add app
Browse files- app.py +309 -0
- attentions.py +329 -0
- commons.py +162 -0
- config.json +72 -0
- flow.py +120 -0
- models.py +489 -0
- modules.py +356 -0
- packages.txt +1 -0
- requirements.txt +3 -0
- vbx_phone_set.json +1 -0
app.py
ADDED
@@ -0,0 +1,309 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch # isort:skip
|
2 |
+
|
3 |
+
torch.manual_seed(42)
|
4 |
+
import json
|
5 |
+
import re
|
6 |
+
import unicodedata
|
7 |
+
from types import SimpleNamespace
|
8 |
+
|
9 |
+
import gradio as gr
|
10 |
+
import numpy as np
|
11 |
+
import regex
|
12 |
+
|
13 |
+
from models import DurationNet, SynthesizerTrn
|
14 |
+
|
15 |
+
title = "LightSpeed: Vietnamese Male Voice TTS"
|
16 |
+
description = "Vietnam Male Voice TTS."
|
17 |
+
config_file = "config.json"
|
18 |
+
duration_model_path = "vbx_duration_model.pth"
|
19 |
+
lightspeed_model_path = "gen_141k.pth"
|
20 |
+
phone_set_file = "vbx_phone_set.json"
|
21 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
22 |
+
with open(config_file, "rb") as f:
|
23 |
+
hps = json.load(f, object_hook=lambda x: SimpleNamespace(**x))
|
24 |
+
|
25 |
+
# load phone set json file
|
26 |
+
with open(phone_set_file, "r") as f:
|
27 |
+
phone_set = json.load(f)
|
28 |
+
|
29 |
+
assert phone_set[0][1:-1] == "SEP"
|
30 |
+
assert "sil" in phone_set
|
31 |
+
sil_idx = phone_set.index("sil")
|
32 |
+
|
33 |
+
vietnamese_characters = [
|
34 |
+
"a",
|
35 |
+
"à",
|
36 |
+
"á",
|
37 |
+
"ả",
|
38 |
+
"ã",
|
39 |
+
"ạ",
|
40 |
+
"ă",
|
41 |
+
"ằ",
|
42 |
+
"ắ",
|
43 |
+
"ẳ",
|
44 |
+
"ẵ",
|
45 |
+
"ặ",
|
46 |
+
"â",
|
47 |
+
"ầ",
|
48 |
+
"ấ",
|
49 |
+
"ẩ",
|
50 |
+
"ẫ",
|
51 |
+
"ậ",
|
52 |
+
"e",
|
53 |
+
"è",
|
54 |
+
"é",
|
55 |
+
"ẻ",
|
56 |
+
"ẽ",
|
57 |
+
"ẹ",
|
58 |
+
"ê",
|
59 |
+
"ề",
|
60 |
+
"ế",
|
61 |
+
"ể",
|
62 |
+
"ễ",
|
63 |
+
"ệ",
|
64 |
+
"i",
|
65 |
+
"ì",
|
66 |
+
"í",
|
67 |
+
"ỉ",
|
68 |
+
"ĩ",
|
69 |
+
"ị",
|
70 |
+
"o",
|
71 |
+
"ò",
|
72 |
+
"ó",
|
73 |
+
"ỏ",
|
74 |
+
"õ",
|
75 |
+
"ọ",
|
76 |
+
"ô",
|
77 |
+
"ồ",
|
78 |
+
"ố",
|
79 |
+
"ổ",
|
80 |
+
"ỗ",
|
81 |
+
"ộ",
|
82 |
+
"ơ",
|
83 |
+
"ờ",
|
84 |
+
"ớ",
|
85 |
+
"ở",
|
86 |
+
"ỡ",
|
87 |
+
"ợ",
|
88 |
+
"u",
|
89 |
+
"ù",
|
90 |
+
"ú",
|
91 |
+
"ủ",
|
92 |
+
"ũ",
|
93 |
+
"ụ",
|
94 |
+
"ư",
|
95 |
+
"ừ",
|
96 |
+
"ứ",
|
97 |
+
"ử",
|
98 |
+
"ữ",
|
99 |
+
"ự",
|
100 |
+
"y",
|
101 |
+
"ỳ",
|
102 |
+
"ý",
|
103 |
+
"ỷ",
|
104 |
+
"ỹ",
|
105 |
+
"ỵ",
|
106 |
+
"b",
|
107 |
+
"c",
|
108 |
+
"d",
|
109 |
+
"đ",
|
110 |
+
"g",
|
111 |
+
"h",
|
112 |
+
"k",
|
113 |
+
"l",
|
114 |
+
"m",
|
115 |
+
"n",
|
116 |
+
"p",
|
117 |
+
"q",
|
118 |
+
"r",
|
119 |
+
"s",
|
120 |
+
"t",
|
121 |
+
"v",
|
122 |
+
"x",
|
123 |
+
]
|
124 |
+
alphabet = "".join(vietnamese_characters)
|
125 |
+
space_re = regex.compile(r"\s+")
|
126 |
+
number_re = regex.compile("([0-9]+)")
|
127 |
+
digits = ["không", "một", "hai", "ba", "bốn", "năm", "sáu", "bảy", "tám", "chín"]
|
128 |
+
num_re = regex.compile(r"([0-9.,]*[0-9])")
|
129 |
+
keep_text_and_num_re = regex.compile(rf"[^\s{alphabet}.,0-9]")
|
130 |
+
keep_text_re = regex.compile(rf"[^\s{alphabet}]")
|
131 |
+
|
132 |
+
|
133 |
+
def read_number(num: str) -> str:
|
134 |
+
if len(num) == 1:
|
135 |
+
return digits[int(num)]
|
136 |
+
elif len(num) == 2 and num.isdigit():
|
137 |
+
n = int(num)
|
138 |
+
end = digits[n % 10]
|
139 |
+
if n == 10:
|
140 |
+
return "mười"
|
141 |
+
if n % 10 == 5:
|
142 |
+
end = "lăm"
|
143 |
+
if n % 10 == 0:
|
144 |
+
return digits[n // 10] + " mươi"
|
145 |
+
elif n < 20:
|
146 |
+
return "mười " + end
|
147 |
+
else:
|
148 |
+
if n % 10 == 1:
|
149 |
+
end = "mốt"
|
150 |
+
return digits[n // 10] + " mươi " + end
|
151 |
+
elif len(num) == 3 and num.isdigit():
|
152 |
+
n = int(num)
|
153 |
+
if n % 100 == 0:
|
154 |
+
return digits[n // 100] + " trăm"
|
155 |
+
elif num[1] == "0":
|
156 |
+
return digits[n // 100] + " trăm lẻ " + digits[n % 100]
|
157 |
+
else:
|
158 |
+
return digits[n // 100] + " trăm " + read_number(num[1:])
|
159 |
+
elif len(num) >= 4 and len(num) <= 6 and num.isdigit():
|
160 |
+
n = int(num)
|
161 |
+
n1 = n // 1000
|
162 |
+
return read_number(str(n1)) + " ngàn " + read_number(num[-3:])
|
163 |
+
elif "," in num:
|
164 |
+
n1, n2 = num.split(",")
|
165 |
+
return read_number(n1) + " phẩy " + read_number(n2)
|
166 |
+
elif "." in num:
|
167 |
+
parts = num.split(".")
|
168 |
+
if len(parts) == 2:
|
169 |
+
if parts[1] == "000":
|
170 |
+
return read_number(parts[0]) + " ngàn"
|
171 |
+
elif parts[1].startswith("00"):
|
172 |
+
end = digits[int(parts[1][2:])]
|
173 |
+
return read_number(parts[0]) + " ngàn lẻ " + end
|
174 |
+
else:
|
175 |
+
return read_number(parts[0]) + " ngàn " + read_number(parts[1])
|
176 |
+
elif len(parts) == 3:
|
177 |
+
return (
|
178 |
+
read_number(parts[0])
|
179 |
+
+ " triệu "
|
180 |
+
+ read_number(parts[1])
|
181 |
+
+ " ngàn "
|
182 |
+
+ read_number(parts[2])
|
183 |
+
)
|
184 |
+
return num
|
185 |
+
|
186 |
+
|
187 |
+
def text_to_phone_idx(text):
|
188 |
+
# lowercase
|
189 |
+
text = text.lower()
|
190 |
+
# unicode normalize
|
191 |
+
text = unicodedata.normalize("NFKC", text)
|
192 |
+
text = text.replace(".", " . ")
|
193 |
+
text = text.replace(",", " , ")
|
194 |
+
text = text.replace(";", " ; ")
|
195 |
+
text = text.replace(":", " : ")
|
196 |
+
text = text.replace("!", " ! ")
|
197 |
+
text = text.replace("?", " ? ")
|
198 |
+
text = text.replace("(", " ( ")
|
199 |
+
|
200 |
+
text = num_re.sub(r" \1 ", text)
|
201 |
+
words = text.split()
|
202 |
+
words = [read_number(w) if num_re.fullmatch(w) else w for w in words]
|
203 |
+
text = " ".join(words)
|
204 |
+
|
205 |
+
# remove redundant spaces
|
206 |
+
text = re.sub(r"\s+", " ", text)
|
207 |
+
# remove leading and trailing spaces
|
208 |
+
text = text.strip()
|
209 |
+
# convert words to phone indices
|
210 |
+
tokens = []
|
211 |
+
for c in text:
|
212 |
+
# if c is "," or ".", add <sil> phone
|
213 |
+
if c in ":,.!?;(":
|
214 |
+
tokens.append(sil_idx)
|
215 |
+
elif c in phone_set:
|
216 |
+
tokens.append(phone_set.index(c))
|
217 |
+
elif c == " ":
|
218 |
+
# add <sep> phone
|
219 |
+
tokens.append(0)
|
220 |
+
if tokens[0] != sil_idx:
|
221 |
+
# insert <sil> phone at the beginning
|
222 |
+
tokens = [sil_idx, 0] + tokens
|
223 |
+
if tokens[-1] != sil_idx:
|
224 |
+
tokens = tokens + [0, sil_idx]
|
225 |
+
return tokens
|
226 |
+
|
227 |
+
|
228 |
+
def text_to_speech(text):
|
229 |
+
# prevent too long text
|
230 |
+
if len(text) > 500:
|
231 |
+
text = text[:500]
|
232 |
+
|
233 |
+
phone_idx = text_to_phone_idx(text)
|
234 |
+
batch = {
|
235 |
+
"phone_idx": np.array([phone_idx]),
|
236 |
+
"phone_length": np.array([len(phone_idx)]),
|
237 |
+
}
|
238 |
+
|
239 |
+
# predict phoneme duration
|
240 |
+
duration_net = DurationNet(hps.data.vocab_size, 64, 4).to(device)
|
241 |
+
duration_net.load_state_dict(torch.load(duration_model_path, map_location=device))
|
242 |
+
duration_net = duration_net.eval()
|
243 |
+
phone_length = torch.from_numpy(batch["phone_length"].copy()).long().to(device)
|
244 |
+
phone_idx = torch.from_numpy(batch["phone_idx"].copy()).long().to(device)
|
245 |
+
with torch.inference_mode():
|
246 |
+
phone_duration = duration_net(phone_idx, phone_length)[:, :, 0] * 1000
|
247 |
+
phone_duration = torch.where(
|
248 |
+
phone_idx == sil_idx, torch.clamp_min(phone_duration, 200), phone_duration
|
249 |
+
)
|
250 |
+
phone_duration = torch.where(phone_idx == 0, 0, phone_duration)
|
251 |
+
|
252 |
+
generator = SynthesizerTrn(
|
253 |
+
hps.data.vocab_size,
|
254 |
+
hps.data.filter_length // 2 + 1,
|
255 |
+
hps.train.segment_size // hps.data.hop_length,
