Introduce Docker and Windows CI Workflow, Pre-commit Formatting, and Language Resource Auto-Download (#2351)
* Docker Auto-Build Workflow * Rename * Update * Fix Bugs * Disable Progress Bar When workflows triggered * Fix Wget * Fix Bugs * Fix Bugs * Update Wget * Update Workflows * Accelerate Docker Image Building * Fix Install.sh * Add Skip-Check For Action Runner * Fix Dockerfile * . * . * . * . * Delete File in Runner * Add Sort * Delete More Files * Delete More * . * . * . * Add Pre-Commit Hook Update Docker * Add Code Spell Check * [pre-commit.ci] trigger * [pre-commit.ci] trigger * [pre-commit.ci] trigger * Fix Bugs * . * Disable Progress Bar and Logs while using GitHub Actions * . * . * Fix Bugs * update conda * fix bugs * Fix Bugs * fix bugs * . * . * Quiet Installation * fix bugs * . * fix bug * . * Fix pre-commit.ci and Docker * fix bugs * . * Update Docker & Pre-Commit * fix bugs * Update Req * Update Req * Update OpenCC * update precommit * . * Update .pre-commit-config.yaml * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Update Docs and fix bugs * Fix \ * Fix MacOS * . * test * . * Add Tag Alias * . * fix bugs * fix bugs * make image smaller * update pre-commit config * . * . * fix bugs * use miniconda * Fix Wrong Path * . * debug * debug * revert * Fix Bugs * Update Docs, Add Dict Auto Download in install.sh * update docker_build * Update Docs for Install.sh * update docker docs about architecture * Add Xcode-Commandline-Tool Installation * Update Docs 1. Add Missing VC17 2. Modufied the Order of FFmpeg Installation and Requirements Installation 3. Remove Duplicate FFmpeg * Fix Wrong Cuda Version * Update TESTED ENV * Add PYTHONNOUSERSITE(-s) * Fix Wrapper * Update install.sh For Robustness * Ignore .git * Preload CUDNN For Ctranslate2 * Remove Gradio Warnings * Update Colab * Fix OpenCC Problems * Update Win DLL Strategy * Fix Onnxruntime-gpu NVRTC Error * Fix Path Problems * Add Windows Packages Workflow * WIP * WIP * WIP * WIP * WIP * WIP * . * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * WIP * Fix Path * Fix Path * Enable Logging * Set 7-Zip compression level to maximum (-mx=9) * Use Multithread in ONNX Session * Fix Tag Bugs * Add Time * Add Time * Add Time * Compress More * Copy DLL to Solve VC Runtime DLL Missing Issues * Expose FFmpeg Errors, Copy Only Part of Visual C++ Runtime * Update build_windows_packages.ps1 * Update build_windows_packages.ps1 * Update build_windows_packages.ps1 * Update build_windows_packages.ps1 * WIP * WIP * WIP * Update build_windows_packages.ps1 * Update install.sh * Update build_windows_packages.ps1 * Update docker-publish.yaml * Update install.sh * Update Dockerfile * Update docker_build.sh * Update miniconda_install.sh * Update README.md * Update README.md * Update README.md * Update README.md * Update README.md * Update README.md * Update Colab-WebUI.ipynb * Update Colab-Inference.ipynb * Update docker-compose.yaml * 更新 build_windows_packages.ps1 * Update install.sh --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
This commit is contained in:
XXXXRT666
@@ -108,7 +108,7 @@ resample_transform_dict = {}
|
||||
|
||||
def resample(audio_tensor, sr0, sr1, device):
|
||||
global resample_transform_dict
|
||||
key="%s-%s"%(sr0,sr1)
|
||||
key = "%s-%s" % (sr0, sr1)
|
||||
if key not in resample_transform_dict:
|
||||
resample_transform_dict[key] = torchaudio.transforms.Resample(sr0, sr1).to(device)
|
||||
return resample_transform_dict[key](audio_tensor)
|
||||
@@ -252,7 +252,6 @@ class TTS_Config:
|
||||
"cnhuhbert_base_path": "GPT_SoVITS/pretrained_models/chinese-hubert-base",
|
||||
"bert_base_path": "GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large",
|
||||
},
|
||||
|
||||
}
|
||||
configs: dict = None
|
||||
v1_languages: list = ["auto", "en", "zh", "ja", "all_zh", "all_ja"]
|
||||
@@ -432,7 +431,6 @@ class TTS:
|
||||
"aux_ref_audio_paths": [],
|
||||
}
|
||||
|
||||
|
||||
self.stop_flag: bool = False
|
||||
self.precision: torch.dtype = torch.float16 if self.configs.is_half else torch.float32
|
||||
|
||||
