt, prompt_text, prompt_wav, gt_text, gt_wav in tqdm(metainfo, desc="Processing prompts..."):

        # Audio
        ref_audio, ref_sr = torchaudio.load(prompt_wav)
        ref_rms = torch.sqrt(torch.mean(torch.square(ref_audio)))
        if ref_rms < target_rms:
            ref_audio = ref_audio * target_rms / ref_rms
        assert ref_audio.shape[-1] > 5000, f"Empty prompt wav: {prompt_wav}, or torchaudio backend issue."
        if ref_sr != target_sample_rate:
            resampler = torchaudio.transforms.Resample(ref_sr, target_sample_rate)
            ref_audio = resampler(ref_audio)

        # Text
        if len(prompt_text[-1].encode('utf-8')) == 1:
            prompt_text = prompt_text + " "
        text = [prompt_text + gt_text]
        if tokenizer == "pinyin":
            text_list = convert_char_to_pinyin(text, polyphone = polyphone)
        else:
            text_list = text

        # Duration, mel frame length
        ref_mel_len = ref_audio.shape[-1] // hop_length
        if use_truth_duration:
            gt_audio, gt_sr = torchaudio.load(gt_wav)
            if gt_sr != target_sample_rate:
                resampler = torchaudio.transforms.Resample(gt_sr, target_sample_rate)
                gt_audio = resampler(gt_audio)
            total_mel_len = ref_mel_len + int(gt_audio.shape[-1] / hop_length / speed)

            # # test vocoder resynthesis
            # ref_audio = gt_audio
        else:
            zh_pause_punc = r"。，、；：？！"
            ref_text_len = len(prompt_text.encode('utf-8')) + 3 * len(re.findall(zh_pause_punc, prompt_text))
            gen_text_len = len(gt_text.encode('utf-8')) + 3 * len(re.findall(zh_pause_punc, gt_text))
            total_mel_len = ref_mel_len + int(ref_mel_len / ref_text_len * gen_text_len / speed)

        # to mel spectrogram
        ref_mel = mel_spectrogram(ref_audio)
        ref_mel = rearrange(ref_mel, '1 d n -> d n')

        # deal with batch
        assert infer_batch_size > 0, "infer_batch_size should be greater than 0."
        assert min_tokens <= total_mel_len <= max_tokens, \
            f"Audio {utt} has duration {total_mel_len*hop_length//target_sample_rate}s out of range [{min_secs}, {max_secs}]."
        bucket_i = math.floor((total_mel_len - min_tokens) / (max_tokens - min_tokens + 1) * num_buckets)

        utts[bucket_i].append(utt)
        ref_rms_list[bucket_i].append(ref_rms)
        ref_mels[bucket_i].append(ref_mel)
        ref_mel_lens[bucket_i].append(ref_mel_len)
        total_mel_lens[bucket_i].append(total_mel_len)
        final_text_list[bucket_i].extend(text_list)

        batch_accum[bucket_i] += total_mel_len

        if batch_accum[bucket_i] >= infer_batch_size:
            # print(f"\n{len(ref_mels[bucket_i][0][0])}\n{ref_mel_lens[bucket_i]}\n{total_mel_lens[bucket_i]}")
            prompts_all.append((
                utts[bucket_i], 
                ref_rms_list[bucket_i], 
                padded_mel_batch(ref_mels[bucket_i]), 
                ref_mel_lens[bucket_i], 
                total_mel_lens[bucket_i], 
                final_text_list[bucket_i]
            ))
            batch_accum[bucket_i] = 0
            utts[bucket_i], ref_rms_list[bucket_i], ref_mels[bucket_i], ref_mel_lens[bucket_i], total_mel_lens[bucket_i], final_text_list[bucket_i] = [], [], [], [], [], []

