tuna2134/sbv2-v2
1
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity

Merge branch 'dev' of github.com:litagin02/Style-Bert-VITS2

Browse Source
This commit is contained in:
tsukumi
2024-10-25 22:16:19 +09:00
parent d397a58874 a8b8c760ff
commit d762bf4401
6 changed files with 309 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

215
gradio_tabs/inference.py
View File

@@ -1,6 +1,6 @@
import datetime
import json
from typing import Optional
from typing import Any, Optional, Union
import gradio as gr
@@ -186,6 +186,10 @@ style_md = f"""
- どのくらいに強さがいいかはモデルやスタイルによって異なるようです。
- 音声ファイルを入力する場合は、学習データと似た声音の話者（特に同じ性別）でないとよい効果が出ないかもしれません。
"""
voice_keys = ["dec"]
voice_pitch_keys = ["flow"]
speech_style_keys = ["enc_p"]
tempo_keys = ["sdp", "dp"]
def make_interactive():
@@ -203,6 +207,35 @@ def gr_util(item):
return (gr.update(visible=False), gr.update(visible=True))
null_models_frame = 0
def change_null_model_row(
null_model_index: int,
null_model_name: str,
null_model_path: str,
null_voice_weights: float,
null_voice_pitch_weights: float,
null_speech_style_weights: float,
null_tempo_weights: float,
null_models: dict[int, dict[str, Any]],
):
null_models[null_model_index] = {
"name": null_model_name,
"path": null_model_path,
"weight": null_voice_weights,
"pitch": null_voice_pitch_weights,
"style": null_speech_style_weights,
"tempo": null_tempo_weights,
}
if len(null_models) > null_models_frame:
keys_to_keep = list(range(null_models_frame))
result = {k: null_models[k] for k in keys_to_keep}
else:
result = null_models
return result, True
def create_inference_app(model_holder: TTSModelHolder) -> gr.Blocks:
def tts_fn(
model_name,
@@ -226,9 +259,12 @@ def create_inference_app(model_holder: TTSModelHolder) -> gr.Blocks:
speaker,
pitch_scale,
intonation_scale,
null_models: dict[int, dict[str, Union[str, float]]],
force_reload_model: bool,
):
model_holder.get_model(model_name, model_path)
assert model_holder.current_model is not None
logger.debug(f"Null models setting: {null_models}")
wrong_tone_message = ""
kata_tone: Optional[list[tuple[str, int]]] = None
@@ -283,6 +319,8 @@ def create_inference_app(model_holder: TTSModelHolder) -> gr.Blocks:
speaker_id=speaker_id,
pitch_scale=pitch_scale,
intonation_scale=intonation_scale,
null_model_params=null_models,
force_reload_model=force_reload_model,
)
except InvalidToneError as e:
logger.error(f"Tone error: {e}")
@@ -304,7 +342,10 @@ def create_inference_app(model_holder: TTSModelHolder) -> gr.Blocks:
message = f"Success, time: {duration} seconds."
if wrong_tone_message != "":
message = wrong_tone_message + "\n" + message
return message, (sr, audio), kata_tone_json_str
return message, (sr, audio), kata_tone_json_str, False
def get_model_files(model_name: str):
return [str(f) for f in model_holder.model_files_dict[model_name]]
model_names = model_holder.model_names
if len(model_names) == 0:
@@ -317,13 +358,13 @@ def create_inference_app(model_holder: TTSModelHolder) -> gr.Blocks:
)
return app
initial_id = 0
initial_pth_files = [
str(f) for f in model_holder.model_files_dict[model_names[initial_id]]
]
initial_pth_files = get_model_files(model_names[initial_id])
with gr.Blocks(theme=GRADIO_THEME) as app:
gr.Markdown(initial_md)
gr.Markdown(terms_of_use_md)
null_models = gr.State({})
force_reload_model = gr.State(False)
with gr.Accordion(label="使い方", open=False):
gr.Markdown(how_to_md)
with gr.Row():
@@ -437,6 +478,166 @@ def create_inference_app(model_holder: TTSModelHolder) -> gr.Blocks:
inputs=[use_assist_text],
