Commit
·
2b1b519
1
Parent(s):
2914730
Update inference.py and meldataset,py
Browse files- inference.py +48 -9
- meldataset.py +129 -40
inference.py
CHANGED
@@ -65,9 +65,31 @@ class StyleTTS2(torch.nn.Module):
|
|
65 |
super().__init__()
|
66 |
self.register_buffer("get_device", torch.empty(0))
|
67 |
self.preprocess = Preprocess()
|
68 |
-
|
69 |
-
config = yaml.safe_load(open(config_path))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
70 |
args = self.__recursive_munch(config['model_params'])
|
|
|
|
|
|
|
71 |
|
72 |
assert args.decoder.type in ['hifigan'], 'Decoder type unknown'
|
73 |
|
@@ -186,7 +208,7 @@ class StyleTTS2(torch.nn.Module):
|
|
186 |
speed = min(max(speed, 0.0001), 2) #speed range [0, 2]
|
187 |
|
188 |
phonem = ' '.join(word_tokenize(phonem))
|
189 |
-
tokens =
|
190 |
tokens.insert(0, 0)
|
191 |
tokens.append(0)
|
192 |
tokens = torch.LongTensor(tokens).to(device).unsqueeze(0)
|
@@ -231,17 +253,34 @@ class StyleTTS2(torch.nn.Module):
|
|
231 |
|
232 |
return out.squeeze().cpu().numpy(), duration.mean()
|
233 |
|
234 |
-
def get_styles(self, speaker, denoise=0.3, avg_style=True):
|
235 |
-
if
|
236 |
-
|
237 |
-
|
238 |
-
|
|
|
|
|
|
|
239 |
style = {
|
240 |
-
'style': ref_s,
|
241 |
'path': speaker['path'],
|
242 |
'speed': speaker['speed'],
|
243 |
}
|
244 |
return style
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
245 |
|
246 |
def generate(self, phonem, style, stabilize=True, n_merge=16):
|
247 |
if stabilize: smooth_value=0.2
|
|
|
65 |
super().__init__()
|
66 |
self.register_buffer("get_device", torch.empty(0))
|
67 |
self.preprocess = Preprocess()
|
68 |
+
self.ref_s = None
|
69 |
+
config = yaml.safe_load(open(config_path, "r", encoding="utf-8"))
|
70 |
+
|
71 |
+
try:
|
72 |
+
symbols = (
|
73 |
+
list(config['symbol']['pad']) +
|
74 |
+
list(config['symbol']['punctuation']) +
|
75 |
+
list(config['symbol']['letters']) +
|
76 |
+
list(config['symbol']['letters_ipa']) +
|
77 |
+
list(config['symbol']['extend'])
|
78 |
+
)
|
79 |
+
symbol_dict = {}
|
80 |
+
for i in range(len((symbols))):
|
81 |
+
symbol_dict[symbols[i]] = i
|
82 |
+
|
83 |
+
n_token = len(symbol_dict) + 1
|
84 |
+
print("\nFound:", n_token, "symbols")
|
85 |
+
except Exception as e:
|
86 |
+
print(f"\nERROR: Cannot find {e} in config file!\nYour config file is likely outdated, please download updated version from the repository.")
|
87 |
+
raise SystemExit(1)
|
88 |
+
|
89 |
args = self.__recursive_munch(config['model_params'])
|
90 |
+
args['n_token'] = n_token
|
91 |
+
|
92 |
+
self.cleaner = TextCleaner(symbol_dict, debug=False)
|
93 |
|
94 |
assert args.decoder.type in ['hifigan'], 'Decoder type unknown'
|
95 |
|
|
|
208 |
speed = min(max(speed, 0.0001), 2) #speed range [0, 2]
|
209 |
|
210 |
phonem = ' '.join(word_tokenize(phonem))
|
211 |
+
tokens = self.cleaner(phonem)
|
212 |
tokens.insert(0, 0)
|
213 |
tokens.append(0)
|
214 |
tokens = torch.LongTensor(tokens).to(device).unsqueeze(0)
|
|
|
253 |
|
254 |
return out.squeeze().cpu().numpy(), duration.mean()
|
255 |
|
256 |
+
def get_styles(self, speaker, denoise=0.3, avg_style=True, load_styles=False):
|
257 |
+
if not load_styles:
|
258 |
+
if avg_style: split_dur = 3
|
259 |
+
else: split_dur = 0
|
260 |
+
self.ref_s = self.__compute_style(speaker['path'], denoise=denoise, split_dur=split_dur)
|
261 |
+
else:
|
262 |
+
if self.ref_s is None:
|
263 |
+
raise Exception("Have to compute or load the styles first!")