|
256 |
+
**vars(hps.model),
|
257 |
+
).to(device)
|
258 |
+
del generator.enc_q
|
259 |
+
ckpt = torch.load(lightspeed_model_path, map_location=device)
|
260 |
+
params = {}
|
261 |
+
for k, v in ckpt["net_g"].items():
|
262 |
+
k = k[7:] if k.startswith("module.") else k
|
263 |
+
params[k] = v
|
264 |
+
generator.load_state_dict(params, strict=False)
|
265 |
+
del ckpt, params
|
266 |
+
generator = generator.eval()
|
267 |
+
# mininum 1 frame for each phone
|
268 |
+
# phone_duration = torch.clamp_min(phone_duration, hps.data.hop_length * 1000 / hps.data.sampling_rate)
|
269 |
+
# phone_duration = torch.where(phone_idx == 0, 0, phone_duration)
|
270 |
+
end_time = torch.cumsum(phone_duration, dim=-1)
|
271 |
+
start_time = end_time - phone_duration
|
272 |
+
start_frame = start_time / 1000 * hps.data.sampling_rate / hps.data.hop_length
|
273 |
+
end_frame = end_time / 1000 * hps.data.sampling_rate / hps.data.hop_length
|
274 |
+
spec_length = end_frame.max(dim=-1).values
|
275 |
+
pos = torch.arange(0, spec_length.item(), device=device)
|
276 |
+
attn = torch.logical_and(
|
277 |
+
pos[None, :, None] >= start_frame[:, None, :],
|
278 |
+
pos[None, :, None] < end_frame[:, None, :],
|
279 |
+
).float()
|
280 |
+
with torch.inference_mode():
|
281 |
+
y_hat = generator.infer(
|
282 |
+
phone_idx, phone_length, spec_length, attn, max_len=None, noise_scale=0.667
|
283 |
+
)[0]
|
284 |
+
wave = y_hat[0, 0].data.cpu().numpy()
|
285 |
+
return (wave * (2**15)).astype(np.int16)
|
286 |
+
|
287 |
+
|
288 |
+
def speak(text):
|
289 |
+
y = text_to_speech(text)
|
290 |
+
return hps.data.sampling_rate, y
|
291 |
+
|
292 |
+
|
293 |
+
gr.Interface(
|
294 |
+
fn=speak,
|
295 |
+
inputs="text",
|
296 |
+
outputs="audio",
|
297 |
+
title=title,
|
298 |
+
examples=[
|
299 |
+
"Trăm năm trong cõi người ta, chữ tài chữ mệnh khéo là ghét nhau.",
|
300 |
+
"Đoạn trường tân thanh, thường được biết đến với cái tên đơn giản là Truyện Kiều, là một truyện thơ của đại thi hào Nguyễn Du",
|
301 |
+
"Lục Vân Tiên quê ở huyện Đông Thành, khôi ngô tuấn tú, tài kiêm văn võ. Nghe tin triều đình mở khoa thi, Vân Tiên từ giã thầy xuống núi đua tài.",
|
302 |
+
"Lê Quý Đôn, tên thuở nhỏ là Lê Danh Phương, là vị quan thời Lê trung hưng, cũng là nhà thơ và được mệnh danh là nhà bác học lớn của Việt Nam trong thời phong kiến",
|
303 |
+
"Tất cả mọi người đều sinh ra có quyền bình đẳng. Tạo hóa cho họ những quyền không ai có thể xâm phạm được; trong những quyền ấy, có quyền được sống, quyền tự do và quyền mưu cầu hạnh phúc.",
|
304 |
+
],
|
305 |
+
description=description,
|
306 |
+
theme="default",
|
307 |
+
allow_screenshot=False,
|
308 |
+
allow_flagging="never",
|
309 |
+
).launch(debug=False)
|
attentions.py
ADDED
@@ -0,0 +1,329 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
|
3 |
+
import torch
|
4 |
+
from torch import nn
|
5 |
+
from torch.nn import functional as F
|
6 |
+
|
7 |
+
import commons
|
8 |
+
from modules import LayerNorm
|
9 |
+
|
10 |
+
|
11 |
+
class Encoder(nn.Module):
|
12 |
+
def __init__(
|
13 |
+
self,
|
14 |
+
hidden_channels,
|
15 |
+
filter_channels,
|
16 |
+
n_heads,
|
17 |
+
n_layers,
|
18 |
+
kernel_size=1,
|
19 |
+
p_dropout=0.0,
|
20 |
+
window_size=4,
|
21 |
+
**kwargs
|
22 |
+
):
|
23 |
+
super().__init__()
|
24 |
+
self.hidden_channels = hidden_channels
|
25 |
+
self.filter_channels = filter_channels
|
26 |
+
self.n_heads = n_heads
|
27 |
+
self.n_layers = n_layers
|
28 |
+
self.kernel_size = kernel_size
|
29 |
+
self.p_dropout = p_dropout
|
30 |
+
self.window_size = window_size
|
31 |
+
|
32 |
+
self.drop = nn.Dropout(p_dropout)
|
33 |
+
self.attn_layers = nn.ModuleList()
|
34 |
+
self.norm_layers_1 = nn.ModuleList()
|
35 |
+
self.ffn_layers = nn.ModuleList()
|
36 |
+
self.norm_layers_2 = nn.ModuleList()
|
37 |
+
for i in range(self.n_layers):
|
38 |
+
self.attn_layers.append(
|
39 |
+
MultiHeadAttention(
|
40 |
+
hidden_channels,
|
41 |
+
hidden_channels,
|
42 |
+
n_heads,
|
43 |
+
p_dropout=p_dropout,
|
44 |
+
window_size=window_size,
|
45 |
+
)
|
46 |
+
)
|
47 |
+
self.norm_layers_1.append(LayerNorm(hidden_channels))
|
48 |
+
self.ffn_layers.append(
|
49 |
+
FFN(
|
50 |
+
hidden_channels,
|
51 |
+
hidden_channels,
|
52 |
+
filter_channels,
|
53 |
+
kernel_size,
|
54 |
+
p_dropout=p_dropout,
|
55 |
+
)
|
56 |
+
)
|
57 |
+
self.norm_layers_2.append(LayerNorm(hidden_channels))
|
58 |
+
|
59 |
+
def forward(self, x, x_mask):
|
60 |
+
attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
|
61 |
+
x = x * x_mask
|
62 |
+
for i in range(self.n_layers):
|
63 |
+
y = self.attn_layers[i](x, x, attn_mask)
|
64 |
+
y = self.drop(y)
|
65 |
+
x = self.norm_layers_1[i](x + y)
|
66 |
+
|
67 |
+
y = self.ffn_layers[i](x, x_mask)
|
68 |
+
y = self.drop(y)
|
69 |
+
x = self.norm_layers_2[i](x + y)
|
70 |
+
x = x * x_mask
|
71 |
+
return x
|
72 |
+
|
73 |
+
|
74 |
+
class MultiHeadAttention(nn.Module):
|
75 |
+
def __init__(
|
76 |
+
self,
|
77 |
+
channels,
|
78 |
+
out_channels,
|
79 |
+
n_heads,
|
80 |
+
p_dropout=0.0,
|
81 |
+
window_size=None,
|
82 |
+
heads_share=True,
|
83 |
+
block_length=None,
|
84 |
+
proximal_bias=False,
|
85 |
+
proximal_init=False,
|
86 |
+
):
|
87 |
+
super().__init__()
|
88 |
+
assert channels % n_heads == 0
|
89 |
+
|
90 |
+
self.channels = channels
|
91 |
+
self.out_channels = out_channels
|
92 |
+
self.n_heads = n_heads
|
93 |
+
self.p_dropout = p_dropout
|
94 |
+
self.window_size = window_size
|
95 |
+
self.heads_share = heads_share
|
96 |
+
self.block_length = block_length
|
97 |
+
self.proximal_bias = proximal_bias
|
98 |
+
self.proximal_init = proximal_init
|
99 |
+
# self.attn = None
|
100 |
+
|
101 |
+
self.k_channels = channels // n_heads
|
102 |
+
self.conv_q = nn.Conv1d(channels, channels, 1)
|
103 |
+
self.conv_k = nn.Conv1d(channels, channels, 1)
|
104 |
+
self.conv_v = nn.Conv1d(channels, channels, 1)
|
105 |
+
self.conv_o = nn.Conv1d(channels, out_channels, 1)
|
106 |
+
self.drop = nn.Dropout(p_dropout)
|
107 |
+
|
108 |
+
if window_size is not None:
|
109 |
+
n_heads_rel = 1 if heads_share else n_heads
|
110 |
+
rel_stddev = self.k_channels**-0.5
|
111 |
+
self.emb_rel_k = nn.Parameter(
|
112 |
+
torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
|
113 |
+
* rel_stddev
|
114 |
+
)
|
115 |
+
self.emb_rel_v = nn.Parameter(
|
116 |
+
torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
|
117 |
+
* rel_stddev
|
118 |
+
)
|
119 |
+
|
120 |
+
nn.init.xavier_uniform_(self.conv_q.weight)
|
121 |
+
nn.init.xavier_uniform_(self.conv_k.weight)
|
122 |
+
nn.init.xavier_uniform_(self.conv_v.weight)
|
123 |
+
if proximal_init:
|
124 |
+
with torch.no_grad():
|
125 |
+
self.conv_k.weight.copy_(self.conv_q.weight)
|
126 |
+
self.conv_k.bias.copy_(self.conv_q.bias)
|
127 |
+
|
128 |
+
def forward(self, x, c, attn_mask=None):
|
129 |
+
q = self.conv_q(x)
|
130 |
+
k = self.conv_k(c)
|
131 |
+
v = self.conv_v(c)
|
132 |
+
|
133 |
+
x, _ = self.attention(q, k, v, mask=attn_mask)
|
134 |
+
|
135 |
+
x = self.conv_o(x)
|
136 |
+
return x
|
137 |
+
|
138 |
+
def attention(self, query, key, value, mask=None):
|
139 |
+
# reshape [b, d, t] -> [b, n_h, t, d_k]
|
140 |
+
b, d, t_s, t_t = (*key.size(), query.size(2))
|
141 |
+
query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
|
142 |
+
key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
|
143 |
+
value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
|
144 |
+
|
145 |
+
scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
|
146 |
+
if self.window_size is not None:
|
147 |
+
assert (
|
148 |
+
t_s == t_t
|
149 |
+
), "Relative attention is only available for self-attention."