@@ -468,7 +466,7 @@ class TTS:
|
||||
path_sovits = self.configs.default_configs[model_version]["vits_weights_path"]
|
||||
|
||||
if if_lora_v3 == True and os.path.exists(path_sovits) == False:
|
||||
info = path_sovits + i18n("SoVITS %s 底模缺失，无法加载相应 LoRA 权重"%model_version)
|
||||
info = path_sovits + i18n("SoVITS %s 底模缺失，无法加载相应 LoRA 权重" % model_version)
|
||||
raise FileExistsError(info)
|
||||
|
||||
# dict_s2 = torch.load(weights_path, map_location=self.configs.device,weights_only=False)
|
||||
@@ -507,7 +505,7 @@ class TTS:
|
||||
)
|
||||
self.configs.use_vocoder = False
|
||||
else:
|
||||
kwargs["version"]=model_version
|
||||
kwargs["version"] = model_version
|
||||
vits_model = SynthesizerTrnV3(
|
||||
self.configs.filter_length // 2 + 1,
|
||||
self.configs.segment_size // self.configs.hop_length,
|
||||
@@ -572,7 +570,7 @@ class TTS:
|
||||
self.vocoder.cpu()
|
||||
del self.vocoder
|
||||
self.empty_cache()
|
||||
|
||||
|
||||
self.vocoder = BigVGAN.from_pretrained(
|
||||
"%s/GPT_SoVITS/pretrained_models/models--nvidia--bigvgan_v2_24khz_100band_256x" % (now_dir,),
|
||||
use_cuda_kernel=False,
|
||||
@@ -595,18 +593,21 @@ class TTS:
|
||||
self.empty_cache()
|
||||
|
||||
self.vocoder = Generator(
|
||||
initial_channel=100,
|
||||
resblock="1",
|
||||
resblock_kernel_sizes=[3, 7, 11],
|
||||
resblock_dilation_sizes=[[1, 3, 5], [1, 3, 5], [1, 3, 5]],
|
||||
upsample_rates=[10, 6, 2, 2, 2],
|
||||
upsample_initial_channel=512,
|
||||
upsample_kernel_sizes=[20, 12, 4, 4, 4],
|
||||
gin_channels=0, is_bias=True
|
||||
)
|
||||
initial_channel=100,
|
||||
resblock="1",
|
||||
resblock_kernel_sizes=[3, 7, 11],
|
||||
resblock_dilation_sizes=[[1, 3, 5], [1, 3, 5], [1, 3, 5]],
|
||||
upsample_rates=[10, 6, 2, 2, 2],
|
||||
upsample_initial_channel=512,
|
||||
upsample_kernel_sizes=[20, 12, 4, 4, 4],
|
||||
gin_channels=0,
|
||||
is_bias=True,
|
||||
)
|
||||
self.vocoder.remove_weight_norm()
|
||||
state_dict_g = torch.load("%s/GPT_SoVITS/pretrained_models/gsv-v4-pretrained/vocoder.pth" % (now_dir,), map_location="cpu")
|
||||
print("loading vocoder",self.vocoder.load_state_dict(state_dict_g))
|
||||
state_dict_g = torch.load(
|
||||
"%s/GPT_SoVITS/pretrained_models/gsv-v4-pretrained/vocoder.pth" % (now_dir,), map_location="cpu"
|
||||
)
|
||||
print("loading vocoder", self.vocoder.load_state_dict(state_dict_g))
|
||||
|
||||
self.vocoder_configs["sr"] = 48000
|
||||
self.vocoder_configs["T_ref"] = 500
|
||||
@@ -614,9 +615,6 @@ class TTS:
|
||||
self.vocoder_configs["upsample_rate"] = 480
|
||||
self.vocoder_configs["overlapped_len"] = 12
|
||||
|
||||
|
||||
|
||||
|
||||
self.vocoder = self.vocoder.eval()
|
||||
if self.configs.is_half == True:
|
||||
self.vocoder = self.vocoder.half().to(self.configs.device)
|
||||
@@ -1439,7 +1437,7 @@ class TTS:
|
||||
ref_audio = ref_audio.to(self.configs.device).float()
|
||||
if ref_audio.shape[0] == 2:
|
||||
ref_audio = ref_audio.mean(0).unsqueeze(0)
|
||||
|
||||
|
||||
# tgt_sr = self.vocoder_configs["sr"]
|
||||
tgt_sr = 24000 if self.configs.version == "v3" else 32000
|
||||
if ref_sr != tgt_sr:
|
||||
|
||||
@@ -7,11 +7,17 @@
|
||||
全部按日文识别
|
||||
"""
|
||||
|
||||
import json
|
||||
import logging
|
||||
import os
|
||||
import re
|
||||
import sys
|
||||
import traceback
|
||||
import warnings
|
||||
|
||||
import torch
|
||||
import torchaudio
|
||||
from text.LangSegmenter import LangSegmenter
|
||||
|
||||
logging.getLogger("markdown_it").setLevel(logging.ERROR)
|
||||
logging.getLogger("urllib3").setLevel(logging.ERROR)
|
||||
@@ -23,20 +29,6 @@ logging.getLogger("torchaudio._extension").setLevel(logging.ERROR)
|
||||
logging.getLogger("multipart.multipart").setLevel(logging.ERROR)
|
||||
warnings.simplefilter(action="ignore", category=FutureWarning)
|
||||
|
||||
import json
|
||||
import os
|
||||
import re
|
||||
import sys
|
||||
|
||||
import torch
|
||||
from text.LangSegmenter import LangSegmenter
|
||||
|
||||
try:
|
||||
import gradio.analytics as analytics
|
||||
|
||||
analytics.version_check = lambda: None
|
||||
except:
|
||||
...
