Skip to content

Commit 97bd5f9

Browse files
erogolerogol
committed
[ci skip] config update #3 WIP
1 parent a21c0b5 commit 97bd5f9

File tree

4 files changed

+159
-221
lines changed

4 files changed

+159
-221
lines changed

‎TTS/bin/train_tacotron.py

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -14,6 +14,7 @@
1414
from TTS.tts.datasets.preprocess import load_meta_data
1515
from TTS.tts.datasets.TTSDataset import MyDataset
1616
from TTS.tts.layers.losses import TacotronLoss
17+
from TTS.tts.configs.tacotron_config import TacotronConfig
1718
from TTS.tts.utils.generic_utils import setup_model
1819
from TTS.tts.utils.io import save_best_model, save_checkpoint
1920
from TTS.tts.utils.measures import alignment_diagonal_score

‎TTS/tts/configs/config.json

Lines changed: 118 additions & 165 deletions
Original file line numberDiff line numberDiff line change
@@ -1,173 +1,126 @@
11
{
2-
"model": "Tacotron2",
3-
"run_name": "ljspeech-ddc",
4-
"run_description": "tacotron2 with DDC and differential spectral loss.",
5-
6-
// AUDIO PARAMETERS
7-
"audio":{
8-
// stft parameters
9-
"fft_size": 1024, // number of stft frequency levels. Size of the linear spectogram frame.
10-
"win_length": 1024, // stft window length in ms.
11-
"hop_length": 256, // stft window hop-lengh in ms.
12-
"frame_length_ms": null, // stft window length in ms.If null, 'win_length' is used.
13-
"frame_shift_ms": null, // stft window hop-lengh in ms. If null, 'hop_length' is used.
14-
15-
// Audio processing parameters
16-
"sample_rate": 22050, // DATASET-RELATED: wav sample-rate.
17-
"preemphasis": 0.0, // pre-emphasis to reduce spec noise and make it more structured. If 0.0, no -pre-emphasis.
18-
"ref_level_db": 20, // reference level db, theoretically 20db is the sound of air.
19-
20-
// Silence trimming
21-
"do_trim_silence": true,// enable trimming of slience of audio as you load it. LJspeech (true), TWEB (false), Nancy (true)
22-
"trim_db": 60, // threshold for timming silence. Set this according to your dataset.
23-
24-
// Griffin-Lim
25-
"power": 1.5, // value to sharpen wav signals after GL algorithm.
26-
"griffin_lim_iters": 60,// #griffin-lim iterations. 30-60 is a good range. Larger the value, slower the generation.
27-
28-
// MelSpectrogram parameters
29-
"num_mels": 80, // size of the mel spec frame.
30-
"mel_fmin": 50.0, // minimum freq level for mel-spec. ~50 for male and ~95 for female voices. Tune for dataset!!
31-
"mel_fmax": 7600.0, // maximum freq level for mel-spec. Tune for dataset!!
2+
"attention_heads": 4,
3+
"attention_norm": "sigmoid",
4+
"attention_type": "original",
5+
"audio_config": {
6+
"clip_norm": true,
7+
"do_trim_silence": true,
8+
"fft_size": 1024,
9+
"frame_length_ms": null,
10+
"frame_shift_ms": null,
11+
"griffin_lim_iters": 60,
12+
"hop_length": 256,
13+
"max_norm": 4,
14+
"mel_fmax": 7600,
15+
"mel_fmin": 50,
16+
"min_level_db": -100,
17+
"num_mels": 80,
18+
"power": 1.5,
19+
"preemphasis": 0,
20+
"ref_level_db": 20,
21+
"sample_rate": 22050,
22+
"signal_norm": true,
3223
"spec_gain": 1,
33-
34-
// Normalization parameters
35-
"signal_norm": true, // normalize spec values. Mean-Var normalization if 'stats_path' is defined otherwise range normalization defined by the other params.
36-
"min_level_db": -100, // lower bound for normalization
37-
"symmetric_norm": true, // move normalization to range [-1, 1]
38-
"max_norm": 4.0, // scale normalization to range [-max_norm, max_norm] or [0, max_norm]
39-
"clip_norm": true, // clip normalized values into the range.
40-
"stats_path": "/home/erogol/Data/LJSpeech-1.1/scale_stats.npy" // DO NOT USE WITH MULTI_SPEAKER MODEL. scaler stats file computed by 'compute_statistics.py'. If it is defined, mean-std based notmalization is used and other normalization params are ignored