|
150 |
+
key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
|
151 |
+
rel_logits = self._matmul_with_relative_keys(
|
152 |
+
query / math.sqrt(self.k_channels), key_relative_embeddings
|
153 |
+
)
|
154 |
+
scores_local = self._relative_position_to_absolute_position(rel_logits)
|
155 |
+
scores = scores + scores_local
|
156 |
+
if self.proximal_bias:
|
157 |
+
assert t_s == t_t, "Proximal bias is only available for self-attention."
|
158 |
+
scores = scores + self._attention_bias_proximal(t_s).to(
|
159 |
+
device=scores.device, dtype=scores.dtype
|
160 |
+
)
|
161 |
+
if mask is not None:
|
162 |
+
scores = scores.masked_fill(mask == 0, -1e4)
|
163 |
+
if self.block_length is not None:
|
164 |
+
assert (
|
165 |
+
t_s == t_t
|
166 |
+
), "Local attention is only available for self-attention."
|
167 |
+
block_mask = (
|
168 |
+
torch.ones_like(scores)
|
169 |
+
.triu(-self.block_length)
|
170 |
+
.tril(self.block_length)
|
171 |
+
)
|
172 |
+
scores = scores.masked_fill(block_mask == 0, -1e4)
|
173 |
+
p_attn = F.softmax(scores, dim=-1) # [b, n_h, t_t, t_s]
|
174 |
+
p_attn = self.drop(p_attn)
|
175 |
+
output = torch.matmul(p_attn, value)
|
176 |
+
if self.window_size is not None:
|
177 |
+
relative_weights = self._absolute_position_to_relative_position(p_attn)
|
178 |
+
value_relative_embeddings = self._get_relative_embeddings(
|
179 |
+
self.emb_rel_v, t_s
|
180 |
+
)
|
181 |
+
output = output + self._matmul_with_relative_values(
|
182 |
+
relative_weights, value_relative_embeddings
|
183 |
+
)
|
184 |
+
output = (
|
185 |
+
output.transpose(2, 3).contiguous().view(b, d, t_t)
|
186 |
+
) # [b, n_h, t_t, d_k] -> [b, d, t_t]
|
187 |
+
return output, p_attn
|
188 |
+
|
189 |
+
def _matmul_with_relative_values(self, x, y):
|
190 |
+
"""
|
191 |
+
x: [b, h, l, m]
|
192 |
+
y: [h or 1, m, d]
|
193 |
+
ret: [b, h, l, d]
|
194 |
+
"""
|
195 |
+
ret = torch.matmul(x, y.unsqueeze(0))
|
196 |
+
return ret
|
197 |
+
|
198 |
+
def _matmul_with_relative_keys(self, x, y):
|
199 |
+
"""
|
200 |
+
x: [b, h, l, d]
|
201 |
+
y: [h or 1, m, d]
|
202 |
+
ret: [b, h, l, m]
|
203 |
+
"""
|
204 |
+
ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
|
205 |
+
return ret
|
206 |
+
|
207 |
+
def _get_relative_embeddings(self, relative_embeddings, length):
|
208 |
+
max_relative_position = 2 * self.window_size + 1
|
209 |
+
# Pad first before slice to avoid using cond ops.
|
210 |
+
pad_length = max(length - (self.window_size + 1), 0)
|
211 |
+
slice_start_position = max((self.window_size + 1) - length, 0)
|
212 |
+
slice_end_position = slice_start_position + 2 * length - 1
|
213 |
+
if pad_length > 0:
|
214 |
+
padded_relative_embeddings = F.pad(
|
215 |
+
relative_embeddings,
|
216 |
+
commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]),
|
217 |
+
)
|
218 |
+
else:
|
219 |
+
padded_relative_embeddings = relative_embeddings
|
220 |
+
used_relative_embeddings = padded_relative_embeddings[
|
221 |
+
:, slice_start_position:slice_end_position
|
222 |
+
]
|
223 |
+
return used_relative_embeddings
|
224 |
+
|
225 |
+
def _relative_position_to_absolute_position(self, x):
|
226 |
+
"""
|
227 |
+
x: [b, h, l, 2*l-1]
|
228 |
+
ret: [b, h, l, l]
|
229 |
+
"""
|
230 |
+
batch, heads, length, _ = x.size()
|
231 |
+
# Concat columns of pad to shift from relative to absolute indexing.
|
232 |
+
x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, 1]]))
|
233 |
+
|
234 |
+
# Concat extra elements so to add up to shape (len+1, 2*len-1).
|
235 |
+
x_flat = x.view([batch, heads, length * 2 * length])
|
236 |
+
x_flat = F.pad(
|
237 |
+
x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [0, length - 1]])
|
238 |
+
)
|
239 |
+
|
240 |
+
# Reshape and slice out the padded elements.
|
241 |
+
x_final = x_flat.view([batch, heads, length + 1, 2 * length - 1])[
|
242 |
+
:, :, :length, length - 1 :
|
243 |
+
]
|
244 |
+
return x_final
|
245 |
+
|
246 |
+
def _absolute_position_to_relative_position(self, x):
|
247 |
+
"""
|
248 |
+
x: [b, h, l, l]
|
249 |
+
ret: [b, h, l, 2*l-1]
|
250 |
+
"""
|
251 |
+
batch, heads, length, _ = x.shape
|
252 |
+
# padd along column
|
253 |
+
x = F.pad(
|
254 |
+
x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length - 1]])
|
255 |
+
)
|
256 |
+
x_flat = x.view([batch, heads, length * length + length * (length - 1)])
|
257 |
+
# add 0's in the beginning that will skew the elements after reshape
|
258 |
+
x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]]))
|
259 |
+
x_final = x_flat.view([batch, heads, length, 2 * length])[:, :, :, 1:]
|
260 |
+
return x_final
|
261 |
+
|
262 |
+
def _attention_bias_proximal(self, length):
|
263 |
+
"""Bias for self-attention to encourage attention to close positions.
|
264 |
+
Args:
|
265 |
+
length: an integer scalar.
|
266 |
+
Returns:
|
267 |
+
a Tensor with shape [1, 1, length, length]
|
268 |
+
"""
|
269 |
+
r = torch.arange(length, dtype=torch.float32)
|
270 |
+
diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
|
271 |
+
return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
|
272 |
+
|
273 |
+
|
274 |
+
class FFN(nn.Module):
|
275 |
+
def __init__(
|
276 |
+
self,
|
277 |
+
in_channels,
|
278 |
+
out_channels,
|
279 |
+
filter_channels,
|
280 |
+
kernel_size,
|
281 |
+
p_dropout=0.0,
|
282 |
+
activation=None,
|
283 |
+
causal=False,
|
284 |
+
):
|
285 |
+
super().__init__()
|
286 |
+
self.in_channels = in_channels
|
287 |
+
self.out_channels = out_channels
|
288 |
+
self.filter_channels = filter_channels
|
289 |
+
self.kernel_size = kernel_size
|
290 |
+
self.p_dropout = p_dropout
|
291 |
+
self.activation = activation
|
292 |
+
self.causal = causal
|
293 |
+
|
294 |
+
if causal:
|
295 |
+
self.padding = self._causal_padding
|
296 |
+
else:
|
297 |
+
self.padding = self._same_padding
|
298 |
+
|
299 |
+
self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)
|
300 |
+
self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)
|
301 |
+
self.drop = nn.Dropout(p_dropout)
|
302 |
+
|
303 |
+
def forward(self, x, x_mask):
|
304 |
+
x = self.conv_1(self.padding(x * x_mask))
|
305 |
+
if self.activation == "gelu":
|
306 |
+
x = x * torch.sigmoid(1.702 * x)
|
307 |
+
else:
|
308 |
+
x = torch.relu(x)
|
309 |
+
x = self.drop(x)
|
310 |
+
x = self.conv_2(self.padding(x * x_mask))
|
311 |
+
return x * x_mask
|
312 |
+
|
313 |
+
def _causal_padding(self, x):
|
314 |
+
if self.kernel_size == 1:
|
315 |
+
return x
|
316 |
+
pad_l = self.kernel_size - 1
|
317 |
+
pad_r = 0
|
318 |
+
padding = [[0, 0], [0, 0], [pad_l, pad_r]]
|
319 |
+
x = F.pad(x, commons.convert_pad_shape(padding))
|
320 |
+
return x
|
321 |
+
|
322 |
+
def _same_padding(self, x):
|
323 |
+
if self.kernel_size == 1:
|
324 |
+
return x
|
325 |
+
pad_l = (self.kernel_size - 1) // 2
|
326 |
+
pad_r = self.kernel_size // 2
|
327 |
+
padding = [[0, 0], [0, 0], [pad_l, pad_r]]
|
328 |
+
x = F.pad(x, commons.convert_pad_shape(padding))
|
329 |
+
return x
|
commons.py
ADDED
@@ -0,0 +1,162 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
|
3 |
+
import torch
|
4 |
+
from torch.nn import functional as F
|
5 |
+
|
6 |
+
|
7 |
+
def init_weights(m, mean=0.0, std=0.01):
|
8 |
+
classname = m.__class__.__name__
|
9 |
+
if classname.find("Conv") != -1:
|
10 |
+
m.weight.data.normal_(mean, std)
|
11 |
+
|
12 |
+
|
13 |
+
def get_padding(kernel_size, dilation=1):
|
14 |
+
return int((kernel_size * dilation - dilation) / 2)
|
15 |
+
|
16 |
+
|
17 |
+
def convert_pad_shape(pad_shape):
|
18 |
+
l = pad_shape[::-1]
|
19 |
+
pad_shape = [item for sublist in l for item in sublist]
|
20 |
+
return pad_shape
|
21 |
+
|
22 |
+
|
23 |
+
def intersperse(lst, item):
|
24 |
+
result = [item] * (len(lst) * 2 + 1)
|
25 |
+
result[1::2] = lst
|
26 |
+
return result
|
27 |
+
|
28 |
+
|
29 |
+
def kl_divergence(m_p, logs_p, m_q, logs_q):
|
30 |
+
"""KL(P||Q)"""
|
31 |
+
kl = (logs_q - logs_p) - 0.5
|
32 |
+
kl += (
|
33 |
+
0.5 * (torch.exp(2.0 * logs_p) + ((m_p - m_q) ** 2)) * torch.exp(-2.0 * logs_q)
|
34 |
+
)
|
35 |
+
return kl
|
36 |
+
|
37 |
+
|
38 |
+
def rand_gumbel(shape):
|
39 |
+
"""Sample from the Gumbel distribution, protect from overflows."""
