Spaces:
Running on Zero
Running on Zero
File size: 13,703 Bytes
cc348e2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 | import random
import torch
import torch.multiprocessing
from torch.utils.data import ConcatDataset
from tqdm import tqdm
from TrainingInterfaces.Text_to_Spectrogram.FastSpeech2.FastSpeech2 import FastSpeech2
from TrainingInterfaces.Text_to_Spectrogram.FastSpeech2.meta_train_loop import train_loop
from Utility.corpus_preparation import prepare_fastspeech_corpus
from Utility.path_to_transcript_dicts import *
def run(gpu_id, resume_checkpoint, finetune, model_dir, resume, remove_faulty_samples=False):
# It is not recommended training this yourself or to finetune this, but you can.
# The recommended use is to download the pretrained model from the github release
# page and finetune to your desired data similar to how it is showcased in
# FastSpeech2_finetune_to_German.py
torch.manual_seed(131714)
random.seed(131714)
torch.random.manual_seed(131714)
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = f"{gpu_id}"
datasets = list()
base_dir = os.path.join("Models", "FastSpeech2_Meta")
if model_dir is not None:
meta_save_dir = model_dir
else:
meta_save_dir = base_dir
os.makedirs(meta_save_dir, exist_ok=True)
print("Preparing")
english_datasets = list()
german_datasets = list()
greek_datasets = list()
spanish_datasets = list()
finnish_datasets = list()
russian_datasets = list()
hungarian_datasets = list()
dutch_datasets = list()
french_datasets = list()
portuguese_datasets = list()
polish_datasets = list()
italian_datasets = list()
chinese_datasets = list()
vietnamese_datasets = list()
english_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_nancy(),
corpus_dir=os.path.join("Corpora", "Nancy"),
lang="en"))
english_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_ljspeech(),
corpus_dir=os.path.join("Corpora", "LJSpeech"),
lang="en"))
english_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_libritts_all_clean(),
corpus_dir=os.path.join("Corpora", "libri_all_clean"),
lang="en"))
english_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_vctk(),
corpus_dir=os.path.join("Corpora", "vctk"),
lang="en"))
english_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_nvidia_hifitts(),
corpus_dir=os.path.join("Corpora", "hifi"),
lang="en"))
german_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_karlsson(),
corpus_dir=os.path.join("Corpora", "Karlsson"),
lang="de"))
german_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_eva(),
corpus_dir=os.path.join("Corpora", "Eva"),
lang="de"))
german_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_hokus(),
corpus_dir=os.path.join("Corpora", "Hokus"),
lang="de"))
german_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_bernd(),
corpus_dir=os.path.join("Corpora", "Bernd"),
lang="de"))
german_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_hui_others(),
corpus_dir=os.path.join("Corpora", "hui_others"),
lang="de"))
german_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_thorsten(),
corpus_dir=os.path.join("Corpora", "Thorsten"),
lang="de"))
greek_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_css10el(),
corpus_dir=os.path.join("Corpora", "meta_Greek"),
lang="el"))
spanish_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_spanish_blizzard_train(),
corpus_dir=os.path.join("Corpora", "spanish_blizzard"),
lang="es"))
spanish_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_css10es(),
corpus_dir=os.path.join("Corpora", "meta_Spanish"),
lang="es"))
spanish_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_mls_spanish(),
corpus_dir=os.path.join("Corpora", "mls_spanish"),
lang="es"))
finnish_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_css10fi(),
corpus_dir=os.path.join("Corpora", "meta_Finnish"),
lang="fi"))
russian_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_css10ru(),
corpus_dir=os.path.join("Corpora", "meta_Russian"),
lang="ru"))
hungarian_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_css10hu(),
corpus_dir=os.path.join("Corpora", "meta_Hungarian"),
lang="hu"))
dutch_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_css10nl(),