outputs=[assist_text, assist_text_weight],
)
with gr.Accordion(label="ヌルモデル", open=False):
with gr.Row():
null_models_count = gr.Number(
label="ヌルモデルの数", value=0, step=1
)
with gr.Column(variant="panel"):
@gr.render(inputs=[null_models_count])
def render_null_models(
null_models_count: int,
):
global null_models_frame
null_models_frame = null_models_count
for i in range(null_models_count):
with gr.Row():
null_model_index = gr.Number(
value=i,
key=f"null_model_index_{i}",
visible=False,
)
null_model_name = gr.Dropdown(
label="モデル一覧",
choices=model_names,
key=f"null_model_name_{i}",
value=model_names[initial_id],
)
null_model_path = gr.Dropdown(
label="モデルファイル",
key=f"null_model_path_{i}",
# FIXME: 再レンダー時に選択肢が消えるのでどうにかしたい
# 現在は再レンダーでvalueは保存されるが選択肢は保存されないので選択肢が空になる
# そのときに選択肢にない値となるので、それを許す
allow_custom_value=True,
)
null_voice_weights = gr.Slider(
minimum=0,
maximum=1,
value=1,
step=0.1,
key=f"null_voice_weights_{i}",
label="声質",
)
null_voice_pitch_weights = gr.Slider(
minimum=0,
maximum=1,
value=1,
step=0.1,
key=f"null_voice_pitch_weights_{i}",
label="声の高さ",
)
null_speech_style_weights = gr.Slider(
minimum=0,
maximum=1,
value=1,
step=0.1,
key=f"null_speech_style_weights_{i}",
label="話し方",
)
null_tempo_weights = gr.Slider(
minimum=0,
maximum=1,
value=1,
step=0.1,
key=f"null_tempo_weights_{i}",
label="テンポ",
)
null_model_name.change(
model_holder.update_model_files_for_gradio,
inputs=[null_model_name],
outputs=[null_model_path],
)
null_model_path.change(
make_non_interactive, outputs=[tts_button]
)
# 愚直すぎるのでもう少しなんとかしたい
null_model_path.change(
change_null_model_row,
inputs=[
null_model_index,
null_model_name,
null_model_path,
null_voice_weights,
null_voice_pitch_weights,
null_speech_style_weights,
null_tempo_weights,
null_models,
],
outputs=[null_models, force_reload_model],
)
null_voice_weights.change(
change_null_model_row,
inputs=[
null_model_index,
null_model_name,
null_model_path,
null_voice_weights,
null_voice_pitch_weights,
null_speech_style_weights,
null_tempo_weights,
null_models,
],
outputs=[null_models, force_reload_model],
)
null_voice_pitch_weights.change(
change_null_model_row,
inputs=[
null_model_index,
null_model_name,
null_model_path,
null_voice_weights,
null_voice_pitch_weights,
null_speech_style_weights,
null_tempo_weights,
null_models,
],
outputs=[null_models, force_reload_model],
)
null_speech_style_weights.change(
change_null_model_row,
inputs=[
null_model_index,
null_model_name,
null_model_path,
null_voice_weights,
null_voice_pitch_weights,
null_speech_style_weights,
null_tempo_weights,
null_models,
],
outputs=[null_models, force_reload_model],
)
null_tempo_weights.change(
change_null_model_row,
inputs=[
null_model_index,
null_model_name,
null_model_path,
null_voice_weights,
null_voice_pitch_weights,
null_speech_style_weights,
null_tempo_weights,
null_models,
],
outputs=[null_models, force_reload_model],
)
add_btn = gr.Button("ヌルモデルを増やす")
del_btn = gr.Button("ヌルモデルを減らす")
add_btn.click(
lambda x: x + 1,
inputs=[null_models_count],
outputs=[null_models_count],
)
del_btn.click(
lambda x: x - 1 if x > 0 else 0,
inputs=[null_models_count],
outputs=[null_models_count],
)
with gr.Column():
with gr.Accordion("スタイルについて詳細", open=False):
gr.Markdown(style_md)
@@ -494,8 +695,10 @@ def create_inference_app(model_holder: TTSModelHolder) -> gr.Blocks:
speaker,
pitch_scale,
intonation_scale,
null_models,
force_reload_model,
],
outputs=[text_output, audio_output, tone],
outputs=[text_output, audio_output, tone, force_reload_model],
)
model_name.change(