|
264 |
style = {
|
265 |
+
'style': self.ref_s,
|
266 |
'path': speaker['path'],
|
267 |
'speed': speaker['speed'],
|
268 |
}
|
269 |
return style
|
270 |
+
|
271 |
+
def save_styles(self, save_dir):
|
272 |
+
if self.ref_s is not None:
|
273 |
+
torch.save(self.ref_s, save_dir)
|
274 |
+
print("Saved styles!")
|
275 |
+
else:
|
276 |
+
raise Exception("Have to compute the styles before saving it.")
|
277 |
+
|
278 |
+
def load_styles(self, save_dir):
|
279 |
+
try:
|
280 |
+
self.ref_s = torch.load(save_dir)
|
281 |
+
print("Loaded styles!")
|
282 |
+
except Exception as e:
|
283 |
+
print(e)
|
284 |
|
285 |
def generate(self, phonem, style, stabilize=True, n_merge=16):
|
286 |
if stabilize: smooth_value=0.2
|
meldataset.py
CHANGED
@@ -1,7 +1,5 @@
|
|
1 |
#coding: utf-8
|
2 |
-
import os
|
3 |
import os.path as osp
|
4 |
-
import time
|
5 |
import random
|
6 |
import numpy as np
|
7 |
import random
|
@@ -9,10 +7,10 @@ import soundfile as sf
|
|
9 |
import librosa
|
10 |
|
11 |
import torch
|
12 |
-
from torch import nn
|
13 |
-
import torch.nn.functional as F
|
14 |
import torchaudio
|
15 |
-
|
|
|
|
|
16 |
|
17 |
import logging
|
18 |
logger = logging.getLogger(__name__)
|
@@ -20,33 +18,19 @@ logger.setLevel(logging.DEBUG)
|
|
20 |
|
21 |
import pandas as pd
|
22 |
|
23 |
-
##########################################################
|
24 |
-
_pad = "$"
|
25 |
-
_punctuation = ';:,.!?¡¿—…"«»“” '
|
26 |
-
_letters = 'ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz'
|
27 |
-
_letters_ipa = "ɑɐɒæɓʙβɔɕçɗɖðʤəɘɚɛɜɝɞɟʄɡɠɢʛɦɧħɥʜɨɪʝɭɬɫɮʟɱɯɰŋɳɲɴøɵɸθœɶʘɹɺɾɻʀʁɽʂʃʈʧʉʊʋⱱʌɣɤʍχʎʏʑʐʒʔʡʕʢǀǁǂǃˈˌːˑʼʴʰʱʲʷˠˤ˞↓↑→↗↘'̩'ᵻ"
|
28 |
-
_extend = "" #ADD MORE SYMBOLS HERE
|
29 |
-
|
30 |
-
# Export all symbols:
|
31 |
-
symbols = [_pad] + list(_punctuation) + list(_letters) + list(_letters_ipa) + list(_extend)
|
32 |
-
|
33 |
-
dicts = {}
|
34 |
-
for i in range(len((symbols))):
|
35 |
-
dicts[symbols[i]] = i
|
36 |
-
|
37 |
-
# Copy this code somewhere else then run with print(len(dicts) + 1) to check total symbols
|
38 |
-
##########################################################
|
39 |
-
|
40 |
class TextCleaner:
|
41 |
-
def __init__(self,
|
42 |
-
self.word_index_dictionary =
|
|
|
43 |
def __call__(self, text):
|
44 |
indexes = []
|
45 |
for char in text:
|
46 |
try:
|
47 |
indexes.append(self.word_index_dictionary[char])
|
48 |
except KeyError as e:
|
49 |
-
|
|
|
|
|
50 |
continue
|
51 |
return indexes
|
52 |
|
@@ -75,17 +59,16 @@ class FilePathDataset(torch.utils.data.Dataset):
|
|
75 |
def __init__(self,
|
76 |
data_list,
|
77 |
root_path,
|
|
|
78 |
sr=24000,
|
79 |
data_augmentation=False,
|
80 |
-
validation=False
|
|
|
81 |
):
|
82 |
|
83 |
-
spect_params = SPECT_PARAMS
|
84 |
-
mel_params = MEL_PARAMS
|
85 |
-
|
86 |
_data_list = [l.strip().split('|') for l in data_list]
|
87 |
self.data_list = _data_list #[data if len(data) == 3 else (*data, 0) for data in _data_list] #append speakerid=0 for all
|
88 |
-
self.text_cleaner = TextCleaner()
|
89 |
