diff --git a/codes/data/audio/paired_voice_audio_dataset.py b/codes/data/audio/paired_voice_audio_dataset.py index 9d50d1c4..d6123229 100644 --- a/codes/data/audio/paired_voice_audio_dataset.py +++ b/codes/data/audio/paired_voice_audio_dataset.py @@ -15,39 +15,49 @@ from models.tacotron2.text import text_to_sequence, sequence_to_text from utils.util import opt_get -def parse_libri(line, base_path, split="|"): - fpt = line.strip().split(split) - fpt[0] = os.path.join(base_path, fpt[0]) - return fpt +def load_tsv(filename): + with open(filename, encoding='utf-8') as f: + components = [line.strip().split('\t') for line in f] + base = os.path.dirname(filename) + filepaths_and_text = [[os.path.join(base, f'{component[1]}'), component[0]] for component in components] + return filepaths_and_text -def parse_tsv(line, base_path): - fpt = line.strip().split('\t') - return os.path.join(base_path, f'{fpt[1]}'), fpt[0] +def load_tsv_aligned_codes(filename): + with open(filename, encoding='utf-8') as f: + components = [line.strip().split('\t') for line in f] + base = os.path.dirname(filename) + 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) + filepaths_and_text = [[os.path.join(base, f'{component[1]}'), component[0], convert_string_list_to_tensor(component[2])] for component in components] + return filepaths_and_text -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]) +def load_mozilla_cv(filename): + with open(filename, encoding='utf-8') as f: + components = [line.strip().split('\t') for line in f][1:] # First line is the header + base = os.path.dirname(filename) + filepaths_and_text = [[os.path.join(base, f'clips/{component[1]}'), component[2]] for component in components] + return filepaths_and_text -def parse_mozilla_cv(line, base_path): - components = line.strip().split('\t') - return os.path.join(base_path, f'clips/{components[1]}'), components[2] - - -def parse_voxpopuli(line, base_path): - line = line.strip().split('\t') - file, raw_text, norm_text, speaker_id, split, gender = line - year = file[:4] - return os.path.join(base_path, year, f'{file}.ogg.wav'), raw_text +def load_voxpopuli(filename): + with open(filename, encoding='utf-8') as f: + lines = [line.strip().split('\t') for line in f][1:] # First line is the header + base = os.path.dirname(filename) + filepaths_and_text = [] + for line in lines: + if len(line) == 0: + continue + file, raw_text, norm_text, speaker_id, split, gender = line + year = file[:4] + filepaths_and_text.append([os.path.join(base, year, f'{file}.ogg.wav'), raw_text]) + return filepaths_and_text class CharacterTokenizer: @@ -60,16 +70,14 @@ class CharacterTokenizer: class TextWavLoader(torch.utils.data.Dataset): 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.path = hparams['path'] + if not isinstance(self.path, list): + self.path = [self.path] - self.fetcher_mode = opt_get(hparams, ['fetcher_mode'], 'lj') - if not isinstance(self.fetcher_mode, list): - self.fetcher_mode = [self.fetcher_mode] - assert len(self.paths) == len(self.fetcher_mode) + fetcher_mode = opt_get(hparams, ['fetcher_mode'], 'lj') + if not isinstance(fetcher_mode, list): + fetcher_mode = [fetcher_mode] + assert len(self.path) == len(fetcher_mode) self.load_conditioning = opt_get(hparams, ['load_conditioning'], False) self.conditioning_candidates = opt_get(hparams, ['num_conditioning_candidates'], 1) @@ -77,8 +85,25 @@ class TextWavLoader(torch.utils.data.Dataset): self.debug_failures = opt_get(hparams, ['debug_loading_failures'], False) self.load_aligned_codes = opt_get(hparams, ['load_aligned_codes'], False) self.aligned_codes_to_audio_ratio = opt_get(hparams, ['aligned_codes_ratio'], 443) + self.audiopaths_and_text = [] + for p, fm in zip(self.path, fetcher_mode): + if fm == 'lj' or fm == 'libritts': + fetcher_fn = load_filepaths_and_text + elif fm == 'tsv': + fetcher_fn = load_tsv_aligned_codes