chenjgtea
commited on
Commit
•
5c0140c
1
Parent(s):
8dce793
新增gpu模式下chattts代码
Browse files- web/app_gpu.py +76 -51
web/app_gpu.py
CHANGED
@@ -1,6 +1,8 @@
|
|
1 |
import os, sys
|
2 |
import spaces
|
3 |
|
|
|
|
|
4 |
if sys.platform == "darwin":
|
5 |
os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
|
6 |
now_dir = os.getcwd()
|
@@ -158,21 +160,17 @@ def main(args):
|
|
158 |
|
159 |
# reload_chat_button.click()
|
160 |
|
161 |
-
generate_button.click(fn=
|
162 |
-
|
163 |
text_seed_input,
|
164 |
-
refine_text_checkBox
|
165 |
-
],
|
166 |
-
outputs=[text_output]
|
167 |
-
).then(fn=get_chat_infer_audio,
|
168 |
-
inputs=[text_output,
|
169 |
temperature_slider,
|
170 |
top_p_slider,
|
171 |
top_k_slider,
|
172 |
audio_seed_input,
|
173 |
spk_emb_text
|
174 |
],
|
175 |
-
outputs=[audio_output])
|
176 |
# 初始化 spk_emb_text 数值
|
177 |
spk_emb_text.value = on_audio_seed_change(audio_seed_input.value)
|
178 |
logger.info("元素初始化完成,启动gradio服务=======")
|
@@ -195,13 +193,16 @@ def main(args):
|
|
195 |
简而言之,"spk_embedding"关注的是对话参与者的身份特征,而"temperature"是用于调整生成文本不确定性的一个超参数。
|
196 |
'''
|
197 |
@spaces.GPU
|
198 |
-
def get_chat_infer_audio(
|
199 |
-
|
200 |
-
|
201 |
-
|
202 |
-
|
203 |
-
|
204 |
-
|
|
|
|
|
|
|
205 |
#音频参数设置
|
206 |
# params_infer_code = Chat2TTS.Chat.InferCodeParams(
|
207 |
# spk_emb=spk_emb_text, # add sampled speaker
|
@@ -209,45 +210,69 @@ def get_chat_infer_audio(chat_txt,
|
|
209 |
# top_P=top_p_slider, # top P decode
|
210 |
# top_K=top_k_slider, # top K decode
|
211 |
# )
|
212 |
-
|
213 |
-
# rand_spk = torch.randn(768)
|
214 |
-
params_infer_code = {
|
215 |
-
'spk_emb': None,
|
216 |
-
'temperature': temperature_slider,
|
217 |
-
'top_P': top_p_slider,
|
218 |
-
'top_K': top_k_slider,
|
219 |
-
}
|
220 |
-
torch.manual_seed(audio_seed_input)
|
221 |
-
wav = chat.infer(
|
222 |
-
text=chat_txt,
|
223 |
-
skip_refine_text=True, #跳过文本优化
|
224 |
-
params_infer_code=params_infer_code,
|
225 |
-
)
|
226 |
-
yield 24000, float_to_int16(wav[0]).T
|
227 |
-
|
228 |
-
@spaces.GPU
|
229 |
-
def get_chat_infer_text(text,seed,refine_text_checkBox):
|
230 |
-
|
231 |
-
logger.info("========开始优化文本内容2=====")
|
232 |
-
global chat
|
233 |
if not refine_text_checkBox:
|
234 |
logger.info("========文本内容无需优化=====")
|
235 |
-
|
236 |
-
|
237 |
-
|
238 |
-
|
239 |
-
|
240 |
-
|
241 |
-
|
242 |
-
|
243 |
-
|
244 |
-
|
245 |
-
|
246 |
-
refine_text_only=True, #仅返回优化后文本内容
|
247 |
-
params_refine_text=params_refine_text,
|
248 |
-
)
|
249 |
|
250 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
251 |
|
252 |
@spaces.GPU
|
253 |
def on_audio_seed_change(audio_seed_input):
|
|
|
1 |
import os, sys
|
2 |
import spaces
|
3 |
|
4 |
+
from tool import TorchSeedContext
|
5 |
+
|
6 |
if sys.platform == "darwin":
|
7 |
os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
|
8 |
now_dir = os.getcwd()
|
|
|
160 |
|
161 |
# reload_chat_button.click()
|
162 |
|
163 |
+
generate_button.click(fn=get_chat_infer_audio,
|
164 |
+
inputs=[text_input,
|
165 |
text_seed_input,
|
166 |
+
refine_text_checkBox,
|
|
|
|
|
|
|
|
|
167 |
temperature_slider,
|
168 |
top_p_slider,
|
169 |
top_k_slider,
|
170 |
audio_seed_input,
|
171 |
spk_emb_text
|
172 |
],
|
173 |
+