self.sr = sr
|
90 |
|
91 |
self.df = pd.DataFrame(self.data_list)
|
@@ -195,9 +178,13 @@ class Collater(object):
|
|
195 |
return waves, texts, input_lengths, mels, output_lengths
|
196 |
|
197 |
|
|
|
|
|
|
|
198 |
|
199 |
def build_dataloader(path_list,
|
200 |
root_path,
|
|
|
201 |
validation=False,
|
202 |
batch_size=4,
|
203 |
num_workers=1,
|
@@ -205,14 +192,116 @@ def build_dataloader(path_list,
|
|
205 |
collate_config={},
|
206 |
dataset_config={}):
|
207 |
|
208 |
-
dataset = FilePathDataset(path_list, root_path, validation=validation, **dataset_config)
|
209 |
collate_fn = Collater(**collate_config)
|
210 |
-
|
211 |
-
|
212 |
-
|
213 |
-
|
214 |
-
|
215 |
-
|
216 |
-
|
217 |
-
|
218 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
#coding: utf-8
|
|
|
2 |
import os.path as osp
|
|
|
3 |
import random
|
4 |
import numpy as np
|
5 |
import random
|
|
|
7 |
import librosa
|
8 |
|
9 |
import torch
|
|
|
|
|
10 |
import torchaudio
|
11 |
+
import torch.utils.data
|
12 |
+
import torch.distributed as dist
|
13 |
+
from multiprocessing import Pool
|
14 |
|
15 |
import logging
|
16 |
logger = logging.getLogger(__name__)
|
|
|
18 |
|
19 |
import pandas as pd
|
20 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
class TextCleaner:
|
22 |
+
def __init__(self, symbol_dict, debug=True):
|
23 |
+
self.word_index_dictionary = symbol_dict
|
24 |
+
self.debug = debug
|
25 |
def __call__(self, text):
|
26 |
indexes = []
|
27 |
for char in text:
|
28 |
try:
|
29 |
indexes.append(self.word_index_dictionary[char])
|
30 |
except KeyError as e:
|
31 |
+
if self.debug:
|
32 |
+
print("\nWARNING UNKNOWN IPA CHARACTERS/LETTERS: ", char)
|
33 |
+
print("To ignore set 'debug' to false in the config")
|
34 |
continue
|
35 |
return indexes
|
36 |
|
|
|
59 |
def __init__(self,
|
60 |
data_list,
|
61 |
root_path,
|
62 |
+
symbol_dict,
|
63 |
sr=24000,
|
64 |
data_augmentation=False,
|
65 |
+
validation=False,
|
66 |
+
debug=True
|
67 |
):
|
68 |
|
|
|
|
|
|
|
69 |
_data_list = [l.strip().split('|') for l in data_list]
|
70 |
self.data_list = _data_list #[data if len(data) == 3 else (*data, 0) for data in _data_list] #append speakerid=0 for all
|
71 |
+
self.text_cleaner = TextCleaner(symbol_dict, debug)
|
72 |
self.sr = sr
|
73 |
|
74 |
self.df = pd.DataFrame(self.data_list)
|
|
|
178 |
return waves, texts, input_lengths, mels, output_lengths
|
179 |
|
180 |
|
181 |
+
def get_length(wave_path, root_path):
|
182 |
+
info = sf.info(osp.join(root_path, wave_path))
|
183 |
+
return info.frames * (24000 / info.samplerate)
|
184 |
|
185 |
def build_dataloader(path_list,
|
186 |
root_path,
|
187 |
+
symbol_dict,
|
188 |
validation=False,
|
189 |
batch_size=4,
|
190 |
num_workers=1,
|
|
|
192 |
collate_config={},
|
193 |
dataset_config={}):
|
194 |
|
195 |
+
dataset = FilePathDataset(path_list, root_path, symbol_dict, validation=validation, **dataset_config)
|
196 |
collate_fn = Collater(**collate_config)
|
197 |
+
|
198 |
+
print("Getting sample lengths...")
