import os import os import random import sys import torch import torch.nn.functional as F import torch.utils.data import torchaudio from tqdm import tqdm from data.audio.paired_voice_audio_dataset import CharacterTokenizer from data.audio.unsupervised_audio_dataset import load_audio, load_similar_clips from models.tacotron2.taco_utils import load_filepaths_and_text from models.tacotron2.text import text_to_sequence, sequence_to_text from utils.util import opt_get def parse_tsv_aligned_codes(line, base_path): fpt = line.strip().split('\t') def convert_string_list_to_tensor(strlist): if strlist.startswith('['): strlist = strlist[1:] if strlist.endswith(']'): strlist = strlist[:-1] as_ints = [int(s) for s in strlist.split(', ')] return torch.tensor(as_ints) return os.path.join(base_path, f'{fpt[1]}'), fpt[0], convert_string_list_to_tensor(fpt[2]) class FastPairedVoiceDataset(torch.utils.data.Dataset): """ This dataset is derived from paired_voice_audio, but it only supports loading from TSV files generated from the ocotillo transcription engine, which includes alignment codes. To support the vastly larger TSV files, this dataset uses an indexing mechanism which randomly selects offsets within the translation file to seek to. The data returned is relative to these offsets. In practice, this means two things: 1) Index {i} of this dataset means nothing: fetching from the same index will almost always return different data. As a result, this dataset should not be used for validation or test runs. 2) This dataset has a slight bias for items with longer text or longer filenames. The upshot is that this dataset loads extremely quickly and consumes almost no system memory. """ def __init__(self, hparams): self.paths = hparams['path'] if not isinstance(self.paths, list): self.paths = [self.paths] self.paths_size_bytes = [os.path.getsize(p) for p in self.paths] self.total_size_bytes = sum(self.paths_size_bytes) self.load_conditioning = opt_get(hparams, ['load_conditioning'], False) self.conditioning_candidates = opt_get(hparams, ['num_conditioning_candidates'], 1) self.conditioning_length = opt_get(hparams, ['conditioning_length'], 44100) self.debug_failures = opt_get(hparams, ['debug_loading_failures'], False) self.aligned_codes_to_audio_ratio = opt_get(hparams, ['aligned_codes_ratio'], 443) self.text_cleaners = hparams.text_cleaners self.sample_rate = hparams.sample_rate self.max_wav_len = opt_get(hparams, ['max_wav_length'], None) if self.max_wav_len is not None: self.max_aligned_codes = self.max_wav_len // self.aligned_codes_to_audio_ratio self.max_text_len = opt_get(hparams, ['max_text_length'], None) assert self.max_wav_len is not None and self.max_text_len is not None self.use_bpe_tokenizer = opt_get(hparams, ['use_bpe_tokenizer'], False) if self.use_bpe_tokenizer: from data.audio.voice_tokenizer import VoiceBpeTokenizer self.tokenizer = VoiceBpeTokenizer(opt_get(hparams, ['tokenizer_vocab'], '../experiments/bpe_lowercase_asr_256.json')) else: self.tokenizer = CharacterTokenizer() self.skipped_items = 0 # records how many items are skipped when accessing an index. def get_wav_text_pair(self, audiopath_and_text): # separate filename and text audiopath, text = audiopath_and_text[0], audiopath_and_text[1] text_seq = self.get_text(text) wav = load_audio(audiopath, self.sample_rate) return (text_seq, wav, text, audiopath_and_text[0]) def get_text(self, text): tokens = self.tokenizer.encode(text) tokens = torch.IntTensor(tokens) if self.use_bpe_tokenizer: # Assert if any UNK,start tokens encountered. assert not torch.any(tokens == 1) # The stop token should always be sacred. assert not torch.any(tokens == 0) return tokens def load_random_line(self, depth=0): assert depth < 10 rand_offset = random.randint(0, self.total_size_bytes) for i in range(len(self.paths)): if rand_offset < self.paths_size_bytes[i]: break else: rand_offset -= self.paths_size_bytes[i] path = self.paths[i] with open(path, 'r', encoding='utf-8') as f: f.seek(rand_offset) # Read the rest of the line we seeked to, then the line after that. try: # This can fail when seeking to a UTF-8 escape byte. f.readline() except: return self.load_random_line(depth=depth + 1) # On failure, just recurse and try again. l2 = f.readline() if l2: try: base_path = os.path.dirname(path) return parse_tsv_aligned_codes(l2, base_path) except: print(f"error parsing random offset: {sys.exc_info()}") return self.load_random_line(depth=depth+1) # On failure, just recurse and try again. def __getitem__(self, index): self.skipped_items += 1 apt = self.load_random_line() try: tseq, wav, text, path = self.get_wav_text_pair(apt) if text is None or len(text.strip()) == 0: raise ValueError cond, cond_is_self = load_similar_clips(apt[0], self.conditioning_length, self.sample_rate, n=self.conditioning_candidates) if self.load_conditioning else (None, False) except: if self.skipped_items > 100: raise # Rethrow if we have nested too far. if self.debug_failures: print(f"error loading {apt[0]} {sys.exc_info()}") return self[(index+1) % len(self)] aligned_codes = apt[2] actually_skipped_items = self.skipped_items self.skipped_items = 0 if wav is None or \ (self.max_wav_len is not None and wav.shape[-1] > self.max_wav_len) or \ (self.max_text_len is not None and tseq.shape[0] > self.max_text_len): # Basically, this audio file is nonexistent or too long to be supported by the dataset. # It's hard to handle this situation properly. Best bet is to return the a random valid token and skew the dataset somewhat as a result. if self.debug_failures: print(f"error loading {path}: ranges are out of bounds; {wav.shape[-1]}, {tseq.shape[0]}") rv = random.randint(0,len(self)-1) return self[rv] orig_output = wav.shape[-1] orig_text_len = tseq.shape[0] orig_aligned_code_length = aligned_codes.shape[0] if wav.shape[-1] != self.max_wav_len: wav = F.pad(wav, (0, self.max_wav_len - wav.shape[-1])) # These codes are aligned to audio inputs, so make sure to pad them as well. aligned_codes = F.pad(aligned_codes, (0, self.max_aligned_codes-aligned_codes.shape[0])) if tseq.shape[0] != self.max_text_len: tseq = F.pad(tseq, (0, self.max_text_len - tseq.shape[0])) res = { 'real_text': text, 'padded_text': tseq, 'aligned_codes': aligned_codes, 'aligned_code_lengths': orig_aligned_code_length, 'text_lengths': torch.tensor(orig_text_len, dtype=torch.long), 'wav': wav, 'wav_lengths': torch.tensor(orig_output, dtype=torch.long), 'filenames': path, 'skipped_items': actually_skipped_items, } if self.load_conditioning: res['conditioning'] = cond res['conditioning_contains_self'] = cond_is_self return res def __len__(self): return self.total_size_bytes // 1000 # 1000 cuts down a TSV file to the actual length pretty well. if __name__ == '__main__': batch_sz = 16 params = { 'mode': 'fast_paired_voice_audio', 'path': ['Y:\\libritts\\train-clean-360\\transcribed-w2v.tsv', 'Y:\\clips\\books1\\transcribed-w2v.tsv'], 'phase': 'train', 'n_workers': 0, 'batch_size': batch_sz, 'max_wav_length': 255995, 'max_text_length': 200, 'sample_rate': 22050, 'load_conditioning': True, 'num_conditioning_candidates': 1, 'conditioning_length': 44000, 'use_bpe_tokenizer': False, 'load_aligned_codes': True, } from data import create_dataset, create_dataloader def save(b, i, ib, key, c=None): if c is not None: torchaudio.save(f'{i}_clip_{ib}_{key}_{c}.wav', b[key][ib][c], 22050) else: torchaudio.save(f'{i}_clip_{ib}_{key}.wav', b[key][ib], 22050) ds, c = create_dataset(params, return_collate=True) dl = create_dataloader(ds, params, collate_fn=c) i = 0 m = None for i, b in tqdm(enumerate(dl)): for ib in range(batch_sz): print(f'{i} {ib} {b["real_text"][ib]}') save(b, i, ib, 'wav') if i > 5: break