|
||||
version = model_version = os.environ.get("version", "v2")
|
||||
path_sovits_v3 = "GPT_SoVITS/pretrained_models/s2Gv3.pth"
|
||||
path_sovits_v4 = "GPT_SoVITS/pretrained_models/gsv-v4-pretrained/s2Gv4.pth"
|
||||
@@ -106,7 +98,7 @@ cnhubert.cnhubert_base_path = cnhubert_base_path
|
||||
|
||||
import random
|
||||
|
||||
from GPT_SoVITS.module.models import SynthesizerTrn, SynthesizerTrnV3,Generator
|
||||
from GPT_SoVITS.module.models import Generator, SynthesizerTrn, SynthesizerTrnV3
|
||||
|
||||
|
||||
def set_seed(seed):
|
||||
@@ -226,9 +218,9 @@ else:
|
||||
resample_transform_dict = {}
|
||||
|
||||
|
||||
def resample(audio_tensor, sr0,sr1):
|
||||
def resample(audio_tensor, sr0, sr1):
|
||||
global resample_transform_dict
|
||||
key="%s-%s"%(sr0,sr1)
|
||||
key = "%s-%s" % (sr0, sr1)
|
||||
if key not in resample_transform_dict:
|
||||
resample_transform_dict[key] = torchaudio.transforms.Resample(sr0, sr1).to(device)
|
||||
return resample_transform_dict[key](audio_tensor)
|
||||
@@ -238,14 +230,18 @@ def resample(audio_tensor, sr0,sr1):
|
||||
# symbol_version-model_version-if_lora_v3
|
||||
from process_ckpt import get_sovits_version_from_path_fast, load_sovits_new
|
||||
|
||||
v3v4set={"v3","v4"}
|
||||
v3v4set = {"v3", "v4"}
|
||||
|
||||
|
||||
def change_sovits_weights(sovits_path, prompt_language=None, text_language=None):
|
||||
global vq_model, hps, version, model_version, dict_language, if_lora_v3
|
||||
version, model_version, if_lora_v3 = get_sovits_version_from_path_fast(sovits_path)
|
||||
print(sovits_path,version, model_version, if_lora_v3)
|
||||
is_exist=is_exist_s2gv3 if model_version=="v3"else is_exist_s2gv4
|
||||
print(sovits_path, version, model_version, if_lora_v3)
|
||||
is_exist = is_exist_s2gv3 if model_version == "v3" else is_exist_s2gv4
|
||||
if if_lora_v3 == True and is_exist == False:
|
||||
info = "GPT_SoVITS/pretrained_models/s2Gv3.pth" + i18n("SoVITS %s 底模缺失，无法加载相应 LoRA 权重"%model_version)
|
||||
info = "GPT_SoVITS/pretrained_models/s2Gv3.pth" + i18n(
|
||||
"SoVITS %s 底模缺失，无法加载相应 LoRA 权重" % model_version
|
||||
)
|
||||
gr.Warning(info)
|
||||
raise FileExistsError(info)
|
||||
dict_language = dict_language_v1 if version == "v1" else dict_language_v2
|
||||
@@ -276,10 +272,15 @@ def change_sovits_weights(sovits_path, prompt_language=None, text_language=None)
|
||||
prompt_language_update,
|
||||
text_update,
|
||||
text_language_update,
|
||||
{"__type__": "update", "visible": visible_sample_steps, "value": 32 if model_version=="v3"else 8,"choices":[4, 8, 16, 32,64,128]if model_version=="v3"else [4, 8, 16, 32]},
|
||||
{
|
||||
"__type__": "update",
|
||||
"visible": visible_sample_steps,
|
||||
"value": 32 if model_version == "v3" else 8,
|
||||
"choices": [4, 8, 16, 32, 64, 128] if model_version == "v3" else [4, 8, 16, 32],
|
||||
},
|
||||
{"__type__": "update", "visible": visible_inp_refs},
|
||||
{"__type__": "update", "value": False, "interactive": True if model_version not in v3v4set else False},
|
||||
{"__type__": "update", "visible": True if model_version =="v3" else False},
|
||||
{"__type__": "update", "visible": True if model_version == "v3" else False},
|
||||
{"__type__": "update", "value": i18n("模型加载中，请等待"), "interactive": False},
|
||||
)
|
||||
|
||||
@@ -304,7 +305,7 @@ def change_sovits_weights(sovits_path, prompt_language=None, text_language=None)
|
||||
)
|
||||
model_version = version
|
||||
else:
|
||||
hps.model.version=model_version
|
||||
hps.model.version = model_version
|
||||
vq_model = SynthesizerTrnV3(
|
||||
hps.data.filter_length // 2 + 1,
|
||||
hps.train.segment_size // hps.data.hop_length,