|
40 |
+
uniform_samples = torch.rand(shape) * 0.99998 + 0.00001
|
41 |
+
return -torch.log(-torch.log(uniform_samples))
|
42 |
+
|
43 |
+
|
44 |
+
def rand_gumbel_like(x):
|
45 |
+
g = rand_gumbel(x.size()).to(dtype=x.dtype, device=x.device)
|
46 |
+
return g
|
47 |
+
|
48 |
+
|
49 |
+
def slice_segments(x, ids_str, segment_size=4):
|
50 |
+
ret = torch.zeros_like(x[:, :, :segment_size])
|
51 |
+
for i in range(x.size(0)):
|
52 |
+
idx_str = ids_str[i]
|
53 |
+
idx_end = idx_str + segment_size
|
54 |
+
ret[i] = x[i, :, idx_str:idx_end]
|
55 |
+
return ret
|
56 |
+
|
57 |
+
|
58 |
+
def rand_slice_segments(x, x_lengths=None, segment_size=4):
|
59 |
+
b, d, t = x.size()
|
60 |
+
if x_lengths is None:
|
61 |
+
x_lengths = t
|
62 |
+
ids_str_max = x_lengths - segment_size + 1
|
63 |
+
ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
|
64 |
+
ret = slice_segments(x, ids_str, segment_size)
|
65 |
+
return ret, ids_str
|
66 |
+
|
67 |
+
|
68 |
+
def get_timing_signal_1d(length, channels, min_timescale=1.0, max_timescale=1.0e4):
|
69 |
+
position = torch.arange(length, dtype=torch.float)
|
70 |
+
num_timescales = channels // 2
|
71 |
+
log_timescale_increment = math.log(float(max_timescale) / float(min_timescale)) / (
|
72 |
+
num_timescales - 1
|
73 |
+
)
|
74 |
+
inv_timescales = min_timescale * torch.exp(
|
75 |
+
torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment
|
76 |
+
)
|
77 |
+
scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)
|
78 |
+
signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0)
|
79 |
+
signal = F.pad(signal, [0, 0, 0, channels % 2])
|
80 |
+
signal = signal.view(1, channels, length)
|
81 |
+
return signal
|
82 |
+
|
83 |
+
|
84 |
+
def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4):
|
85 |
+
b, channels, length = x.size()
|
86 |
+
signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
|
87 |
+
return x + signal.to(dtype=x.dtype, device=x.device)
|
88 |
+
|
89 |
+
|
90 |
+
def cat_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4, axis=1):
|
91 |
+
b, channels, length = x.size()
|
92 |
+
signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
|
93 |
+
return torch.cat([x, signal.to(dtype=x.dtype, device=x.device)], axis)
|
94 |
+
|
95 |
+
|
96 |
+
def subsequent_mask(length):
|
97 |
+
mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)
|
98 |
+
return mask
|
99 |
+
|
100 |
+
|
101 |
+
@torch.jit.script
|
102 |
+
def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
|
103 |
+
n_channels_int = n_channels[0]
|
104 |
+
in_act = input_a + input_b
|
105 |
+
t_act = torch.tanh(in_act[:, :n_channels_int, :])
|
106 |
+
s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
|
107 |
+
acts = t_act * s_act
|
108 |
+
return acts
|
109 |
+
|
110 |
+
|
111 |
+
def convert_pad_shape(pad_shape):
|
112 |
+
l = pad_shape[::-1]
|
113 |
+
pad_shape = [item for sublist in l for item in sublist]
|
114 |
+
return pad_shape
|
115 |
+
|
116 |
+
|
117 |
+
def shift_1d(x):
|
118 |
+
x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]
|
119 |
+
return x
|
120 |
+
|
121 |
+
|
122 |
+
def sequence_mask(length, max_length=None):
|
123 |
+
if max_length is None:
|
124 |
+
max_length = length.max()
|
125 |
+
x = torch.arange(max_length, dtype=length.dtype, device=length.device)
|
126 |
+
return x.unsqueeze(0) < length.unsqueeze(1)
|
127 |
+
|
128 |
+
|
129 |
+
def generate_path(duration, mask):
|
130 |
+
"""
|
131 |
+
duration: [b, 1, t_x]
|
132 |
+
mask: [b, 1, t_y, t_x]
|
133 |
+
"""
|
134 |
+
device = duration.device
|
135 |
+
|
136 |
+
b, _, t_y, t_x = mask.shape
|
137 |
+
cum_duration = torch.cumsum(duration, -1)
|
138 |
+
|
139 |
+
cum_duration_flat = cum_duration.view(b * t_x)
|
140 |
+
path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype)
|
141 |
+
path = path.view(b, t_x, t_y)
|
142 |
+
path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]
|
143 |
+
path = path.unsqueeze(1).transpose(2, 3) * mask
|
144 |
+
return path
|
145 |
+
|
146 |
+
|
147 |
+
def clip_grad_value_(parameters, clip_value, norm_type=2):
|
148 |
+
if isinstance(parameters, torch.Tensor):
|
149 |
+
parameters = [parameters]
|
150 |
+
parameters = list(filter(lambda p: p.grad is not None, parameters))
|
151 |
+
norm_type = float(norm_type)
|
152 |
+
if clip_value is not None:
|
153 |
+
clip_value = float(clip_value)
|
154 |
+
|
155 |
+
total_norm = 0
|
156 |
+
for p in parameters:
|
157 |
+
param_norm = p.grad.data.norm(norm_type)
|
158 |
+
total_norm += param_norm.item() ** norm_type
|
159 |
+
if clip_value is not None:
|
160 |
+
p.grad.data.clamp_(min=-clip_value, max=clip_value)
|
161 |
+
total_norm = total_norm ** (1.0 / norm_type)
|
162 |
+
return total_norm
|
config.json
ADDED
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"train": {
|
3 |
+
"learning_rate": 2e-4,
|
4 |
+
"betas": [
|
5 |
+
0.8,
|
6 |
+
0.99
|
7 |
+
],
|
8 |
+
"eps": 1e-9,
|
9 |
+
"lr_decay": 0.999875,
|
10 |
+
"segment_size": 8192,
|
11 |
+
"c_mel": 45,
|
12 |
+
"c_kl": 1.0
|
13 |
+
},
|
14 |
+
"data": {
|
15 |
+
"vocab_size": 256,
|
16 |
+
"max_wav_value": 32768.0,
|
17 |
+
"sampling_rate": 16000,
|
18 |
+
"filter_length": 1024,
|
19 |
+
"hop_length": 256,
|
20 |
+
"win_length": 1024,
|
21 |
+
"n_mel_channels": 80,
|
22 |
+
"mel_fmin": 0.0,
|
23 |
+
"mel_fmax": null
|
24 |
+
},
|
25 |
+
"model": {
|
26 |
+
"inter_channels": 192,
|
27 |
+
"hidden_channels": 192,
|
28 |
+
"filter_channels": 768,
|
29 |
+
"n_heads": 2,
|
30 |
+
"n_layers": 6,
|
31 |
+
"kernel_size": 3,
|
32 |
+
"p_dropout": 0.1,
|
33 |
+
"resblock": "1",
|
34 |
+
"resblock_kernel_sizes": [
|
35 |
+
3,
|
36 |
+
7,
|
37 |
+
11
|
38 |
+
],
|
39 |
+
"resblock_dilation_sizes": [
|
40 |
+
[
|
41 |
+
1,
|
42 |
+
3,
|
43 |
+
5
|
44 |
+
],
|
45 |
+
[
|
46 |
+
1,
|
47 |
+
3,
|
48 |
+
5
|
49 |
+
],
|
50 |
+
[
|
51 |
+
1,
|
52 |
+
3,
|
53 |
+
5
|
54 |
+
]
|
55 |
+
],
|
56 |
+
"upsample_rates": [
|
57 |
+
8,
|
58 |
+
8,
|
59 |
+
2,
|
60 |
+
2
|
61 |
+
],
|
62 |
+
"upsample_initial_channel": 512,
|
63 |
+
"upsample_kernel_sizes": [
|
64 |
+
16,
|
65 |
+
16,
|
66 |
+
4,
|
67 |
+
4
|
68 |
+
],
|
69 |
+
"n_layers_q": 3,
|
70 |
+
"use_spectral_norm": false
|
71 |
+
}
|
72 |
+
}
|
flow.py
ADDED
@@ -0,0 +1,120 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from torch import nn
|
3 |
+
|
4 |
+
from modules import WN
|
5 |
+
|
6 |
+
DEFAULT_MIN_BIN_WIDTH = 1e-3
|
7 |
+
DEFAULT_MIN_BIN_HEIGHT = 1e-3
|
8 |
+
DEFAULT_MIN_DERIVATIVE = 1e-3
|
9 |
+
|
10 |
+
|
11 |
+
class ResidualCouplingLayer(nn.Module):
|
12 |
+
def __init__(
|
13 |
+
self,
|
14 |
+
channels,
|
15 |
+
hidden_channels,
|
16 |
+
kernel_size,
|
17 |
+
dilation_rate,
|
18 |
+
n_layers,
|
19 |
+
p_dropout=0,
|
20 |
+
gin_channels=0,
|
21 |
+
mean_only=False,
|
22 |
+
):
|
23 |
+
assert channels % 2 == 0, "channels should be divisible by 2"
|
24 |
+
super().__init__()
|
25 |
+
self.channels = channels
|
26 |
+
self.hidden_channels = hidden_channels
|
27 |
+
self.kernel_size = kernel_size
|
28 |
+
self.dilation_rate = dilation_rate
|
29 |
+
self.n_layers = n_layers
|
30 |
+
self.half_channels = channels // 2
|
31 |
+
self.mean_only = mean_only
|
32 |
+
|
33 |
+
self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
|
34 |
+
self.enc = WN(
|
35 |
+
hidden_channels,
|
36 |
+
kernel_size,
|
37 |
+
dilation_rate,
|
38 |
+
n_layers,
|
39 |
+
p_dropout=p_dropout,
|
40 |
+
gin_channels=gin_channels,
|
41 |
+
)
|
42 |
+
self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
|
43 |
+
self.post.weight.data.zero_()
|
44 |
+
self.post.bias.data.zero_()
|
45 |
+
|
46 |
+
def forward(self, x, x_mask, g=None, reverse=False):
|
47 |
+
x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
|
48 |
+
h = self.pre(x0) * x_mask
|
49 |
+
h = self.enc(h, x_mask, g=g)
|
50 |
+
stats = self.post(h) * x_mask
|
51 |
+
if not self.mean_only:
|
52 |
+
m, logs = torch.split(stats, [self.half_channels] * 2, 1)
|
53 |
+
else:
|
54 |
+
m = stats
|
55 |
+
logs = torch.zeros_like(m)
|
56 |
+
|
57 |
+
if not reverse:
|
58 |
+
x1 = m + x1 * torch.exp(logs) * x_mask
|
59 |
+