if self.load_aligned_codes else load_tsv + elif fm == 'mozilla_cv': + assert not self.load_conditioning # Conditioning inputs are incompatible with mozilla_cv + fetcher_fn = load_mozilla_cv + elif fm == 'voxpopuli': + assert not self.load_conditioning # Conditioning inputs are incompatible with voxpopuli + fetcher_fn = load_voxpopuli + else: + raise NotImplementedError() + self.audiopaths_and_text.extend(fetcher_fn(p)) self.text_cleaners = hparams.text_cleaners self.sample_rate = hparams.sample_rate + random.seed(hparams.seed) + random.shuffle(self.audiopaths_and_text) 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 @@ -109,64 +134,23 @@ class TextWavLoader(torch.utils.data.Dataset): 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] - fm = self.fetcher_mode[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) - if fm == 'lj' or fm == 'libritts': - return parse_libri(l2, base_path) - elif fm == 'tsv': - return parse_tsv_aligned_codes(l2, base_path) if self.load_aligned_codes else parse_tsv(l2, base_path) - elif fm == 'mozilla_cv': - assert not self.load_conditioning # Conditioning inputs are incompatible with mozilla_cv - return parse_mozilla_cv(l2, base_path) - elif fm == 'voxpopuli': - assert not self.load_conditioning # Conditioning inputs are incompatible with voxpopuli - return parse_voxpopuli(l2, base_path) - else: - raise NotImplementedError() - 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) + tseq, wav, text, path = self.get_wav_text_pair(self.audiopaths_and_text[index]) 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, + cond, cond_is_self = load_similar_clips(self.audiopaths_and_text[index][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()}") + print(f"error loading {self.audiopaths_and_text[index][0]} {sys.exc_info()}") return self[(index+1) % len(self)] if self.load_aligned_codes: - aligned_codes = apt[2] + aligned_codes = self.audiopaths_and_text[index][2] actually_skipped_items = self.skipped_items self.skipped_items = 0 @@ -205,7 +189,7 @@ class TextWavLoader(torch.utils.data.Dataset): return res def __len__(self): - return self.total_size_bytes // 1000 # 1000 cuts down a TSV file to the actual length pretty well, but doesn't work with the other formats. + return len(self.audiopaths_and_text) class PairedVoiceDebugger: @@ -240,23 +224,22 @@ class PairedVoiceDebugger: if __name__ == '__main__': - batch_sz = 16 + batch_sz = 8 params = { 'mode': 'paired_voice_audio', - #'path': ['Y:\\clips\\books1\\transcribed-w2v.tsv'], - 'path': ['Y:\\bigasr_dataset\\mozcv\\en\\train.tsv'], - 'fetcher_mode': ['mozilla_cv'], + 'path': ['Y:\\clips\\books1\\transcribed-w2v.tsv'], + 'fetcher_mode': ['tsv'], 'phase': 'train', 'n_workers': 0, 'batch_size': batch_sz, 'max_wav_length': 255995, 'max_text_length': 200, 'sample_rate': 22050, - 'load_conditioning': False, + 'load_conditioning': True, 'num_conditioning_candidates': 2, 'conditioning_length': 44000, 'use_bpe_tokenizer': True, - 'load_aligned_codes': False, + 'load_aligned_codes': True, } from data import create_dataset, create_dataloader @@ -273,7 +256,7 @@ if __name__ == '__main__': 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 + save(b, i, ib, 'wav') + if i > 5: + break diff --git a/codes/train.py b/codes/train.py index 0e6e6699..d249b3cd 100644 --- a/codes/train.py +++ b/codes/train.py @@ -300,7 +300,7 @@ class Trainer: if __name__ == '__main__': parser = argparse.ArgumentParser() - parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../experiments/train_gpt_asr_mass_hf2_audio_only_fp32/train_gpt_asr_mass_hf2.yml') + parser.add_argument('-opt', type=str, help='Path to option YAML file.', default='../options/train_diffusion_tts.yml') parser.add_argument('--launcher', choices=['none', 'pytorch'], default='none', help='job launcher') parser.add_argument('--local_rank', type=int, default=0) args = parser.parse_args()