|
||||
@@ -326,7 +327,7 @@ def change_sovits_weights(sovits_path, prompt_language=None, text_language=None)
|
||||
else:
|
||||
path_sovits = path_sovits_v3 if model_version == "v3" else path_sovits_v4
|
||||
print(
|
||||
"loading sovits_%spretrained_G"%model_version,
|
||||
"loading sovits_%spretrained_G" % model_version,
|
||||
vq_model.load_state_dict(load_sovits_new(path_sovits)["weight"], strict=False),
|
||||
)
|
||||
lora_rank = dict_s2["lora_rank"]
|
||||
@@ -337,7 +338,7 @@ def change_sovits_weights(sovits_path, prompt_language=None, text_language=None)
|
||||
init_lora_weights=True,
|
||||
)
|
||||
vq_model.cfm = get_peft_model(vq_model.cfm, lora_config)
|
||||
print("loading sovits_%s_lora%s" % (model_version,lora_rank))
|
||||
print("loading sovits_%s_lora%s" % (model_version, lora_rank))
|
||||
vq_model.load_state_dict(dict_s2["weight"], strict=False)
|
||||
vq_model.cfm = vq_model.cfm.merge_and_unload()
|
||||
# torch.save(vq_model.state_dict(),"merge_win.pth")
|
||||
@@ -350,10 +351,15 @@ def change_sovits_weights(sovits_path, prompt_language=None, text_language=None)
|
||||
prompt_language_update,
|
||||
text_update,
|
||||
text_language_update,
|
||||
{"__type__": "update", "visible": visible_sample_steps, "value":32 if model_version=="v3"else 8,"choices":[4, 8, 16, 32,64,128]if model_version=="v3"else [4, 8, 16, 32]},
|
||||
{
|
||||
"__type__": "update",
|
||||
"visible": visible_sample_steps,
|
||||
"value": 32 if model_version == "v3" else 8,
|
||||
"choices": [4, 8, 16, 32, 64, 128] if model_version == "v3" else [4, 8, 16, 32],
|
||||
},
|
||||
{"__type__": "update", "visible": visible_inp_refs},
|
||||
{"__type__": "update", "value": False, "interactive": True if model_version not in v3v4set else False},
|
||||
{"__type__": "update", "visible": True if model_version =="v3" else False},
|
||||
{"__type__": "update", "visible": True if model_version == "v3" else False},
|
||||
{"__type__": "update", "value": i18n("合成语音"), "interactive": True},
|
||||
)
|
||||
with open("./weight.json") as f:
|
||||
@@ -400,7 +406,7 @@ now_dir = os.getcwd()
|
||||
|
||||
|
||||
def init_bigvgan():
|
||||
global bigvgan_model,hifigan_model
|
||||
global bigvgan_model, hifigan_model
|
||||
from BigVGAN import bigvgan
|
||||
|
||||
bigvgan_model = bigvgan.BigVGAN.from_pretrained(
|
||||
@@ -411,17 +417,20 @@ def init_bigvgan():
|
||||
bigvgan_model.remove_weight_norm()
|
||||
bigvgan_model = bigvgan_model.eval()
|
||||
if hifigan_model:
|
||||
hifigan_model=hifigan_model.cpu()
|
||||
hifigan_model=None
|
||||
try:torch.cuda.empty_cache()
|
||||
except:pass
|
||||
hifigan_model = hifigan_model.cpu()
|
||||
hifigan_model = None
|
||||
try:
|
||||
torch.cuda.empty_cache()
|
||||
except:
|
||||
pass
|
||||
if is_half == True:
|
||||
bigvgan_model = bigvgan_model.half().to(device)
|
||||
else:
|
||||
bigvgan_model = bigvgan_model.to(device)
|
||||
|
||||
|
||||
def init_hifigan():
|
||||
global hifigan_model,bigvgan_model
|
||||
global hifigan_model, bigvgan_model
|
||||
hifigan_model = Generator(
|
||||
initial_channel=100,
|
||||
resblock="1",
|
||||
@@ -430,26 +439,32 @@ def init_hifigan():
|
||||
upsample_rates=[10, 6, 2, 2, 2],
|
||||
upsample_initial_channel=512,
|
||||
upsample_kernel_sizes=[20, 12, 4, 4, 4],
|
||||
gin_channels=0, is_bias=True
|
||||
gin_channels=0,
|
||||
is_bias=True,
|
||||
)
|
||||
hifigan_model.eval()
|
||||
hifigan_model.remove_weight_norm()
|
||||
state_dict_g = torch.load("%s/GPT_SoVITS/pretrained_models/gsv-v4-pretrained/vocoder.pth" % (now_dir,), map_location="cpu")
|
||||
print("loading vocoder",hifigan_model.load_state_dict(state_dict_g))