x = torch.cat([x0, x1], 1)
|
60 |
+
logdet = torch.sum(logs, [1, 2])
|
61 |
+
return x, logdet
|
62 |
+
else:
|
63 |
+
x1 = (x1 - m) * torch.exp(-logs) * x_mask
|
64 |
+
x = torch.cat([x0, x1], 1)
|
65 |
+
return x
|
66 |
+
|
67 |
+
|
68 |
+
class Flip(nn.Module):
|
69 |
+
def forward(self, x, *args, reverse=False, **kwargs):
|
70 |
+
x = torch.flip(x, [1])
|
71 |
+
if not reverse:
|
72 |
+
logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
|
73 |
+
return x, logdet
|
74 |
+
else:
|
75 |
+
return x
|
76 |
+
|
77 |
+
|
78 |
+
class ResidualCouplingBlock(nn.Module):
|
79 |
+
def __init__(
|
80 |
+
self,
|
81 |
+
channels,
|
82 |
+
hidden_channels,
|
83 |
+
kernel_size,
|
84 |
+
dilation_rate,
|
85 |
+
n_layers,
|
86 |
+
n_flows=4,
|
87 |
+
gin_channels=0,
|
88 |
+
):
|
89 |
+
super().__init__()
|
90 |
+
self.channels = channels
|
91 |
+
self.hidden_channels = hidden_channels
|
92 |
+
self.kernel_size = kernel_size
|
93 |
+
self.dilation_rate = dilation_rate
|
94 |
+
self.n_layers = n_layers
|
95 |
+
self.n_flows = n_flows
|
96 |
+
self.gin_channels = gin_channels
|
97 |
+
|
98 |
+
self.flows = nn.ModuleList()
|
99 |
+
for i in range(n_flows):
|
100 |
+
self.flows.append(
|
101 |
+
ResidualCouplingLayer(
|
102 |
+
channels,
|
103 |
+
hidden_channels,
|
104 |
+
kernel_size,
|
105 |
+
dilation_rate,
|
106 |
+
n_layers,
|
107 |
+
gin_channels=gin_channels,
|
108 |
+
mean_only=True,
|
109 |
+
)
|
110 |
+
)
|
111 |
+
self.flows.append(Flip())
|
112 |
+
|
113 |
+
def forward(self, x, x_mask, g=None, reverse=False):
|
114 |
+
if not reverse:
|
115 |
+
for flow in self.flows:
|
116 |
+
x, _ = flow(x, x_mask, g=g, reverse=reverse)
|
117 |
+
else:
|
118 |
+
for flow in reversed(self.flows):
|
119 |
+
x = flow(x, x_mask, g=g, reverse=reverse)
|
120 |
+
return x
|
models.py
ADDED
@@ -0,0 +1,489 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
|
3 |
+
import torch
|
4 |
+
from torch import nn
|
5 |
+
from torch.nn import Conv1d, Conv2d, ConvTranspose1d
|
6 |
+
from torch.nn import functional as F
|
7 |
+
from torch.nn.utils import remove_weight_norm, spectral_norm, weight_norm
|
8 |
+
from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence
|
9 |
+
|
10 |
+
import attentions
|
11 |
+
import commons
|
12 |
+
import modules
|
13 |
+
from commons import get_padding, init_weights
|
14 |
+
from flow import ResidualCouplingBlock
|
15 |
+
|
16 |
+
|
17 |
+
class PriorEncoder(nn.Module):
|
18 |
+
def __init__(
|
19 |
+
self,
|
20 |
+
n_vocab,
|
21 |
+
out_channels,
|
22 |
+
hidden_channels,
|
23 |
+
filter_channels,
|
24 |
+
n_heads,
|
25 |
+
n_layers,
|
26 |
+
kernel_size,
|
27 |
+
p_dropout,
|
28 |
+
):
|
29 |
+
super().__init__()
|
30 |
+
self.n_vocab = n_vocab
|
31 |
+
self.out_channels = out_channels
|
32 |
+
self.hidden_channels = hidden_channels
|
33 |
+
self.filter_channels = filter_channels
|
34 |
+
self.n_heads = n_heads
|
35 |
+
self.n_layers = n_layers
|
36 |
+
self.kernel_size = kernel_size
|
37 |
+
self.p_dropout = p_dropout
|
38 |
+
|
39 |
+
self.emb = nn.Embedding(n_vocab, hidden_channels)
|
40 |
+
nn.init.normal_(self.emb.weight, 0.0, hidden_channels**-0.5)
|
41 |
+
self.pre_attn_encoder = attentions.Encoder(
|
42 |
+
hidden_channels,
|
43 |
+
filter_channels,
|
44 |
+
n_heads,
|
45 |
+
n_layers // 2,
|
46 |
+
kernel_size,
|
47 |
+
p_dropout,
|
48 |
+
)
|
49 |
+
self.post_attn_encoder = attentions.Encoder(
|
50 |
+
hidden_channels,
|
51 |
+
filter_channels,
|
52 |
+
n_heads,
|
53 |
+
n_layers - n_layers // 2,
|
54 |
+
kernel_size,
|
55 |
+
p_dropout,
|
56 |
+
)
|
57 |
+
self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
|
58 |
+
|
59 |
+
def forward(self, x, x_lengths, y_lengths, attn):
|
60 |
+
x = self.emb(x) * math.sqrt(self.hidden_channels) # [b, t, h]
|
61 |
+
x = torch.transpose(x, 1, -1) # [b, h, t]
|
62 |
+
x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(
|
63 |
+
x.dtype
|
64 |
+
)
|
65 |
+
x = self.pre_attn_encoder(x * x_mask, x_mask)
|
66 |
+
y = torch.einsum("bht,blt->bhl", x, attn)
|
67 |
+
y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, y.size(2)), 1).to(
|
68 |
+
y.dtype
|
69 |
+
)
|
70 |
+
y = self.post_attn_encoder(y * y_mask, y_mask)
|
71 |
+
stats = self.proj(y) * y_mask
|
72 |
+
|
73 |
+
m, logs = torch.split(stats, self.out_channels, dim=1)
|
74 |
+
return y, m, logs, y_mask
|
75 |
+
|
76 |
+
|
77 |
+
class PosteriorEncoder(nn.Module):
|
78 |
+
def __init__(
|
79 |
+
self,
|
80 |
+
in_channels,
|
81 |
+
out_channels,
|
82 |
+
hidden_channels,
|
83 |
+
kernel_size,
|
84 |
+
dilation_rate,
|
85 |
+
n_layers,
|
86 |
+
gin_channels=0,
|
87 |
+
):
|
88 |
+
super().__init__()
|
89 |
+
self.in_channels = in_channels
|
90 |
+
self.out_channels = out_channels
|
91 |
+
self.hidden_channels = hidden_channels
|
92 |
+
self.kernel_size = kernel_size
|
93 |
+
self.dilation_rate = dilation_rate
|
94 |
+
self.n_layers = n_layers
|
95 |
+
self.gin_channels = gin_channels
|
96 |
+
|
97 |
+
self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
|
98 |
+
self.enc = modules.WN(
|
99 |
+
hidden_channels,
|
100 |
+
kernel_size,
|
101 |
+
dilation_rate,
|
102 |
+
n_layers,
|
103 |
+
gin_channels=gin_channels,
|
104 |
+
)
|
105 |
+
self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
|
106 |
+
|
107 |
+
def forward(self, x, x_lengths, g=None):
|
108 |
+
x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(
|
109 |
+
x.dtype
|
110 |
+
)
|
111 |
+
x = self.pre(x) * x_mask
|
112 |
+
x = self.enc(x, x_mask, g=g)
|
113 |
+
stats = self.proj(x) * x_mask
|
114 |
+
m, logs = torch.split(stats, self.out_channels, dim=1)
|
115 |
+
z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
|
116 |
+
return z, m, logs, x_mask
|
117 |
+
|
118 |
+
|
119 |
+
class Generator(torch.nn.Module):
|
120 |
+
def __init__(
|
121 |
+
self,
|
122 |
+
initial_channel,
|
123 |
+
resblock,
|
124 |
+
resblock_kernel_sizes,
|
125 |
+
resblock_dilation_sizes,
|
126 |
+
upsample_rates,
|
127 |
+
upsample_initial_channel,
|
128 |
+
upsample_kernel_sizes,
|
129 |
+
gin_channels=0,
|
130 |
+
):
|
131 |
+
super(Generator, self).__init__()
|
132 |
+
self.num_kernels = len(resblock_kernel_sizes)
|
133 |
+
self.num_upsamples = len(upsample_rates)
|
134 |
+
self.conv_pre = Conv1d(
|
135 |
+
initial_channel, upsample_initial_channel, 7, 1, padding=3
|
136 |
+
)
|
137 |
+
resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
|
138 |
+
|
139 |
+
self.ups = nn.ModuleList()
|
140 |
+
for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
|
141 |
+
self.ups.append(
|
142 |
+
weight_norm(
|
143 |
+
ConvTranspose1d(
|
144 |
+
upsample_initial_channel // (2**i),
|
145 |
+
upsample_initial_channel // (2 ** (i + 1)),
|
146 |
+
k,
|
147 |
+
u,
|
148 |
+
padding=(k - u) // 2,
|
149 |
+
)
|
150 |
+
)
|
151 |
+
)
|
152 |
+
|
153 |
+
self.resblocks = nn.ModuleList()
|
154 |
+
for i in range(len(self.ups)):
|
155 |
+
ch = upsample_initial_channel // (2 ** (i + 1))
|
156 |
+
for j, (k, d) in enumerate(
|
157 |
+
zip(resblock_kernel_sizes, resblock_dilation_sizes)
|
158 |
+
):
|
159 |
+
self.resblocks.append(resblock(ch, k, d))
|
160 |
+
|
161 |
+
self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
|
162 |
+
self.ups.apply(init_weights)
|
163 |
+
|
164 |
+
if gin_channels != 0:
|
165 |
+
self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
|
166 |
+
|
167 |
+
def forward(self, x, g=None):
|
168 |
+
x = self.conv_pre(x)
|
169 |
+
if g is not None:
|
170 |
+
x = x + self.cond(g)
|
171 |
+
|
172 |
+
for i in range(self.num_upsamples):
|
173 |
+
x = F.leaky_relu(x, modules.LRELU_SLOPE)
|
174 |
+
x = self.ups[i](x)
|
175 |
+
xs = None
|
176 |
+
for j in range(self.num_kernels):
|
177 |
+
if xs is None:
|
178 |
+
xs = self.resblocks[i * self.num_kernels + j](x)
|
179 |
+
else:
|
180 |
+
xs += self.resblocks[i * self.num_kernels + j](x)
|
181 |
+
x = xs / self.num_kernels
|
182 |
+
x = F.leaky_relu(x)
|
183 |
+
x = self.conv_post(x)
|
184 |
+
x = torch.tanh(x)
|
185 |
+
|
186 |
+
return x
|
187 |
+
|
188 |
+
def remove_weight_norm(self):
|
189 |
+
print("Removing weight norm...")