outputs=[text_output,audio_output])
|
174 |
# 初始化 spk_emb_text 数值
|
175 |
spk_emb_text.value = on_audio_seed_change(audio_seed_input.value)
|
176 |
logger.info("元素初始化完成,启动gradio服务=======")
|
|
|
193 |
简而言之,"spk_embedding"关注的是对话参与者的身份特征,而"temperature"是用于调整生成文本不确定性的一个超参数。
|
194 |
'''
|
195 |
@spaces.GPU
|
196 |
+
def get_chat_infer_audio(text,
|
197 |
+
text_seed_input,
|
198 |
+
refine_text_checkBox,
|
199 |
+
temperature_slider,
|
200 |
+
top_p_slider,
|
201 |
+
top_k_slider,
|
202 |
+
audio_seed_input,
|
203 |
+
spk_emb_text):
|
204 |
+
|
205 |
+
logger.info("========开始处理TTS模型=====")
|
206 |
#音频参数设置
|
207 |
# params_infer_code = Chat2TTS.Chat.InferCodeParams(
|
208 |
# spk_emb=spk_emb_text, # add sampled speaker
|
|
|
210 |
# top_P=top_p_slider, # top P decode
|
211 |
# top_K=top_k_slider, # top K decode
|
212 |
# )
|
213 |
+
params_refine_text = {'prompt': '[oral_2][laugh_0][break_6]'}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
214 |
if not refine_text_checkBox:
|
215 |
logger.info("========文本内容无需优化=====")
|
216 |
+
chat_txt=text
|
217 |
+
else:
|
218 |
+
logger.info("========开始优化文本内容=====")
|
219 |
+
#torch.manual_seed(text_seed_input)
|
220 |
+
with TorchSeedContext(text_seed_input):
|
221 |
+
chat_txt = chat.infer(
|
222 |
+
text=text,
|
223 |
+
skip_refine_text=False,
|
224 |
+
refine_text_only=True, #仅返回优化后文本内容
|
225 |
+
params_refine_text=params_refine_text,
|
226 |
+
)
|
|
|
|
|
|
|
227 |
|
228 |
+
logger.info("========开始生成音频文件=====")
|
229 |
+
#torch.manual_seed(audio_seed_input)
|
230 |
+
|
231 |
+
with TorchSeedContext(audio_seed_input):
|
232 |
+
rand_spk = torch.randn(768)
|
233 |
+
params_infer_code = {
|
234 |
+
'spk_emb': rand_spk,
|
235 |
+
'temperature': temperature_slider,
|
236 |
+
'top_P': top_p_slider,
|
237 |
+
'top_K': top_k_slider,
|
238 |
+
}
|
239 |
+
wav = chat.infer(
|
240 |
+
text=chat_txt,
|
241 |
+
skip_refine_text=True, #跳过文本优化
|
242 |
+
params_refine_text=params_refine_text,
|
243 |
+
params_infer_code=params_infer_code,
|
244 |
+
)
|
245 |
+
#yield 24000, float_to_int16(wav[0]).T
|
246 |
+
audio_data = np.array(wav[0]).flatten()
|
247 |
+
sample_rate = 24000
|
248 |
+
text_data = text[0] if isinstance(text, list) else text
|
249 |
+
|
250 |
+
return [text_data,(sample_rate, audio_data)]
|
251 |
+
|
252 |
+
|
253 |
+
# @spaces.GPU
|
254 |
+
# def get_chat_infer_text(text,seed,refine_text_checkBox):
|
255 |
+
#
|
256 |
+
# logger.info("========开始优化文本内容2=====")
|
257 |
+
# global chat
|
258 |
+
# if not refine_text_checkBox:
|
259 |
+
# logger.info("========文本内容无需优化=====")
|
260 |
+
# return text
|
261 |
+
#
|
262 |
+
# # params_refine_text = Chat2TTS.Chat.RefineTextParams(
|
263 |
+
# # prompt='[oral_2][laugh_0][break_6]',
|
264 |
+
# # )
|
265 |
+
#
|
266 |
+
# params_refine_text = {'prompt': '[oral_2][laugh_0][break_6]'}
|
267 |
+
# torch.manual_seed(seed)
|
268 |
+
# chat_text = chat.infer(
|
269 |
+
# text=text,
|
270 |
+
# skip_refine_text=False,
|
271 |
+
# refine_text_only=True, #仅返回优化后文本内容
|
272 |
+
# params_refine_text=params_refine_text,
|
273 |
+
# )
|
274 |
+
#
|
275 |
+
# return chat_text[0] if isinstance(chat_text, list) else chat_text
|
276 |
|
277 |
@spaces.GPU
|
278 |
def on_audio_seed_change(audio_seed_input):
|