24+
"stats_path": "/home/erogol/Data/LJSpeech-1.1/scale_stats.npy",
25+
"symmetric_norm": true,
26+
"trim_db": 60,
27+
"win_length": 1024
4128
},
42-
43-
// VOCABULARY PARAMETERS
44-
// if custom character set is not defined,
45-
// default set in symbols.py is used
46-
// "characters":{
47-
// "pad": "_",
48-
// "eos": "~",
49-
// "bos": "^",
50-
// "characters": "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!'(),-.:;? ",
51-
// "punctuations":"!'(),-.:;? ",
52-
// "phonemes":"iyɨʉɯuɪʏʊeøɘəɵɤoɛœɜɞʌɔæɐaɶɑɒᵻʘɓǀɗǃʄǂɠǁʛpbtdʈɖcɟkɡqɢʔɴŋɲɳnɱmʙrʀⱱɾɽɸβfvθðszʃʒʂʐçʝxɣχʁħʕhɦɬɮʋɹɻjɰlɭʎʟˈˌːˑʍwɥʜʢʡɕʑɺɧɚ˞ɫ"
53-
// },
54-
55-
// DISTRIBUTED TRAINING
56-
"distributed":{
57-
"backend": "nccl",
58-
"url": "tcp:\/\/localhost:54321"
59-
},
60-
61-
"reinit_layers": [], // give a list of layer names to restore from the given checkpoint. If not defined, it reloads all heuristically matching layers.
62-
63-
// TRAINING
64-
"batch_size": 32, // Batch size for training. Lower values than 32 might cause hard to learn attention. It is overwritten by 'gradual_training'.
65-
"eval_batch_size":16,
66-
"r": 7, // Number of decoder frames to predict per iteration. Set the initial values if gradual training is enabled.
67-
"gradual_training": [[0, 7, 64], [1, 5, 64], [50000, 3, 32], [130000, 2, 32], [290000, 1, 32]], //set gradual training steps [first_step, r, batch_size]. If it is null, gradual training is disabled. For Tacotron, you might need to reduce the 'batch_size' as you proceeed.
68-
"mixed_precision": true, // level of optimization with NVIDIA's apex feature for automatic mixed FP16/FP32 precision (AMP), NOTE: currently only O1 is supported, and use "O1" to activate.
69-
70-
// LOSS SETTINGS
71-
"loss_masking": true, // enable / disable loss masking against the sequence padding.
72-
"decoder_loss_alpha": 0.5, // original decoder loss weight. If > 0, it is enabled
73-
"postnet_loss_alpha": 0.25, // original postnet loss weight. If > 0, it is enabled
74-
"postnet_diff_spec_alpha": 0.25, // differential spectral loss weight. If > 0, it is enabled
75-
"decoder_diff_spec_alpha": 0.25, // differential spectral loss weight. If > 0, it is enabled
76-
"decoder_ssim_alpha": 0.5, // decoder ssim loss weight. If > 0, it is enabled
77-
"postnet_ssim_alpha": 0.25, // postnet ssim loss weight. If > 0, it is enabled
78-
"ga_alpha": 5.0, // weight for guided attention loss. If > 0, guided attention is enabled.
79-
"stopnet_pos_weight": 15.0, // pos class weight for stopnet loss since there are way more negative samples than positive samples.
80-
81-
82-
// VALIDATION
83-
"run_eval": true,
84-
"test_delay_epochs": 10, //Until attention is aligned, testing only wastes computation time.
85-
"test_sentences_file": null, // set a file to load sentences to be used for testing. If it is null then we use default english sentences.
86-
87-
// OPTIMIZER
88-
"noam_schedule": false, // use noam warmup and lr schedule.
89-
"grad_clip": 1.0, // upper limit for gradients for clipping.
90-
"epochs": 1000, // total number of epochs to train.
91-
"lr": 0.0001, // Initial learning rate. If Noam decay is active, maximum learning rate.
92-
"wd": 0.000001, // Weight decay weight.
93-
"warmup_steps": 4000, // Noam decay steps to increase the learning rate from 0 to "lr"
94-
"seq_len_norm": false, // Normalize eash sample loss with its length to alleviate imbalanced datasets. Use it if your dataset is small or has skewed distribution of sequence lengths.
95-
96-
// TACOTRON PRENET
97-
"memory_size": -1, // ONLY TACOTRON - size of the memory queue used fro storing last decoder predictions for auto-regression. If < 0, memory queue is disabled and decoder only uses the last prediction frame.