|
190 |
+
for l in self.ups:
|
191 |
+
remove_weight_norm(l)
|
192 |
+
for l in self.resblocks:
|
193 |
+
l.remove_weight_norm()
|
194 |
+
|
195 |
+
|
196 |
+
class DiscriminatorP(torch.nn.Module):
|
197 |
+
def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
|
198 |
+
super(DiscriminatorP, self).__init__()
|
199 |
+
self.period = period
|
200 |
+
self.use_spectral_norm = use_spectral_norm
|
201 |
+
norm_f = weight_norm if use_spectral_norm == False else spectral_norm
|
202 |
+
self.convs = nn.ModuleList(
|
203 |
+
[
|
204 |
+
norm_f(
|
205 |
+
Conv2d(
|
206 |
+
1,
|
207 |
+
32,
|
208 |
+
(kernel_size, 1),
|
209 |
+
(stride, 1),
|
210 |
+
padding=(get_padding(kernel_size, 1), 0),
|
211 |
+
)
|
212 |
+
),
|
213 |
+
norm_f(
|
214 |
+
Conv2d(
|
215 |
+
32,
|
216 |
+
128,
|
217 |
+
(kernel_size, 1),
|
218 |
+
(stride, 1),
|
219 |
+
padding=(get_padding(kernel_size, 1), 0),
|
220 |
+
)
|
221 |
+
),
|
222 |
+
norm_f(
|
223 |
+
Conv2d(
|
224 |
+
128,
|
225 |
+
512,
|
226 |
+
(kernel_size, 1),
|
227 |
+
(stride, 1),
|
228 |
+
padding=(get_padding(kernel_size, 1), 0),
|
229 |
+
)
|
230 |
+
),
|
231 |
+
norm_f(
|
232 |
+
Conv2d(
|
233 |
+
512,
|
234 |
+
1024,
|
235 |
+
(kernel_size, 1),
|
236 |
+
(stride, 1),
|
237 |
+
padding=(get_padding(kernel_size, 1), 0),
|
238 |
+
)
|
239 |
+
),
|
240 |
+
norm_f(
|
241 |
+
Conv2d(
|
242 |
+
1024,
|
243 |
+
1024,
|
244 |
+
(kernel_size, 1),
|
245 |
+
1,
|
246 |
+
padding=(get_padding(kernel_size, 1), 0),
|
247 |
+
)
|
248 |
+
),
|
249 |
+
]
|
250 |
+
)
|
251 |
+
self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
|
252 |
+
|
253 |
+
def forward(self, x):
|
254 |
+
fmap = []
|
255 |
+
|
256 |
+
# 1d to 2d
|
257 |
+
b, c, t = x.shape
|
258 |
+
if t % self.period != 0: # pad first
|
259 |
+
n_pad = self.period - (t % self.period)
|
260 |
+
x = F.pad(x, (0, n_pad), "reflect")
|
261 |
+
t = t + n_pad
|
262 |
+
x = x.view(b, c, t // self.period, self.period)
|
263 |
+
|
264 |
+
for l in self.convs:
|
265 |
+
x = l(x)
|
266 |
+
x = F.leaky_relu(x, modules.LRELU_SLOPE)
|
267 |
+
fmap.append(x)
|
268 |
+
x = self.conv_post(x)
|
269 |
+
fmap.append(x)
|
270 |
+
x = torch.flatten(x, 1, -1)
|
271 |
+
|
272 |
+
return x, fmap
|
273 |
+
|
274 |
+
|
275 |
+
class DiscriminatorS(torch.nn.Module):
|
276 |
+
def __init__(self, use_spectral_norm=False):
|
277 |
+
super(DiscriminatorS, self).__init__()
|
278 |
+
norm_f = weight_norm if use_spectral_norm == False else spectral_norm
|
279 |
+
self.convs = nn.ModuleList(
|
280 |
+
[
|
281 |
+
norm_f(Conv1d(1, 16, 15, 1, padding=7)),
|
282 |
+
norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
|
283 |
+
norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
|
284 |
+
norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
|
285 |
+
norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
|
286 |
+
norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
|
287 |
+
]
|
288 |
+
)
|
289 |
+
self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
|
290 |
+
|
291 |
+
def forward(self, x):
|
292 |
+
fmap = []
|
293 |
+
|
294 |
+
for l in self.convs:
|
295 |
+
x = l(x)
|
296 |
+
x = F.leaky_relu(x, modules.LRELU_SLOPE)
|
297 |
+
fmap.append(x)
|
298 |
+
x = self.conv_post(x)
|
299 |
+
fmap.append(x)
|
300 |
+
x = torch.flatten(x, 1, -1)
|
301 |
+
|
302 |
+
return x, fmap
|
303 |
+
|
304 |
+
|
305 |
+
class MultiPeriodDiscriminator(torch.nn.Module):
|
306 |
+
def __init__(self, use_spectral_norm=False):
|
307 |
+
super(MultiPeriodDiscriminator, self).__init__()
|
308 |
+
periods = [2, 3, 5, 7, 11]
|
309 |
+
|
310 |
+
discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
|
311 |
+
discs = discs + [
|
312 |
+
DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods
|
313 |
+
]
|
314 |
+
self.discriminators = nn.ModuleList(discs)
|
315 |
+
|
316 |
+
def forward(self, y, y_hat):
|
317 |
+
y_d_rs = []
|
318 |
+
y_d_gs = []
|
319 |
+
fmap_rs = []
|
320 |
+
fmap_gs = []
|
321 |
+
for i, d in enumerate(self.discriminators):
|
322 |
+
y_d_r, fmap_r = d(y)
|
323 |
+
y_d_g, fmap_g = d(y_hat)
|
324 |
+
y_d_rs.append(y_d_r)
|
325 |
+
y_d_gs.append(y_d_g)
|
326 |
+
fmap_rs.append(fmap_r)
|
327 |
+
fmap_gs.append(fmap_g)
|
328 |
+
|
329 |
+
return y_d_rs, y_d_gs, fmap_rs, fmap_gs
|
330 |
+
|
331 |
+
|
332 |
+
class SynthesizerTrn(nn.Module):
|
333 |
+
"""
|
334 |
+
Synthesizer for Training
|
335 |
+
"""
|
336 |
+
|
337 |
+
def __init__(
|
338 |
+
self,
|
339 |
+
n_vocab,
|
340 |
+
spec_channels,
|
341 |
+
segment_size,
|
342 |
+
inter_channels,
|
343 |
+
hidden_channels,
|
344 |
+
filter_channels,
|
345 |
+
n_heads,
|
346 |
+
n_layers,
|
347 |
+
kernel_size,
|
348 |
+
p_dropout,
|
349 |
+
resblock,
|
350 |
+
resblock_kernel_sizes,
|
351 |
+
resblock_dilation_sizes,
|
352 |
+
upsample_rates,
|
353 |
+
upsample_initial_channel,
|
354 |
+
upsample_kernel_sizes,
|
355 |
+
n_speakers=0,
|
356 |
+
gin_channels=0,
|
357 |
+
**kwargs
|
358 |
+
):
|
359 |
+
super().__init__()
|
360 |
+
self.n_vocab = n_vocab
|
361 |
+
self.spec_channels = spec_channels
|
362 |
+
self.inter_channels = inter_channels
|
363 |
+
self.hidden_channels = hidden_channels
|
364 |
+
self.filter_channels = filter_channels
|
365 |
+
self.n_heads = n_heads
|
366 |
+
self.n_layers = n_layers
|
367 |
+
self.kernel_size = kernel_size
|
368 |
+
self.p_dropout = p_dropout
|
369 |
+
self.resblock = resblock
|
370 |
+
self.resblock_kernel_sizes = resblock_kernel_sizes
|
371 |
+
self.resblock_dilation_sizes = resblock_dilation_sizes
|
372 |
+
self.upsample_rates = upsample_rates
|
373 |
+
self.upsample_initial_channel = upsample_initial_channel
|
374 |
+
self.upsample_kernel_sizes = upsample_kernel_sizes
|
375 |
+
self.segment_size = segment_size
|
376 |
+
self.n_speakers = n_speakers
|
377 |
+
self.gin_channels = gin_channels
|
378 |
+
|
379 |
+
self.enc_p = PriorEncoder(
|
380 |
+
n_vocab,
|
381 |
+
inter_channels,
|
382 |
+
hidden_channels,
|
383 |
+
filter_channels,
|
384 |
+
n_heads,
|
385 |
+
n_layers,
|
386 |
+
kernel_size,
|
387 |
+
p_dropout,
|
388 |
+
)
|
389 |
+
self.dec = Generator(
|
390 |
+
inter_channels,
|
391 |
+
resblock,
|
392 |
+
resblock_kernel_sizes,
|
393 |
+
resblock_dilation_sizes,
|
394 |
+
upsample_rates,
|
395 |
+
upsample_initial_channel,
|
396 |
+
upsample_kernel_sizes,
|
397 |
+
gin_channels=gin_channels,
|
398 |
+
)
|
399 |
+
self.enc_q = PosteriorEncoder(
|
400 |
+
spec_channels,
|
401 |
+
inter_channels,
|
402 |
+
hidden_channels,
|
403 |
+
5,
|
404 |
+
1,
|
405 |
+
16,
|
406 |
+
gin_channels=gin_channels,
|
407 |
+
)
|
408 |
+
self.flow = ResidualCouplingBlock(
|
409 |
+
inter_channels, hidden_channels, 5, 2, 4, gin_channels=gin_channels
|
410 |
+
)
|
411 |
+
|
412 |
+
if n_speakers > 1:
|
413 |
+
self.emb_g = nn.Embedding(n_speakers, gin_channels)
|
414 |
+
|
415 |
+
def forward(self, x, x_lengths, attn, y, y_lengths, sid=None):
|
416 |
+
x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths, y_lengths, attn=attn)
|
417 |
+
if self.n_speakers > 0:
|
418 |
+
g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
|
419 |
+
else:
|
420 |
+
g = None
|
421 |
+
|