|
||||
state_dict_g = torch.load(
|
||||
"%s/GPT_SoVITS/pretrained_models/gsv-v4-pretrained/vocoder.pth" % (now_dir,), map_location="cpu"
|
||||
)
|
||||
print("loading vocoder", hifigan_model.load_state_dict(state_dict_g))
|
||||
if bigvgan_model:
|
||||
bigvgan_model=bigvgan_model.cpu()
|
||||
bigvgan_model=None
|
||||
try:torch.cuda.empty_cache()
|
||||
except:pass
|
||||
bigvgan_model = bigvgan_model.cpu()
|
||||
bigvgan_model = None
|
||||
try:
|
||||
torch.cuda.empty_cache()
|
||||
except:
|
||||
pass
|
||||
if is_half == True:
|
||||
hifigan_model = hifigan_model.half().to(device)
|
||||
else:
|
||||
hifigan_model = hifigan_model.to(device)
|
||||
|
||||
bigvgan_model=hifigan_model=None
|
||||
if model_version=="v3":
|
||||
|
||||
bigvgan_model = hifigan_model = None
|
||||
if model_version == "v3":
|
||||
init_bigvgan()
|
||||
if model_version=="v4":
|
||||
if model_version == "v4":
|
||||
init_hifigan()
|
||||
|
||||
|
||||
@@ -831,17 +846,17 @@ def get_tts_wav(
|
||||
ref_audio = ref_audio.to(device).float()
|
||||
if ref_audio.shape[0] == 2:
|
||||
ref_audio = ref_audio.mean(0).unsqueeze(0)
|
||||
tgt_sr=24000 if model_version=="v3"else 32000
|
||||
tgt_sr = 24000 if model_version == "v3" else 32000
|
||||
if sr != tgt_sr:
|
||||
ref_audio = resample(ref_audio, sr,tgt_sr)
|
||||
ref_audio = resample(ref_audio, sr, tgt_sr)
|
||||
# print("ref_audio",ref_audio.abs().mean())
|
||||
mel2 = mel_fn(ref_audio)if model_version=="v3"else mel_fn_v4(ref_audio)
|
||||
mel2 = mel_fn(ref_audio) if model_version == "v3" else mel_fn_v4(ref_audio)
|
||||
mel2 = norm_spec(mel2)
|
||||
T_min = min(mel2.shape[2], fea_ref.shape[2])
|
||||
mel2 = mel2[:, :, :T_min]
|
||||
fea_ref = fea_ref[:, :, :T_min]
|
||||
Tref=468 if model_version=="v3"else 500
|
||||
Tchunk=934 if model_version=="v3"else 1000
|
||||
Tref = 468 if model_version == "v3" else 500
|
||||
Tchunk = 934 if model_version == "v3" else 1000
|
||||
if T_min > Tref:
|
||||
mel2 = mel2[:, :, -Tref:]
|
||||
fea_ref = fea_ref[:, :, -Tref:]
|
||||
@@ -866,13 +881,13 @@ def get_tts_wav(
|
||||
cfm_resss.append(cfm_res)
|
||||
cfm_res = torch.cat(cfm_resss, 2)
|
||||
cfm_res = denorm_spec(cfm_res)
|
||||
if model_version=="v3":
|
||||
if model_version == "v3":
|
||||
if bigvgan_model == None:
|
||||
init_bigvgan()
|
||||
else:#v4
|
||||
else: # v4
|
||||
if hifigan_model == None:
|
||||
init_hifigan()
|
||||
vocoder_model=bigvgan_model if model_version=="v3"else hifigan_model
|
||||
vocoder_model = bigvgan_model if model_version == "v3" else hifigan_model
|
||||
with torch.inference_mode():
|
||||
wav_gen = vocoder_model(cfm_res)
|
||||
audio = wav_gen[0][0] # .cpu().detach().numpy()
|
||||
@@ -886,9 +901,12 @@ def get_tts_wav(
|
||||
t1 = ttime()
|
||||
print("%.3f\t%.3f\t%.3f\t%.3f" % (t[0], sum(t[1::3]), sum(t[2::3]), sum(t[3::3])))
|
||||
audio_opt = torch.cat(audio_opt, 0) # np.concatenate
|
||||
if model_version in {"v1","v2"}:opt_sr=32000
|
||||
elif model_version=="v3":opt_sr=24000
|
||||
else:opt_sr=48000#v4
|
||||
if model_version in {"v1", "v2"}:
|
||||
opt_sr = 32000
|
||||
elif model_version == "v3":
|
||||
opt_sr = 24000
|
||||
else:
|
||||
opt_sr = 48000 # v4
|
||||
if if_sr == True and opt_sr == 24000:
|
||||
print(i18n("音频超分中"))
|
||||
audio_opt, opt_sr = audio_sr(audio_opt.unsqueeze(0), opt_sr)
|
||||
@@ -1061,7 +1079,7 @@ def html_left(text, label="p"):
|
||||
</div>"""
|
||||
|
||||
|
||||
with gr.Blocks(title="GPT-SoVITS WebUI") as app:
|
||||
with gr.Blocks(title="GPT-SoVITS WebUI", analytics_enabled=False) as app:
|
||||
gr.Markdown(
|
||||
value=i18n("本软件以MIT协议开源, 作者不对软件具备任何控制力, 使用软件者、传播软件导出的声音者自负全责.")