98-
"prenet_type": "original", // "original" or "bn".
99-
"prenet_dropout": true, // enable/disable dropout at prenet.
100-
101-
// TACOTRON ATTENTION
102-
"attention_type": "original", // 'original' , 'graves', 'dynamic_convolution'
103-
"attention_heads": 4, // number of attention heads (only for 'graves')
104-
"attention_norm": "sigmoid", // softmax or sigmoid.
105-
"windowing": false, // Enables attention windowing. Used only in eval mode.
106-
"use_forward_attn": false, // if it uses forward attention. In general, it aligns faster.
107-
"forward_attn_mask": false, // Additional masking forcing monotonicity only in eval mode.
108-
"transition_agent": false, // enable/disable transition agent of forward attention.
109-
"location_attn": true, // enable_disable location sensitive attention. It is enabled for TACOTRON by default.
110-
"bidirectional_decoder": false, // use https://arxiv.org/abs/1907.09006. Use it, if attention does not work well with your dataset.
111-
"double_decoder_consistency": true, // use DDC explained here https://erogol.com/solving-attention-problems-of-tts-models-with-double-decoder-consistency-draft/
112-
"ddc_r": 7, // reduction rate for coarse decoder.
113-
114-
// STOPNET
115-
"stopnet": true, // Train stopnet predicting the end of synthesis.
116-
"separate_stopnet": true, // Train stopnet seperately if 'stopnet==true'. It prevents stopnet loss to influence the rest of the model. It causes a better model, but it trains SLOWER.
117-
118-
// TENSORBOARD and LOGGING
119-
"print_step": 25, // Number of steps to log training on console.
120-
"tb_plot_step": 100, // Number of steps to plot TB training figures.
121-
"print_eval": false, // If True, it prints intermediate loss values in evalulation.
122-
"save_step": 10000, // Number of training steps expected to save traninpg stats and checkpoints.
123-
"checkpoint": true, // If true, it saves checkpoints per "save_step"
124-
"keep_all_best": false, // If true, keeps all best_models after keep_after steps
125-
"keep_after": 10000, // Global step after which to keep best models if keep_all_best is true
126-
"tb_model_param_stats": false, // true, plots param stats per layer on tensorboard. Might be memory consuming, but good for debugging.
127-
128-
// DATA LOADING
129-
"text_cleaner": "phoneme_cleaners",
130-
"enable_eos_bos_chars": false, // enable/disable beginning of sentence and end of sentence chars.
131-
"num_loader_workers": 4, // number of training data loader processes. Don't set it too big. 4-8 are good values.
132-
"num_val_loader_workers": 4, // number of evaluation data loader processes.
133-
"batch_group_size": 4, //Number of batches to shuffle after bucketing.
134-
"min_seq_len": 6, // DATASET-RELATED: minimum text length to use in training
135-
"max_seq_len": 153, // DATASET-RELATED: maximum text length
136-
"compute_input_seq_cache": false, // if true, text sequences are computed before starting training. If phonemes are enabled, they are also computed at this stage.
137-
"use_noise_augment": true,
138-
139-
// PATHS
140-
"output_path": "/home/erogol/Models/LJSpeech/",
141-
142-
// PHONEMES
143-
"phoneme_cache_path": "/home/erogol/Models/phoneme_cache/", // phoneme computation is slow, therefore, it caches results in the given folder.
144-
"use_phonemes": true, // use phonemes instead of raw characters. It is suggested for better pronounciation.
145-
"phoneme_language": "en-us", // depending on your target language, pick one from https://github.com/bootphon/phonemizer#languages
146-
147-
// MULTI-SPEAKER and GST
148-
"use_speaker_embedding": false, // use speaker embedding to enable multi-speaker learning.