422 |
+
z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
|
423 |
+
z_p = self.flow(z, y_mask, g=g)
|
424 |
+
|
425 |
+
z_slice, ids_slice = commons.rand_slice_segments(
|
426 |
+
z, y_lengths, self.segment_size
|
427 |
+
)
|
428 |
+
o = self.dec(z_slice, g=g)
|
429 |
+
l_length = None
|
430 |
+
return (
|
431 |
+
o,
|
432 |
+
l_length,
|
433 |
+
attn,
|
434 |
+
ids_slice,
|
435 |
+
x_mask,
|
436 |
+
y_mask,
|
437 |
+
(z, z_p, m_p, logs_p, m_q, logs_q),
|
438 |
+
)
|
439 |
+
|
440 |
+
def infer(
|
441 |
+
self,
|
442 |
+
x,
|
443 |
+
x_lengths,
|
444 |
+
y_lengths,
|
445 |
+
attn,
|
446 |
+
sid=None,
|
447 |
+
noise_scale=1,
|
448 |
+
max_len=None,
|
449 |
+
):
|
450 |
+
x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths, y_lengths, attn=attn)
|
451 |
+
if self.n_speakers > 0:
|
452 |
+
g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
|
453 |
+
else:
|
454 |
+
g = None
|
455 |
+
|
456 |
+
y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, attn.shape[1]), 1).to(
|
457 |
+
x_mask.dtype
|
458 |
+
)
|
459 |
+
z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
|
460 |
+
z = self.flow(z_p, y_mask, g=g, reverse=True)
|
461 |
+
o = self.dec((z * y_mask)[:, :, :max_len], g=g)
|
462 |
+
return o, attn, y_mask, (z, z_p, m_p, logs_p)
|
463 |
+
|
464 |
+
|
465 |
+
class DurationNet(torch.nn.Module):
|
466 |
+
def __init__(self, vocab_size: int, dim: int, num_layers=2):
|
467 |
+
super().__init__()
|
468 |
+
self.embed = torch.nn.Embedding(vocab_size, embedding_dim=dim)
|
469 |
+
self.rnn = torch.nn.GRU(
|
470 |
+
dim,
|
471 |
+
dim,
|
472 |
+
num_layers=num_layers,
|
473 |
+
batch_first=True,
|
474 |
+
bidirectional=True,
|
475 |
+
dropout=0.2,
|
476 |
+
)
|
477 |
+
self.proj = torch.nn.Linear(2 * dim, 1)
|
478 |
+
|
479 |
+
def forward(self, token, lengths):
|
480 |
+
x = self.embed(token)
|
481 |
+
lengths = lengths.long().cpu()
|
482 |
+
x = pack_padded_sequence(
|
483 |
+
x, lengths=lengths, batch_first=True, enforce_sorted=False
|
484 |
+
)
|
485 |
+
x, _ = self.rnn(x)
|
486 |
+
x, _ = pad_packed_sequence(x, batch_first=True, total_length=token.shape[1])
|
487 |
+
x = self.proj(x)
|
488 |
+
x = torch.nn.functional.softplus(x)
|
489 |
+
return x
|
modules.py
ADDED
@@ -0,0 +1,356 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from torch import nn
|
3 |
+
from torch.nn import Conv1d
|
4 |
+
from torch.nn import functional as F
|
5 |
+
from torch.nn.utils import remove_weight_norm, weight_norm
|
6 |
+
|
7 |
+
import commons
|
8 |
+
from commons import get_padding, init_weights
|
9 |
+
|
10 |
+
LRELU_SLOPE = 0.1
|
11 |
+
|
12 |
+
|
13 |
+
class LayerNorm(nn.Module):
|
14 |
+
def __init__(self, channels, eps=1e-5):
|
15 |
+
super().__init__()
|
16 |
+
self.channels = channels
|
17 |
+
self.eps = eps
|
18 |
+
|
19 |
+
self.gamma = nn.Parameter(torch.ones(channels))
|
20 |
+
self.beta = nn.Parameter(torch.zeros(channels))
|
21 |
+
|
22 |
+
def forward(self, x):
|
23 |
+
x = x.transpose(1, -1)
|
24 |
+
x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
|
25 |
+
return x.transpose(1, -1)
|
26 |
+
|
27 |
+
|
28 |
+
class ConvReluNorm(nn.Module):
|
29 |
+
def __init__(
|
30 |
+
self,
|
31 |
+
in_channels,
|
32 |
+
hidden_channels,
|
33 |
+
out_channels,
|
34 |
+
kernel_size,
|
35 |
+
n_layers,
|
36 |
+
p_dropout,
|
37 |
+
):
|
38 |
+
super().__init__()
|
39 |
+
self.in_channels = in_channels
|
40 |
+
self.hidden_channels = hidden_channels
|
41 |
+
self.out_channels = out_channels
|
42 |
+
self.kernel_size = kernel_size
|
43 |
+
self.n_layers = n_layers
|
44 |
+
self.p_dropout = p_dropout
|
45 |
+
assert n_layers > 1, "Number of layers should be larger than 0."
|
46 |
+
|
47 |
+
self.conv_layers = nn.ModuleList()
|
48 |
+
self.norm_layers = nn.ModuleList()
|
49 |
+
self.conv_layers.append(
|
50 |
+
nn.Conv1d(
|
51 |
+
in_channels, hidden_channels, kernel_size, padding=kernel_size // 2
|
52 |
+
)
|
53 |
+
)
|
54 |
+
self.norm_layers.append(LayerNorm(hidden_channels))
|
55 |
+
self.relu_drop = nn.Sequential(nn.ReLU(), nn.Dropout(p_dropout))
|
56 |
+
for _ in range(n_layers - 1):
|
57 |
+
self.conv_layers.append(
|
58 |
+
nn.Conv1d(
|
59 |
+
hidden_channels,
|
60 |
+
hidden_channels,
|
61 |
+
kernel_size,
|
62 |
+
padding=kernel_size // 2,
|
63 |
+
)
|
64 |
+
)
|
65 |
+
self.norm_layers.append(LayerNorm(hidden_channels))
|
66 |
+
self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
|
67 |
+
self.proj.weight.data.zero_()
|
68 |
+
self.proj.bias.data.zero_()
|
69 |
+
|
70 |
+
def forward(self, x, x_mask):
|
71 |
+
x_org = x
|
72 |
+
for i in range(self.n_layers):
|
73 |
+
x = self.conv_layers[i](x * x_mask)
|
74 |
+
x = self.norm_layers[i](x)
|
75 |
+
x = self.relu_drop(x)
|
76 |
+
x = x_org + self.proj(x)
|
77 |
+
return x * x_mask
|
78 |
+
|
79 |
+
|
80 |
+
class DDSConv(nn.Module):
|
81 |
+
"""
|
82 |
+
Dialted and Depth-Separable Convolution
|
83 |
+
"""
|
84 |
+
|
85 |
+
def __init__(self, channels, kernel_size, n_layers, p_dropout=0.0):
|
86 |
+
super().__init__()
|
87 |
+
self.channels = channels
|
88 |
+
self.kernel_size = kernel_size
|
89 |
+
self.n_layers = n_layers
|
90 |
+
self.p_dropout = p_dropout
|
91 |
+
|
92 |
+
self.drop = nn.Dropout(p_dropout)
|
93 |
+
self.convs_sep = nn.ModuleList()
|
94 |
+
self.convs_1x1 = nn.ModuleList()
|
95 |
+
self.norms_1 = nn.ModuleList()
|
96 |
+
self.norms_2 = nn.ModuleList()
|
97 |
+
for i in range(n_layers):
|
98 |
+
dilation = kernel_size**i
|
99 |
+
padding = (kernel_size * dilation - dilation) // 2
|
100 |
+
self.convs_sep.append(
|
101 |
+
nn.Conv1d(
|
102 |
+
channels,
|
103 |
+
channels,
|
104 |
+
kernel_size,
|
105 |
+
groups=channels,
|
106 |
+
dilation=dilation,
|
107 |
+
padding=padding,
|
108 |
+
)
|
109 |
+
)
|
110 |
+
self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
|
111 |
+
self.norms_1.append(LayerNorm(channels))
|
112 |
+
self.norms_2.append(LayerNorm(channels))
|
113 |
+
|
114 |
+
def forward(self, x, x_mask, g=None):
|
115 |
+
if g is not None:
|
116 |
+
x = x + g
|
117 |
+
for i in range(self.n_layers):
|
118 |
+
y = self.convs_sep[i](x * x_mask)
|
119 |
+
y = self.norms_1[i](y)
|
120 |
+
y = F.gelu(y)
|
121 |
+
y = self.convs_1x1[i](y)
|
122 |
+
y = self.norms_2[i](y)
|
123 |
+
y = F.gelu(y)
|
124 |
+
y = self.drop(y)
|
125 |
+
x = x + y
|
126 |
+
return x * x_mask
|
127 |
+
|
128 |
+
|
129 |
+
class WN(torch.nn.Module):
|
130 |
+
def __init__(
|
131 |
+
self,
|
132 |
+
hidden_channels,
|
133 |
+
kernel_size,
|
134 |
+
dilation_rate,
|
135 |
+
n_layers,
|
136 |
+
gin_channels=0,
|
137 |
+
p_dropout=0,
|
138 |
+
):
|
139 |
+
super(WN, self).__init__()
|
140 |
+
assert kernel_size % 2 == 1
|
141 |
+
self.hidden_channels = hidden_channels
|
142 |
+
self.kernel_size = (kernel_size,)
|
143 |
+
self.dilation_rate = dilation_rate
|
144 |
+
self.n_layers = n_layers
|
145 |
+
self.gin_channels = gin_channels
|
146 |
+
self.p_dropout = p_dropout
|
147 |
+
|
148 |
+
self.in_layers = torch.nn.ModuleList()
|
149 |
+
self.res_skip_layers = torch.nn.ModuleList()
|
150 |
+
self.drop = nn.Dropout(p_dropout)
|
151 |
+
|
152 |
+