    # add residual
    for bucket_i, bucket_frames in enumerate(batch_accum):
        if bucket_frames > 0:
            prompts_all.append((
                utts[bucket_i], 
                ref_rms_list[bucket_i], 
                padded_mel_batch(ref_mels[bucket_i]), 
                ref_mel_lens[bucket_i], 
                total_mel_lens[bucket_i], 
                final_text_list[bucket_i]
            ))
    # not only leave easy work for last workers
    random.seed(666)
    random.shuffle(prompts_all)

    return prompts_all


# get wav_res_ref_text of seed-tts test metalst
# https://github.com/BytedanceSpeech/seed-tts-eval

def get_seed_tts_test(metalst, gen_wav_dir, gpus):
    f = open(metalst)
    lines = f.readlines()
    f.close()

    test_set_ = []
    for line in tqdm(lines):
        if len(line.strip().split('|')) == 5:
            utt, prompt_text, prompt_wav, gt_text, gt_wav = line.strip().split('|')
        elif len(line.strip().split('|')) == 4:
            utt, prompt_text, prompt_wav, gt_text = line.strip().split('|')

        if not os.path.exists(os.path.join(gen_wav_dir, utt + '.wav')):
            continue
        gen_wav = os.path.join(gen_wav_dir, utt + '.wav')
        if not os.path.isabs(prompt_wav):
            prompt_wav = os.path.join(os.path.dirname(metalst), prompt_wav)

        test_set_.append((gen_wav, prompt_wav, gt_text))

    num_jobs = len(gpus)
    if num_jobs == 1:
        return [(gpus[0], test_set_)]
    
    wav_per_job = len(test_set_) // num_jobs + 1
    test_set = []
    for i in range(num_jobs):
        test_set.append((gpus[i], test_set_[i*wav_per_job:(i+1)*wav_per_job]))

    return test_set


# get librispeech test-clean cross sentence test

def get_librispeech_test(metalst, gen_wav_dir, gpus, librispeech_test_clean_path, eval_ground_truth = False):
    f = open(metalst)
    lines = f.readlines()
    f.close()

    test_set_ = []
    for line in tqdm(lines):
        ref_utt, ref_dur, ref_txt, gen_utt, gen_dur, gen_txt = line.strip().split('\t')

        if eval_ground_truth:
            gen_spk_id, gen_chaptr_id, _ =  gen_utt.split('-')
            gen_wav = os.path.join(librispeech_test_clean_path, gen_spk_id, gen_chaptr_id, gen_utt + '.flac')
        else:
            if not os.path.exists(os.path.join(gen_wav_dir, gen_utt + '.wav')):
                raise FileNotFoundError(f"Generated wav not found: {gen_utt}")
            gen_wav = os.path.join(gen_wav_dir, gen_utt + '.wav')

        ref_spk_id, ref_chaptr_id, _ =  ref_utt.split('-')
        ref_wav = os.path.join(librispeech_test_clean_path, ref_spk_id, ref_chaptr_id, ref_utt + '.flac')

        test_set_.append((gen_wav, ref_wav, gen_txt))

    num_jobs = len(gpus)
    if num_jobs == 1:
        return [(gpus[0], test_set_)]
    
    wav_per_job = len(test_set_) // num_jobs + 1
    test_set = []
    for i in range(num_jobs):
        test_set.append((gpus[i], test_set_[i*wav_per_job:(i+1)*wav_per_job]))

    return test_set


# load asr model

def load_asr_model(lang, ckpt_dir = ""):
    if lang == "zh":
        from funasr import AutoModel
        model = AutoModel(
            model = os.path.join(ckpt_dir, "paraformer-zh"), 
            # vad_model = os.path.join(ckpt_dir, "fsmn-vad"), 
            # punc_model = os.path.join(ckpt_dir, "ct-punc"),
            # spk_model = os.path.join(ckpt_dir, "cam++"), 
            disable_update=True,
            )  # following seed-tts setting
    elif lang == "en":
        from faster_whisper import WhisperModel
        model_size = "large-v3" if ckpt_dir == "" else ckpt_dir
        model = WhisperModel(model_size, device="cuda", compute_type="float16")
    return model


# WER Evaluation, the way Seed-TTS does

def run_asr_wer(args):
    rank, lang, test_set, ckpt_dir = args

    if lang == "zh":
        import zhconv
        torch.cuda.set_device(rank)
    elif lang == "en":
        os.environ["CUDA_VISIBLE_DEVICES"] = str(rank)
    else:
        raise NotImplementedError("lang support only 'zh' (funasr paraformer-zh), 'en' (faster-whisper-large-v3), for now.")