149-
"use_gst": false, // use global style tokens
150-
"use_external_speaker_embedding_file": false, // if true, forces the model to use external embedding per sample instead of nn.embeddings, that is, it supports external embeddings such as those used at: https://arxiv.org/abs /1806.04558
151-
"external_speaker_embedding_file": "../../speakers-vctk-en.json", // if not null and use_external_speaker_embedding_file is true, it is used to load a specific embedding file and thus uses these embeddings instead of nn.embeddings, that is, it supports external embeddings such as those used at: https://arxiv.org/abs /1806.04558
152-
"gst": { // gst parameter if gst is enabled
153-
"gst_style_input": null, // Condition the style input either on a
154-
// -> wave file [path to wave] or
155-
// -> dictionary using the style tokens {'token1': 'value', 'token2': 'value'} example {"0": 0.15, "1": 0.15, "5": -0.15}
156-
// with the dictionary being len(dict) <= len(gst_num_style_tokens).
157-
"gst_embedding_dim": 512,
158-
"gst_num_heads": 4,
159-
"gst_num_style_tokens": 10,
160-
"gst_use_speaker_embedding": false
161-
},
162-
163-
// DATASETS
164-
"datasets": // List of datasets. They all merged and they get different speaker_ids.
29+
"bidirectional_decoder": false,
30+
"compute_input_seq_cache": false,
31+
"ddc_r": 7,
32+
"decoder_diff_spec_alpha": 0.25,
33+
"decoder_loss_alpha": 0.5,
34+
"decoder_ssim_alpha": 0.5,
35+
"double_decoder_consistency": true,
36+
"enable_eos_bos_chars": false,
37+
"forward_attn_mask": false,
38+
"ga_alpha": 5,
39+
"grad_clip": 1,
40+
"gradual_training": [
41+
[
42+
0,
43+
7,
44+
64
45+
],
46+
[
47+
1,
48+
5,
49+
64
50+
],
51+
[
52+
50000,
53+
3,
54+
32
55+
],
16556
[
57+
130000,
58+
2,
59+
32
60+
],
61+
[
62+
290000,
63+
1,
64+
32
65+
]
66+
],
67+
"location_attn": true,
68+
"lr": 0.0001,
69+
"memory_size": -1,
70+
"noam_schedule": false,
71+
"phoneme_cache_path": "/home/erogol/Models/phoneme_cache/",
72+
"phoneme_language": "en-us",
73+
"postnet_diff_spec_alpha": 0.25,
74+
"postnet_loss_alpha": 0.25,
75+
"postnet_ssim_alpha": 0.25,
76+
"prenet_dropout": false,
77+
"prenet_type": "original",
78+
"r": 7,
79+
"separate_stopnet": true,
80+
"seq_len_norm": false,
81+
"stopnet": true,
82+
"stopnet_pos_weight": 15,
83+
"test_sentences_file": null,
84+
"text_cleaner": "phoneme_cleaners",
85+
"training_config": {
86+
"batch_group_size": 4,
87+
"batch_size": 32,
88+
"checkpoint": true,
89+
"datasets": [
16690
{
91+
"meta_file_train": "metadata.csv",
92+
"meta_file_val": null,
16793
"name": "ljspeech",
168-
"path": "/home/erogol/Data/LJSpeech-1.1/",
169-
"meta_file_train": "metadata.csv", // for vtck if list, ignore speakers id in list for train, its useful for test cloning with new speakers
170-
"meta_file_val": null
94+
"path": "/home/erogol/Data/LJSpeech-1.1/"
17195
}
172-
]
96+
],
97+
"epochs": 1000,
98+
"eval_batch_size": 16,
99+
"keep_after": 10000,
100+
"keep_all_best": false,
101+
"loss_masking": true,
102+
"max_seq_len": 153,
103+
"min_seq_len": 6,
104+
"mixed_precision": true,
105+
"model": "Tacotron2",
106+
"num_loader_workers": 4,
107+
"num_val_loader_workers": 4,
108+
"output_path": "/home/erogol/Models/LJSpeech/",
109+
"print_eval": false,
110+
"print_step": 25,
111+
"run_description": "tacotron2 with DDC and differential spectral loss.",
112+
"run_eval": true,
113+
"run_name": "ljspeech-ddc",
114+
"save_step": 10000,
115+
"tb_model_param_stats": false,
116+
"tb_plot_step": 100,
117+
"test_delay_epochs": 10,
118+
"use_noise_augment": true
119+
},
120+
"transition_agent": false,
121+
"use_forward_attn": false,
122+
"use_phonemes": true,
123+
"warmup_steps": 4000,
124+
"wd": 0.000001,
125+
"windowing": false
173126
}