if gin_channels != 0:
|
153 |
+
cond_layer = torch.nn.Conv1d(
|
154 |
+
gin_channels, 2 * hidden_channels * n_layers, 1
|
155 |
+
)
|
156 |
+
self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name="weight")
|
157 |
+
|
158 |
+
for i in range(n_layers):
|
159 |
+
dilation = dilation_rate**i
|
160 |
+
padding = int((kernel_size * dilation - dilation) / 2)
|
161 |
+
in_layer = torch.nn.Conv1d(
|
162 |
+
hidden_channels,
|
163 |
+
2 * hidden_channels,
|
164 |
+
kernel_size,
|
165 |
+
dilation=dilation,
|
166 |
+
padding=padding,
|
167 |
+
)
|
168 |
+
in_layer = torch.nn.utils.weight_norm(in_layer, name="weight")
|
169 |
+
self.in_layers.append(in_layer)
|
170 |
+
|
171 |
+
# last one is not necessary
|
172 |
+
if i < n_layers - 1:
|
173 |
+
res_skip_channels = 2 * hidden_channels
|
174 |
+
else:
|
175 |
+
res_skip_channels = hidden_channels
|
176 |
+
|
177 |
+
res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
|
178 |
+
res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name="weight")
|
179 |
+
self.res_skip_layers.append(res_skip_layer)
|
180 |
+
|
181 |
+
def forward(self, x, x_mask, g=None, **kwargs):
|
182 |
+
output = torch.zeros_like(x)
|
183 |
+
n_channels_tensor = torch.IntTensor([self.hidden_channels])
|
184 |
+
|
185 |
+
if g is not None:
|
186 |
+
g = self.cond_layer(g)
|
187 |
+
|
188 |
+
for i in range(self.n_layers):
|
189 |
+
x_in = self.in_layers[i](x)
|
190 |
+
if g is not None:
|
191 |
+
cond_offset = i * 2 * self.hidden_channels
|
192 |
+
g_l = g[:, cond_offset : cond_offset + 2 * self.hidden_channels, :]
|
193 |
+
else:
|
194 |
+
g_l = torch.zeros_like(x_in)
|
195 |
+
|
196 |
+
acts = commons.fused_add_tanh_sigmoid_multiply(x_in, g_l, n_channels_tensor)
|
197 |
+
acts = self.drop(acts)
|
198 |
+
|
199 |
+
res_skip_acts = self.res_skip_layers[i](acts)
|
200 |
+
if i < self.n_layers - 1:
|
201 |
+
res_acts = res_skip_acts[:, : self.hidden_channels, :]
|
202 |
+
x = (x + res_acts) * x_mask
|
203 |
+
output = output + res_skip_acts[:, self.hidden_channels :, :]
|
204 |
+
else:
|
205 |
+
output = output + res_skip_acts
|
206 |
+
return output * x_mask
|
207 |
+
|
208 |
+
def remove_weight_norm(self):
|
209 |
+
if self.gin_channels != 0:
|
210 |
+
torch.nn.utils.remove_weight_norm(self.cond_layer)
|
211 |
+
for l in self.in_layers:
|
212 |
+
torch.nn.utils.remove_weight_norm(l)
|
213 |
+
for l in self.res_skip_layers:
|
214 |
+
torch.nn.utils.remove_weight_norm(l)
|
215 |
+
|
216 |
+
|
217 |
+
class ResBlock1(torch.nn.Module):
|
218 |
+
def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
|
219 |
+
super(ResBlock1, self).__init__()
|
220 |
+
self.convs1 = nn.ModuleList(
|
221 |
+
[
|
222 |
+
weight_norm(
|
223 |
+
Conv1d(
|
224 |
+
channels,
|
225 |
+
channels,
|
226 |
+
kernel_size,
|
227 |
+
1,
|
228 |
+
dilation=dilation[0],
|
229 |
+
padding=get_padding(kernel_size, dilation[0]),
|
230 |
+
)
|
231 |
+
),
|
232 |
+
weight_norm(
|
233 |
+
Conv1d(
|
234 |
+
channels,
|
235 |
+
channels,
|
236 |
+
kernel_size,
|
237 |
+
1,
|
238 |
+
dilation=dilation[1],
|
239 |
+
padding=get_padding(kernel_size, dilation[1]),
|
240 |
+
)
|
241 |
+
),
|
242 |
+
weight_norm(
|
243 |
+
Conv1d(
|
244 |
+
channels,
|
245 |
+
channels,
|
246 |
+
kernel_size,
|
247 |
+
1,
|
248 |
+
dilation=dilation[2],
|
249 |
+
padding=get_padding(kernel_size, dilation[2]),
|
250 |
+
)
|
251 |
+
),
|
252 |
+
]
|
253 |
+
)
|
254 |
+
self.convs1.apply(init_weights)
|
255 |
+
|
256 |
+
self.convs2 = nn.ModuleList(
|
257 |
+
[
|
258 |
+
weight_norm(
|
259 |
+
Conv1d(
|
260 |
+
channels,
|
261 |
+
channels,
|
262 |
+
kernel_size,
|
263 |
+
1,
|
264 |
+
dilation=1,
|
265 |
+
padding=get_padding(kernel_size, 1),
|
266 |
+
)
|
267 |
+
),
|
268 |
+
weight_norm(
|
269 |
+
Conv1d(
|
270 |
+
channels,
|
271 |
+
channels,
|
272 |
+
kernel_size,
|
273 |
+
1,
|
274 |
+
dilation=1,
|
275 |
+
padding=get_padding(kernel_size, 1),
|
276 |
+
)
|
277 |
+
),
|
278 |
+
weight_norm(
|
279 |
+
Conv1d(
|
280 |
+
channels,
|
281 |
+
channels,
|
282 |
+
kernel_size,
|
283 |
+
1,
|
284 |
+
dilation=1,
|
285 |
+
padding=get_padding(kernel_size, 1),
|
286 |
+
)
|
287 |
+
),
|
288 |
+
]
|
289 |
+
)
|
290 |
+
self.convs2.apply(init_weights)
|
291 |
+
|
292 |
+
def forward(self, x, x_mask=None):
|
293 |
+
for c1, c2 in zip(self.convs1, self.convs2):
|
294 |
+
xt = F.leaky_relu(x, LRELU_SLOPE)
|
295 |
+
if x_mask is not None:
|
296 |
+
xt = xt * x_mask
|
297 |
+
xt = c1(xt)
|
298 |
+
xt = F.leaky_relu(xt, LRELU_SLOPE)
|
299 |
+
if x_mask is not None:
|
300 |
+
xt = xt * x_mask
|
301 |
+
xt = c2(xt)
|
302 |
+
x = xt + x
|
303 |
+
if x_mask is not None:
|
304 |
+
x = x * x_mask
|
305 |
+
return x
|
306 |
+
|
307 |
+
def remove_weight_norm(self):
|
308 |
+
for l in self.convs1:
|
309 |
+
remove_weight_norm(l)
|
310 |
+
for l in self.convs2:
|
311 |
+
remove_weight_norm(l)
|
312 |
+
|
313 |
+
|
314 |
+
class ResBlock2(torch.nn.Module):
|
315 |
+
def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
|
316 |
+
super(ResBlock2, self).__init__()
|
317 |
+
self.convs = nn.ModuleList(
|
318 |
+
[
|
319 |
+
weight_norm(
|
320 |
+
Conv1d(
|
321 |
+
channels,
|
322 |
+
channels,
|
323 |
+
kernel_size,
|
324 |
+
1,
|
325 |
+
dilation=dilation[0],
|
326 |
+
padding=get_padding(kernel_size, dilation[0]),
|
327 |
+
)
|
328 |
+
),
|
329 |
+
weight_norm(
|
330 |
+
Conv1d(
|
331 |
+
channels,
|
332 |
+
channels,
|
333 |
+
kernel_size,
|
334 |
+
1,
|
335 |
+
dilation=dilation[1],
|
336 |
+
padding=get_padding(kernel_size, dilation[1]),
|
337 |
+
)
|
338 |
+
),
|
339 |
+
]
|
340 |
+
)
|
341 |
+
self.convs.apply(init_weights)
|
342 |
+
|
343 |
+
def forward(self, x, x_mask=None):
|
344 |
+
for c in self.convs:
|
345 |
+
xt = F.leaky_relu(x, LRELU_SLOPE)
|
346 |
+
if x_mask is not None:
|
347 |
+
xt = xt * x_mask
|
348 |
+
xt = c(xt)
|
349 |
+
x = xt + x
|
350 |
+
if x_mask is not None:
|
351 |
+
x = x * x_mask
|
352 |
+
return x
|
353 |
+
|
354 |
+
def remove_weight_norm(self):
|
355 |
+
for l in self.convs:
|
356 |
+
remove_weight_norm(l)
|
packages.txt
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
libsndfile1-dev
|
requirements.txt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
gradio
|
2 |
+
regex
|
3 |
+
torch
|
vbx_phone_set.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
["<SEP>", "a", "b", "c", "d", "e", "g", "h", "i", "k", "l", "m", "n", "o", "p", "q", "r", "s", "sil", "t", "u", "v", "x", "y", "\u00e0", "\u00e1", "\u00e2", "\u00e3", "\u00e8", "\u00e9", "\u00ea", "\u00ec", "\u00ed", "\u00f2", "\u00f3", "\u00f4", "\u00f5", "\u00f9", "\u00fa", "\u00fd", "\u0103", "\u0111", "\u0129", "\u0169", "\u01a1", "\u01b0", "\u1ea1", "\u1ea3", "\u1ea5", "\u1ea7", "\u1ea9", "\u1eab", "\u1ead", "\u1eaf", "\u1eb1", "\u1eb3", "\u1eb5", "\u1eb7", "\u1eb9", "\u1ebb", "\u1ebd", "\u1ebf", "\u1ec1", "\u1ec3", "\u1ec5", "\u1ec7", "\u1ec9", "\u1ecb", "\u1ecd", "\u1ecf", "\u1ed1", "\u1ed3", "\u1ed5", "\u1ed7", "\u1ed9", "\u1edb", "\u1edd", "\u1edf", "\u1ee1", "\u1ee3", "\u1ee5", "\u1ee7", "\u1ee9", "\u1eeb", "\u1eed", "\u1eef", "\u1ef1", "\u1ef3", "\u1ef5", "\u1ef7", "\u1ef9"]
|