|
||||
+ "<br>"
|
||||
@@ -1131,16 +1149,16 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app:
|
||||
sample_steps = (
|
||||
gr.Radio(
|
||||
label=i18n("采样步数,如果觉得电,提高试试,如果觉得慢,降低试试"),
|
||||
value=32 if model_version=="v3"else 8,
|
||||
choices=[4, 8, 16, 32,64,128]if model_version=="v3"else [4, 8, 16, 32],
|
||||
value=32 if model_version == "v3" else 8,
|
||||
choices=[4, 8, 16, 32, 64, 128] if model_version == "v3" else [4, 8, 16, 32],
|
||||
visible=True,
|
||||
)
|
||||
if model_version in v3v4set
|
||||
else gr.Radio(
|
||||
label=i18n("采样步数,如果觉得电,提高试试,如果觉得慢,降低试试"),
|
||||
choices=[4, 8, 16, 32,64,128]if model_version=="v3"else [4, 8, 16, 32],
|
||||
choices=[4, 8, 16, 32, 64, 128] if model_version == "v3" else [4, 8, 16, 32],
|
||||
visible=False,
|
||||
value=32 if model_version=="v3"else 8,
|
||||
value=32 if model_version == "v3" else 8,
|
||||
)
|
||||
)
|
||||
if_sr_Checkbox = gr.Checkbox(
|
||||
@@ -1148,7 +1166,7 @@ with gr.Blocks(title="GPT-SoVITS WebUI") as app:
|
||||
value=False,
|
||||
interactive=True,
|
||||
show_label=True,
|
||||
visible=False if model_version !="v3" else True,
|
||||
visible=False if model_version != "v3" else True,
|
||||
)
|
||||
gr.Markdown(html_center(i18n("*请填写需要合成的目标文本和语种模式"), "h3"))
|
||||
with gr.Row():
|
||||
|
||||
@@ -14,6 +14,8 @@ import random
|
||||
import re
|
||||
import sys
|
||||
|
||||
import torch
|
||||
|
||||
now_dir = os.getcwd()
|
||||
sys.path.append(now_dir)
|
||||
sys.path.append("%s/GPT_SoVITS" % (now_dir))
|
||||
@@ -25,14 +27,6 @@ logging.getLogger("httpx").setLevel(logging.ERROR)
|
||||
logging.getLogger("asyncio").setLevel(logging.ERROR)
|
||||
logging.getLogger("charset_normalizer").setLevel(logging.ERROR)
|
||||
logging.getLogger("torchaudio._extension").setLevel(logging.ERROR)
|
||||
import torch
|
||||
|
||||
try:
|
||||
import gradio.analytics as analytics
|
||||
|
||||
analytics.version_check = lambda: None
|
||||
except:
|
||||
...
|
||||
|
||||
|
||||
infer_ttswebui = os.environ.get("infer_ttswebui", 9872)
|
||||
@@ -262,15 +256,17 @@ SoVITS_names, GPT_names = get_weights_names(GPT_weight_root, SoVITS_weight_root)
|
||||
|
||||
from process_ckpt import get_sovits_version_from_path_fast
|
||||
|
||||
v3v4set={"v3","v4"}
|
||||
v3v4set = {"v3", "v4"}
|
||||
|
||||
|
||||
def change_sovits_weights(sovits_path, prompt_language=None, text_language=None):
|
||||
global version, model_version, dict_language, if_lora_v3
|
||||
version, model_version, if_lora_v3 = get_sovits_version_from_path_fast(sovits_path)
|
||||
# print(sovits_path,version, model_version, if_lora_v3)
|
||||
is_exist=is_exist_s2gv3 if model_version=="v3"else is_exist_s2gv4
|
||||
is_exist = is_exist_s2gv3 if model_version == "v3" else is_exist_s2gv4
|
||||
path_sovits = path_sovits_v3 if model_version == "v3" else path_sovits_v4
|
||||
if if_lora_v3 == True and is_exist == False:
|
||||
info = path_sovits + i18n("SoVITS %s 底模缺失，无法加载相应 LoRA 权重"%model_version)
|
||||
info = path_sovits + i18n("SoVITS %s 底模缺失，无法加载相应 LoRA 权重" % model_version)
|
||||
gr.Warning(info)
|
||||
raise FileExistsError(info)
|
||||
dict_language = dict_language_v1 if version == "v1" else dict_language_v2
|
||||
@@ -328,7 +324,7 @@ def change_sovits_weights(sovits_path, prompt_language=None, text_language=None)
|
||||
f.write(json.dumps(data))
|
||||
|
||||
|
||||
with gr.Blocks(title="GPT-SoVITS WebUI") as app:
|
||||
with gr.Blocks(title="GPT-SoVITS WebUI", analytics_enabled=False) as app:
|
||||
gr.Markdown(
|
||||
value=i18n("本软件以MIT协议开源, 作者不对软件具备任何控制力, 使用软件者、传播软件导出的声音者自负全责.")