    asr_model = load_asr_model(lang, ckpt_dir = ckpt_dir)
    
    from zhon.hanzi import punctuation
    punctuation_all = punctuation + string.punctuation
    wers = []

    from jiwer import compute_measures
    for gen_wav, prompt_wav, truth in tqdm(test_set):
        if lang == "zh":
            res = asr_model.generate(input=gen_wav, batch_size_s=300, disable_pbar=True)
            hypo = res[0]["text"]
            hypo = zhconv.convert(hypo, 'zh-cn')
        elif lang == "en":
            segments, _ = asr_model.transcribe(gen_wav, beam_size=5, language="en")
            hypo = ''
            for segment in segments:
                hypo = hypo + ' ' + segment.text

        # raw_truth = truth
        # raw_hypo = hypo

        for x in punctuation_all:
            truth = truth.replace(x, '')
            hypo = hypo.replace(x, '')

        truth = truth.replace('  ', ' ')
        hypo = hypo.replace('  ', ' ')

        if lang == "zh":
            truth = " ".join([x for x in truth])
            hypo = " ".join([x for x in hypo])
        elif lang == "en":
            truth = truth.lower()
            hypo = hypo.lower()

        measures = compute_measures(truth, hypo)
        wer = measures["wer"]

        # ref_list = truth.split(" ")
        # subs = measures["substitutions"] / len(ref_list)
        # dele = measures["deletions"] / len(ref_list)
        # inse = measures["insertions"] / len(ref_list)

        wers.append(wer)

    return wers


# SIM Evaluation

def run_sim(args):
    rank, test_set, ckpt_dir = args
    device = f"cuda:{rank}"

    model = ECAPA_TDNN_SMALL(feat_dim=1024, feat_type='wavlm_large', config_path=None)
    state_dict = torch.load(ckpt_dir, weights_only=True, map_location=lambda storage, loc: storage)
    model.load_state_dict(state_dict['model'], strict=False)

    use_gpu=True if torch.cuda.is_available() else False
    if use_gpu:
        model = model.cuda(device)
    model.eval()

    sim_list = []
    for wav1, wav2, truth in tqdm(test_set):

        wav1, sr1 = torchaudio.load(wav1)
        wav2, sr2 = torchaudio.load(wav2)

        resample1 = torchaudio.transforms.Resample(orig_freq=sr1, new_freq=16000)
        resample2 = torchaudio.transforms.Resample(orig_freq=sr2, new_freq=16000)
        wav1 = resample1(wav1)
        wav2 = resample2(wav2)

        if use_gpu:
            wav1 = wav1.cuda(device)
            wav2 = wav2.cuda(device)
        with torch.no_grad():
            emb1 = model(wav1)
            emb2 = model(wav2)
        
        sim = F.cosine_similarity(emb1, emb2)[0].item()
        # print(f"VSim score between two audios: {sim:.4f} (-1.0, 1.0).")
        sim_list.append(sim)
    
    return sim_list


# filter func for dirty data with many repetitions

def repetition_found(text, length = 2, tolerance = 10):
    pattern_count = defaultdict(int)
    for i in range(len(text) - length + 1):
        pattern = text[i:i + length]
        pattern_count[pattern] += 1
    for pattern, count in pattern_count.items():
        if count > tolerance:
            return True
    return False


# load model checkpoint for inference

def load_checkpoint(model, ckpt_path, device, use_ema = True):
    from ema_pytorch import EMA

    ckpt_type = ckpt_path.split(".")[-1]
    if ckpt_type == "safetensors":
        from safetensors.torch import load_file
        checkpoint = load_file(ckpt_path, device=device)
    else:
        checkpoint = torch.load(ckpt_path, weights_only=True, map_location=device)

    if use_ema == True:
        ema_model = EMA(model, include_online_model = False).to(device)
        if ckpt_type == "safetensors":
            ema_model.load_state_dict(checkpoint)
        else:
            ema_model.load_state_dict(checkpoint['ema_model_state_dict'])
        ema_model.copy_params_from_ema_to_model()
    else:
        model.load_state_dict(checkpoint['model_state_dict'])
        
    return model