|
199 |
+
|
200 |
+
num_processes = num_workers * 2
|
201 |
+
if num_processes != 0:
|
202 |
+
list_of_tuples = [(d[0], root_path) for d in dataset.data_list]
|
203 |
+
with Pool(processes=num_processes) as pool:
|
204 |
+
sample_lengths = pool.starmap(get_length, list_of_tuples, chunksize=16)
|
205 |
+
else:
|
206 |
+
sample_lengths = []
|
207 |
+
for d in dataset.data_list:
|
208 |
+
sample_lengths.append(get_length(d[0], root_path))
|
209 |
+
|
210 |
+
data_loader = torch.utils.data.DataLoader(
|
211 |
+
dataset,
|
212 |
+
num_workers=num_workers,
|
213 |
+
batch_sampler=BatchSampler(
|
214 |
+
sample_lengths,
|
215 |
+
batch_size,
|
216 |
+
shuffle=(not validation),
|
217 |
+
drop_last=(not validation),
|
218 |
+
num_replicas=1,
|
219 |
+
rank=0,
|
220 |
+
),
|
221 |
+
collate_fn=collate_fn,
|
222 |
+
pin_memory=(device != "cpu"),
|
223 |
+
)
|
224 |
+
|
225 |
+
return data_loader
|
226 |
+
|
227 |
+
#https://github.com/duerig/StyleTTS2/
|
228 |
+
class BatchSampler(torch.utils.data.Sampler):
|
229 |
+
def __init__(
|
230 |
+
self,
|
231 |
+
sample_lengths,
|
232 |
+
batch_sizes,
|
233 |
+
num_replicas=None,
|
234 |
+
rank=None,
|
235 |
+
shuffle=True,
|
236 |
+
drop_last=False,
|
237 |
+
):
|
238 |
+
self.batch_sizes = batch_sizes
|
239 |
+
if num_replicas is None:
|
240 |
+
self.num_replicas = dist.get_world_size()
|
241 |
+
else:
|
242 |
+
self.num_replicas = num_replicas
|
243 |
+
if rank is None:
|
244 |
+
self.rank = dist.get_rank()
|
245 |
+
else:
|
246 |
+
self.rank = rank
|
247 |
+
self.shuffle = shuffle
|
248 |
+
self.drop_last = drop_last
|
249 |
+
|
250 |
+
self.time_bins = {}
|
251 |
+
self.epoch = 0
|
252 |
+
self.total_len = 0
|
253 |
+
self.last_bin = None
|
254 |
+
|
255 |
+
for i in range(len(sample_lengths)):
|
256 |
+
bin_num = self.get_time_bin(sample_lengths[i])
|
257 |
+
if bin_num != -1:
|
258 |
+
if bin_num not in self.time_bins:
|
259 |
+
self.time_bins[bin_num] = []
|
260 |
+
self.time_bins[bin_num].append(i)
|
261 |
+
|
262 |
+
for key in self.time_bins.keys():
|
263 |
+
val = self.time_bins[key]
|
264 |
+
total_batch = self.batch_sizes * num_replicas
|
265 |
+
self.total_len += len(val) // total_batch
|
266 |
+
if not self.drop_last and len(val) % total_batch != 0:
|
267 |
+
self.total_len += 1
|
268 |
+
|
269 |
+
def __iter__(self):
|
270 |
+
sampler_order = list(self.time_bins.keys())
|
271 |
+
sampler_indices = []
|
272 |
+
|
273 |
+
if self.shuffle:
|
274 |
+
sampler_indices = torch.randperm(len(sampler_order)).tolist()
|
275 |
+
else:
|
276 |
+
sampler_indices = list(range(len(sampler_order)))
|
277 |
+
|
278 |
+
for index in sampler_indices:
|
279 |
+
key = sampler_order[index]
|
280 |
+
current_bin = self.time_bins[key]
|
281 |
+
dist = torch.utils.data.distributed.DistributedSampler(
|
282 |
+
current_bin,
|
283 |
+
num_replicas=self.num_replicas,
|
284 |
+
rank=self.rank,
|
285 |
+
shuffle=self.shuffle,
|
286 |
+
drop_last=self.drop_last,
|
287 |
+
)
|
288 |
+
dist.set_epoch(self.epoch)
|
289 |
+
sampler = torch.utils.data.sampler.BatchSampler(
|
290 |
+
dist, self.batch_sizes, self.drop_last
|
291 |
+
)
|
292 |
+
for item_list in sampler:
|
293 |
+
self.last_bin = key
|
294 |
+
yield [current_bin[i] for i in item_list]
|
295 |
+
|
296 |
+
def __len__(self):
|
297 |
+
return self.total_len
|
298 |
+
|
299 |
+
def set_epoch(self, epoch):
|
300 |
+
self.epoch = epoch
|
301 |
+
|
302 |
+
def get_time_bin(self, sample_count):
|
303 |
+
result = -1
|
304 |
+
frames = sample_count // 300
|
305 |
+
if frames >= 20:
|
306 |
+
result = (frames - 20) // 20
|
307 |
+
return result
|