16
gradio_tabs/merge.py
View File

@@ -105,6 +105,8 @@ def merge_style_usual(
new_config = config_a.copy()
new_config["data"]["num_styles"] = len(new_style2id)
new_config["data"]["style2id"] = new_style2id
if new_config["data"]["n_speakers"] == 1:
new_config["data"]["spk2id"] = { output_name : 0}
new_config["model_name"] = output_name
save_config(new_config, output_name)
@@ -160,6 +162,8 @@ def merge_style_add_diff(
new_config = config_a.copy()
new_config["data"]["num_styles"] = len(new_style2id)
new_config["data"]["style2id"] = new_style2id
if new_config["data"]["n_speakers"] == 1:
new_config["data"]["spk2id"] = { output_name : 0}
new_config["model_name"] = output_name
save_config(new_config, output_name)
@@ -219,6 +223,8 @@ def merge_style_weighted_sum(
new_config = config_a.copy()
new_config["data"]["num_styles"] = len(new_style2id)
new_config["data"]["style2id"] = new_style2id
if new_config["data"]["n_speakers"] == 1:
new_config["data"]["spk2id"] = { output_name : 0}
new_config["model_name"] = output_name
save_config(new_config, output_name)
@@ -268,6 +274,8 @@ def merge_style_add_null(
new_config = config_a.copy()
new_config["data"]["num_styles"] = len(new_style2id)
new_config["data"]["style2id"] = new_style2id
if new_config["data"]["n_speakers"] == 1:
new_config["data"]["spk2id"] = { output_name : 0}
new_config["model_name"] = output_name
save_config(new_config, output_name)
@@ -362,6 +370,8 @@ def merge_models_usual(
new_config["model_name"] = output_name
new_config["data"]["num_styles"] = 1
new_config["data"]["style2id"] = {DEFAULT_STYLE: 0}
if new_config["data"]["n_speakers"] == 1:
new_config["data"]["spk2id"] = { output_name : 0}
save_config(new_config, output_name)
neutral_vector_a = style_vectors_a[0]
@@ -444,6 +454,8 @@ def merge_models_add_diff(
new_config["model_name"] = output_name
new_config["data"]["num_styles"] = 1
new_config["data"]["style2id"] = {DEFAULT_STYLE: 0}
if new_config["data"]["n_speakers"] == 1:
new_config["data"]["spk2id"] = { output_name : 0}
with open(assets_root / output_name / "config.json", "w", encoding="utf-8") as f:
json.dump(new_config, f, indent=2, ensure_ascii=False)
@@ -519,6 +531,8 @@ def merge_models_weighted_sum(
new_config["model_name"] = output_name
new_config["data"]["num_styles"] = 1
new_config["data"]["style2id"] = {DEFAULT_STYLE: 0}
if new_config["data"]["n_speakers"] == 1:
new_config["data"]["spk2id"] = { output_name : 0}
with open(assets_root / output_name / "config.json", "w", encoding="utf-8") as f:
json.dump(new_config, f, indent=2, ensure_ascii=False)
@@ -595,6 +609,8 @@ def merge_models_add_null(
new_config["model_name"] = output_name
new_config["data"]["num_styles"] = 1
new_config["data"]["style2id"] = {DEFAULT_STYLE: 0}
if new_config["data"]["n_speakers"] == 1:
new_config["data"]["spk2id"] = { output_name : 0}
with open(assets_root / output_name / "config.json", "w", encoding="utf-8") as f:
json.dump(new_config, f, indent=2, ensure_ascii=False)