|
||||
+ "<br>"
|
||||
|
||||
@@ -470,6 +470,7 @@ class TextAudioSpeakerCollateV3:
|
||||
# return ssl_padded, spec_padded,mel_padded, ssl_lengths, spec_lengths, text_padded, text_lengths, wav_padded, wav_lengths,mel_lengths
|
||||
return ssl_padded, spec_padded, mel_padded, ssl_lengths, spec_lengths, text_padded, text_lengths, mel_lengths
|
||||
|
||||
|
||||
class TextAudioSpeakerLoaderV4(torch.utils.data.Dataset):
|
||||
"""
|
||||
1) loads audio, speaker_id, text pairs
|
||||
@@ -596,7 +597,7 @@ class TextAudioSpeakerLoaderV4(torch.utils.data.Dataset):
|
||||
audio_norm, self.filter_length, self.sampling_rate, self.hop_length, self.win_length, center=False
|
||||
)
|
||||
spec = torch.squeeze(spec, 0)
|
||||
spec1 = spectrogram_torch(audio_norm, 1280,32000, 320, 1280,center=False)
|
||||
spec1 = spectrogram_torch(audio_norm, 1280, 32000, 320, 1280, center=False)
|
||||
mel = spec_to_mel_torch(spec1, 1280, 100, 32000, 0, None)
|
||||
mel = self.norm_spec(torch.squeeze(mel, 0))
|
||||
return spec, mel
|
||||
@@ -643,7 +644,7 @@ class TextAudioSpeakerCollateV4:
|
||||
mel_lengths = torch.LongTensor(len(batch))
|
||||
|
||||
spec_padded = torch.FloatTensor(len(batch), batch[0][1].size(0), max_spec_len)
|
||||
mel_padded = torch.FloatTensor(len(batch), batch[0][2].size(0), max_spec_len*2)
|
||||
mel_padded = torch.FloatTensor(len(batch), batch[0][2].size(0), max_spec_len * 2)
|
||||
ssl_padded = torch.FloatTensor(len(batch), batch[0][0].size(1), max_ssl_len)
|
||||
text_padded = torch.LongTensor(len(batch), max_text_len)
|
||||
# wav_padded = torch.FloatTensor(len(batch), 1, max_wav_len)
|
||||
|
||||
@@ -39,24 +39,36 @@ hann_window = {}
|
||||
|
||||
def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False):
|
||||
if torch.min(y) < -1.2:
|
||||
print('min value is ', torch.min(y))
|
||||
print("min value is ", torch.min(y))
|
||||
if torch.max(y) > 1.2:
|
||||
print('max value is ', torch.max(y))
|
||||
print("max value is ", torch.max(y))
|
||||
|
||||
global hann_window
|
||||
dtype_device = str(y.dtype) + '_' + str(y.device)
|
||||
dtype_device = str(y.dtype) + "_" + str(y.device)
|
||||
# wnsize_dtype_device = str(win_size) + '_' + dtype_device
|
||||
key = "%s-%s-%s-%s-%s" %(dtype_device,n_fft, sampling_rate, hop_size, win_size)
|
||||
key = "%s-%s-%s-%s-%s" % (dtype_device, n_fft, sampling_rate, hop_size, win_size)
|
||||
# if wnsize_dtype_device not in hann_window:
|
||||
if key not in hann_window:
|
||||
# hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
|
||||
hann_window[key] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
|
||||
|
||||
y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
|
||||
y = torch.nn.functional.pad(
|
||||
y.unsqueeze(1), (int((n_fft - hop_size) / 2), int((n_fft - hop_size) / 2)), mode="reflect"
|
||||
)
|
||||
y = y.squeeze(1)
|
||||
# spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
|
||||
spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[key],
|
||||
center=center, pad_mode='reflect', normalized=False, onesided=True, return_complex=False)
|
||||
spec = torch.stft(
|
||||
y,
|
||||
n_fft,
|
||||
hop_length=hop_size,
|
||||
win_length=win_size,
|
||||
window=hann_window[key],
|
||||
center=center,
|
||||
pad_mode="reflect",
|
||||
normalized=False,
|
||||
onesided=True,
|
||||
return_complex=False,
|
||||
)
|
||||
|
||||
spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-8)
|
||||
return spec
|
||||
@@ -64,9 +76,9 @@ def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False)
|
||||
|
||||
def spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax):
|
||||
global mel_basis
|
||||
dtype_device = str(spec.dtype) + '_' + str(spec.device)
|
||||
dtype_device = str(spec.dtype) + "_" + str(spec.device)
|
||||
# fmax_dtype_device = str(fmax) + '_' + dtype_device
|
||||
key = "%s-%s-%s-%s-%s-%s"%(dtype_device,n_fft, num_mels, sampling_rate, fmin, fmax)
|
||||
key = "%s-%s-%s-%s-%s-%s" % (dtype_device, n_fft, num_mels, sampling_rate, fmin, fmax)
|
||||
# if fmax_dtype_device not in mel_basis:
|
||||
if key not in mel_basis:
|
||||