corpus_dir=os.path.join("Corpora", "meta_Dutch"),
lang="nl"))
dutch_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_mls_dutch(),
corpus_dir=os.path.join("Corpora", "mls_dutch"),
lang="nl"))
french_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_css10fr(),
corpus_dir=os.path.join("Corpora", "meta_French"),
lang="fr"))
french_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_mls_french(),
corpus_dir=os.path.join("Corpora", "mls_french"),
lang="fr"))
portuguese_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_mls_portuguese(),
corpus_dir=os.path.join("Corpora", "mls_porto"),
lang="pt"))
polish_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_mls_polish(),
corpus_dir=os.path.join("Corpora", "mls_polish"),
lang="pl"))
italian_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_mls_italian(),
corpus_dir=os.path.join("Corpora", "mls_italian"),
lang="it"))
chinese_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_css10cmn(),
corpus_dir=os.path.join("Corpora", "css10_chinese"),
lang="cmn"))
chinese_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_aishell3(),
corpus_dir=os.path.join("Corpora", "aishell3"),
lang="cmn"))
vietnamese_datasets.append(prepare_fastspeech_corpus(transcript_dict=build_path_to_transcript_dict_vietTTS(),
corpus_dir=os.path.join("Corpora", "vietTTS"),
lang="vi"))
datasets.append(ConcatDataset(english_datasets))
datasets.append(ConcatDataset(german_datasets))
datasets.append(ConcatDataset(greek_datasets))
datasets.append(ConcatDataset(spanish_datasets))
datasets.append(ConcatDataset(finnish_datasets))
datasets.append(ConcatDataset(russian_datasets))
datasets.append(ConcatDataset(hungarian_datasets))
datasets.append(ConcatDataset(dutch_datasets))
datasets.append(ConcatDataset(french_datasets))
datasets.append(ConcatDataset(portuguese_datasets))
datasets.append(ConcatDataset(polish_datasets))
datasets.append(ConcatDataset(italian_datasets))
datasets.append(ConcatDataset(chinese_datasets))
datasets.append(ConcatDataset(vietnamese_datasets))
if remove_faulty_samples:
find_and_remove_faulty_samples(net=FastSpeech2(lang_embs=100),
datasets=english_datasets +
german_datasets +
greek_datasets +
spanish_datasets +
finnish_datasets +
russian_datasets +
hungarian_datasets +
dutch_datasets +
french_datasets +
portuguese_datasets +
polish_datasets +
italian_datasets +
chinese_datasets +
vietnamese_datasets,
device=torch.device("cuda"),
path_to_checkpoint=resume_checkpoint)
train_loop(net=FastSpeech2(lang_embs=100),
device=torch.device("cuda"),
datasets=datasets,
batch_size=6,
save_directory=meta_save_dir,
steps=300000,
steps_per_checkpoint=1000,
lr=0.001,
path_to_checkpoint=resume_checkpoint,
resume=resume)
@torch.inference_mode()
def find_and_remove_faulty_samples(net,
datasets,
device,
path_to_checkpoint):
net = net.to(device)
torch.multiprocessing.set_sharing_strategy('file_system')
check_dict = torch.load(os.path.join(path_to_checkpoint), map_location=device)
net.load_state_dict(check_dict["model"])
for dataset_index in range(len(datasets)):
nan_ids = list()
for datapoint_index in tqdm(range(len(datasets[dataset_index]))):
loss = net(text_tensors=datasets[dataset_index][datapoint_index][0].unsqueeze(0).to(device),
text_lengths=datasets[dataset_index][datapoint_index][1].to(device),
gold_speech=datasets[dataset_index][datapoint_index][2].unsqueeze(0).to(device),
speech_lengths=datasets[dataset_index][datapoint_index][3].to(device),
gold_durations=datasets[dataset_index][datapoint_index][4].unsqueeze(0).to(device),
gold_pitch=datasets[dataset_index][datapoint_index][6].unsqueeze(0).to(device), # mind the switched order
gold_energy=datasets[dataset_index][datapoint_index][5].unsqueeze(0).to(device), # mind the switched order
utterance_embedding=datasets[dataset_index][datapoint_index][7].unsqueeze(0).to(device),
lang_ids=datasets[dataset_index][datapoint_index][8].unsqueeze(0).to(device),
return_mels=False).squeeze()
if torch.isnan(loss):
print(f"NAN DETECTED: {dataset_index}, {datapoint_index}")
nan_ids.append(datapoint_index)
datasets[dataset_index].remove_samples(nan_ids)
|