‎TTS/utils/arguments.py

Lines changed: 14 additions & 29 deletions
Original file line numberDiff line numberDiff line change
@@ -117,16 +117,11 @@ def get_last_checkpoint(path):
117117
return last_models["checkpoint"], last_models["best_model"]
118118

119119

120-
def process_args(args, model_class):
121-
"""Process parsed comand line arguments based on model class (tts or vocoder).
120+
def process_args(args, config, tb_prefix):
121+
"""Process parsed comand line arguments.
122122
123123
Args:
124124
args (argparse.Namespace or dict like): Parsed input arguments.
125-
model_type (str): Model type used to check config parameters and setup
126-
the TensorBoard logger. One of ['tts', 'vocoder'].
127-
128-
Raises:
129-
ValueError: If `model_type` is not one of implemented choices.
130125
131126
Returns:
132127
c (TTS.utils.io.AttrDict): Config paramaters.
@@ -138,28 +133,21 @@ def process_args(args, model_class):
138133
the TensorBoard loggind.
139134
"""
140135
if args.continue_path:
136+
# continue a previous training from its output folder
141137
args.output_path = args.continue_path
142138
args.config_path = os.path.join(args.continue_path, "config.json")
143139
args.restore_path, best_model = get_last_checkpoint(args.continue_path)
144140
if not args.best_path:
145141
args.best_path = best_model
146-
147142
# setup output paths and read configs
148-
c = load_config(args.config_path)
149-
_ = os.path.dirname(os.path.realpath(__file__))
150-
151-
if "mixed_precision" in c and c.mixed_precision:
143+
c = config.load_json(args.config_path)
144+
if c.mixed_precision:
152145
print(" > Mixed precision mode is ON")
153-
154-
out_path = args.continue_path
155-
if not out_path:
156-
out_path = create_experiment_folder(c.output_path, c.run_name, args.debug)
157-
146+
if not os.path.exists(c.output_path):
147+
out_path = create_experiment_folder(c.output_path, c.run_name,
148+
args.debug)
158149
audio_path = os.path.join(out_path, "test_audios")
159-
160-
c_logger = ConsoleLogger()
161-
tb_logger = None
162-
150+
# setup rank 0 process in distributed training
163151
if args.rank == 0:
164152
os.makedirs(audio_path, exist_ok=True)
165153
new_fields = {}
@@ -169,18 +157,15 @@ def process_args(args, model_class):
169157
# if model characters are not set in the config file
170158
# save the default set to the config file for future
171159
# compatibility.
172-
if model_class == "tts" and "characters" not in c:
160+
if c.has('characters_config'):
173161
used_characters = parse_symbols()
174162
new_fields["characters"] = used_characters
175163
copy_model_files(c, args.config_path, out_path, new_fields)
176164
os.chmod(audio_path, 0o775)
177165
os.chmod(out_path, 0o775)
178-
179166
log_path = out_path
180-
181-
tb_logger = TensorboardLogger(log_path, model_name=model_class.upper())
182-
183-
# write model config to tensorboard
184-
tb_logger.tb_add_text("model-config", f"<pre>{json.dumps(c, indent=4)}</pre>", 0)
185-
167+
tb_logger = TensorboardLogger(log_path, model_name=tb_prefix)
168+
# write model desc to tensorboard
169+
tb_logger.tb_add_text("model-description", c["run_description"], 0)
170+
c_logger = ConsoleLogger()
186171
return c, out_path, audio_path, c_logger, tb_logger

0 commit comments

Comments
 (0)