Create train_glowtts.py
Browse files- train_glowtts.py +117 -0
train_glowtts.py
ADDED
@@ -0,0 +1,117 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
|
3 |
+
# Trainer: Where the ✨️ happens.
|
4 |
+
# TrainingArgs: Defines the set of arguments of the Trainer.
|
5 |
+
from trainer import Trainer, TrainerArgs
|
6 |
+
|
7 |
+
# GlowTTSConfig: all model related values for training, validating and testing.
|
8 |
+
from TTS.tts.configs.glow_tts_config import GlowTTSConfig
|
9 |
+
|
10 |
+
# BaseDatasetConfig: defines name, formatter and path of the dataset.
|
11 |
+
from TTS.tts.configs.shared_configs import BaseDatasetConfig , CharactersConfig
|
12 |
+
from TTS.config.shared_configs import BaseAudioConfig
|
13 |
+
from TTS.tts.datasets import load_tts_samples
|
14 |
+
from TTS.tts.models.glow_tts import GlowTTS
|
15 |
+
from TTS.tts.utils.text.tokenizer import TTSTokenizer
|
16 |
+
from TTS.utils.audio import AudioProcessor
|
17 |
+
|
18 |
+
# we use the same path as this script as our training folder.
|
19 |
+
output_path = os.path.dirname(os.path.abspath(__file__))
|
20 |
+
|
21 |
+
# DEFINE DATASET CONFIG
|
22 |
+
# Set LJSpeech as our target dataset and define its path.
|
23 |
+
# You can also use a simple Dict to define the dataset and pass it to your custom formatter.
|
24 |
+
|
25 |
+
|
26 |
+
dataset_config = BaseDatasetConfig(
|
27 |
+
formatter="mozilla", meta_file_train="metadata.csv", path="/kaggle/input/persian-tts-dataset"
|
28 |
+
)
|
29 |
+
|
30 |
+
audio_config = BaseAudioConfig(
|
31 |
+
sample_rate=22050,
|
32 |
+
do_trim_silence=True,
|
33 |
+
resample=False
|
34 |
+
|
35 |
+
)
|
36 |
+
|
37 |
+
character_config=CharactersConfig(
|
38 |
+
characters='ءابتثجحخدذرزسشصضطظعغفقلمنهويِپچژکگیآأؤإئًَُّ',
|
39 |
+
punctuations='!(),-.:;? ̠،؛؟<>',
|
40 |
+
phonemes='ˈˌːˑpbtdʈɖcɟkɡqɢʔɴŋɲɳnɱmʙrʀⱱɾɽɸβfvθðszʃʒʂʐçʝxɣχʁħʕhɦɬɮʋɹɻjɰlɭʎʟaegiouwyɪʊ̩æɑɔəɚɛɝɨ̃ʉʌʍ0123456789"#$%*+/=ABCDEFGHIJKLMNOPRSTUVWXYZ[]^_{}',
|
41 |
+
pad="<PAD>",
|
42 |
+
eos="<EOS>",
|
43 |
+
bos="<BOS>",
|
44 |
+
blank="<BLNK>",
|
45 |
+
characters_class="TTS.tts.utils.text.characters.IPAPhonemes",
|
46 |
+
)
|
47 |
+
# INITIALIZE THE TRAINING CONFIGURATION
|
48 |
+
# Configure the model. Every config class inherits the BaseTTSConfig.
|
49 |
+
config = GlowTTSConfig(
|
50 |
+
batch_size=8,#batch_size=32,
|
51 |
+
eval_batch_size=4,#eval_batch_size=16,
|
52 |
+
num_loader_workers=0,
|
53 |
+
num_eval_loader_workers=0,
|
54 |
+
run_eval=True,
|
55 |
+
test_delay_epochs=-1,
|
56 |
+
epochs=1000,
|
57 |
+
save_step=1000,
|
58 |
+
text_cleaner="basic_cleaners",
|
59 |
+
use_phonemes=True,
|
60 |
+
phoneme_language="fa",
|
61 |
+
phoneme_cache_path=os.path.join(output_path, "phoneme_cache"),
|
62 |
+
characters=character_config,
|
63 |
+
print_step=25,
|
64 |
+
print_eval=False,
|
65 |
+
mixed_precision=True,
|
66 |
+
output_path=output_path,
|
67 |
+
datasets=[dataset_config],
|
68 |
+
audio=audio_config,
|
69 |
+
test_sentences=[
|
70 |
+
"سلطان محمود در زمستانی سخت به طلخک گفت که: با این جامه ی یک لا در این سرما چه می کنی ",
|
71 |
+
"مردی نزد بقالی آمد و گفت پیاز هم ده تا دهان بدان خو شبوی سازم.",
|
72 |
+
"از مال خود پاره ای گوشت بستان و زیره بایی معطّر بساز",
|
73 |
+
"یک بار هم از جهنم بگویید.",
|
74 |
+
"یکی اسبی به عاریت خواست"
|
75 |
+
],
|
76 |
+
|
77 |
+
|
78 |
+
)
|
79 |
+
|
80 |
+
# INITIALIZE THE AUDIO PROCESSOR
|
81 |
+
# Audio processor is used for feature extraction and audio I/O.
|
82 |
+
# It mainly serves to the dataloader and the training loggers.
|
83 |
+
ap = AudioProcessor.init_from_config(config)
|
84 |
+
|
85 |
+
# INITIALIZE THE TOKENIZER
|
86 |
+
# Tokenizer is used to convert text to sequences of token IDs.
|
87 |
+
# If characters are not defined in the config, default characters are passed to the config
|
88 |
+
tokenizer, config = TTSTokenizer.init_from_config(config)
|
89 |
+
|
90 |
+
# LOAD DATA SAMPLES
|
91 |
+
# Each sample is a list of ```[text, audio_file_path, speaker_name]```
|
92 |
+
# You can define your custom sample loader returning the list of samples.
|
93 |
+
# Or define your custom formatter and pass it to the `load_tts_samples`.
|
94 |
+
# Check `TTS.tts.datasets.load_tts_samples` for more details.
|
95 |
+
train_samples, eval_samples = load_tts_samples(
|
96 |
+
dataset_config,
|
97 |
+
eval_split=True,
|
98 |
+
eval_split_max_size=config.eval_split_max_size,
|
99 |
+
eval_split_size=config.eval_split_size,
|
100 |
+
#formatter=changizer
|
101 |
+
)
|
102 |
+
|
103 |
+
# INITIALIZE THE MODEL
|
104 |
+
# Models take a config object and a speaker manager as input
|
105 |
+
# Config defines the details of the model like the number of layers, the size of the embedding, etc.
|
106 |
+
# Speaker manager is used by multi-speaker models.
|
107 |
+
model = GlowTTS(config, ap, tokenizer, speaker_manager=None)
|
108 |
+
|
109 |
+
# INITIALIZE THE TRAINER
|
110 |
+
# Trainer provides a generic API to train all the 🐸TTS models with all its perks like mixed-precision training,
|
111 |
+
# distributed training, etc.
|
112 |
+
trainer = Trainer(
|
113 |
+
TrainerArgs(), config, output_path, model=model, train_samples=train_samples, eval_samples=eval_samples
|
114 |
+
)
|
115 |
+
|
116 |
+
# AND... 3,2,1... 🚀
|
117 |
+
trainer.fit()
|