mel = librosa_mel_fn(sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax)
|
||||
@@ -78,17 +90,25 @@ def spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax):
|
||||
return spec
|
||||
|
||||
|
||||
|
||||
def mel_spectrogram_torch(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False):
|
||||
if torch.min(y) < -1.2:
|
||||
print('min value is ', torch.min(y))
|
||||
print("min value is ", torch.min(y))
|
||||
if torch.max(y) > 1.2:
|
||||
print('max value is ', torch.max(y))
|
||||
print("max value is ", torch.max(y))
|
||||
|
||||
global mel_basis, hann_window
|
||||
dtype_device = str(y.dtype) + '_' + str(y.device)
|
||||
dtype_device = str(y.dtype) + "_" + str(y.device)
|
||||
# fmax_dtype_device = str(fmax) + '_' + dtype_device
|
||||
fmax_dtype_device = "%s-%s-%s-%s-%s-%s-%s-%s"%(dtype_device,n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax)
|
||||
fmax_dtype_device = "%s-%s-%s-%s-%s-%s-%s-%s" % (
|
||||
dtype_device,
|
||||
n_fft,
|
||||
num_mels,
|
||||
sampling_rate,
|
||||
hop_size,
|
||||
win_size,
|
||||
fmin,
|
||||
fmax,
|
||||
)
|
||||
# wnsize_dtype_device = str(win_size) + '_' + dtype_device
|
||||
wnsize_dtype_device = fmax_dtype_device
|
||||
if fmax_dtype_device not in mel_basis:
|
||||
@@ -97,11 +117,23 @@ def mel_spectrogram_torch(y, n_fft, num_mels, sampling_rate, hop_size, win_size,
|
||||
if wnsize_dtype_device not in hann_window:
|
||||
hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
|
||||
|
||||
y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
|
||||
y = torch.nn.functional.pad(
|
||||
y.unsqueeze(1), (int((n_fft - hop_size) / 2), int((n_fft - hop_size) / 2)), mode="reflect"
|
||||
)
|
||||
y = y.squeeze(1)
|
||||
|
||||
spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
|
||||
center=center, pad_mode='reflect', normalized=False, onesided=True, return_complex=False)
|
||||
spec = torch.stft(
|
||||
y,
|
||||
n_fft,
|
||||
hop_length=hop_size,
|
||||
win_length=win_size,
|
||||
window=hann_window[wnsize_dtype_device],
|
||||
center=center,
|
||||
pad_mode="reflect",
|
||||
normalized=False,
|
||||
onesided=True,
|
||||
return_complex=False,
|
||||
)
|
||||
|
||||
spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-8)
|
||||
|
||||
|
||||
@@ -414,7 +414,8 @@ class Generator(torch.nn.Module):
|
||||
upsample_rates,
|
||||
upsample_initial_channel,
|
||||
upsample_kernel_sizes,
|
||||
gin_channels=0,is_bias=False,
|
||||
gin_channels=0,
|
||||
is_bias=False,
|
||||
):
|
||||
super(Generator, self).__init__()
|
||||
self.num_kernels = len(resblock_kernel_sizes)
|
||||
@@ -1173,7 +1174,7 @@ class SynthesizerTrnV3(nn.Module):
|
||||
quantized = F.interpolate(quantized, scale_factor=2, mode="nearest") ##BCT
|
||||
x, m_p, logs_p, y_mask = self.enc_p(quantized, y_lengths, text, text_lengths, ge)
|
||||
fea = self.bridge(x)
|
||||
fea = F.interpolate(fea, scale_factor=(1.875 if self.version=="v3"else 2), mode="nearest") ##BCT
|
||||
fea = F.interpolate(fea, scale_factor=(1.875 if self.version == "v3" else 2), mode="nearest") ##BCT
|
||||
fea, y_mask_ = self.wns1(
|
||||
fea, mel_lengths, ge
|
||||
) ##If the 1-minute fine-tuning works fine, no need to manually adjust the learning rate.
|
||||
@@ -1196,9 +1197,9 @@ class SynthesizerTrnV3(nn.Module):
|
||||
ge = self.ref_enc(refer[:, :704] * refer_mask, refer_mask)
|
||||
y_lengths = torch.LongTensor([int(codes.size(2) * 2)]).to(codes.device)
|
||||
if speed == 1:
|
||||
sizee = int(codes.size(2) * (3.875 if self.version=="v3"else 4))
|
||||
sizee = int(codes.size(2) * (3.875 if self.version == "v3" else 4))
|
||||
else:
|
||||
sizee = int(codes.size(2) * (3.875 if self.version=="v3"else 4) / speed) + 1
|
||||
sizee = int(codes.size(2) * (3.875 if self.version == "v3" else 4) / speed) + 1
|
||||
y_lengths1 = torch.LongTensor([sizee]).to(codes.device)
|
||||
text_lengths = torch.LongTensor([text.size(-1)]).to(text.device)
|
||||
|
||||
@@ -1207,7 +1208,7 @@ class SynthesizerTrnV3(nn.Module):
|
||||
quantized = F.interpolate(quantized, scale_factor=2, mode="nearest") ##BCT
|
||||
x, m_p, logs_p, y_mask = self.enc_p(quantized, y_lengths, text, text_lengths, ge, speed)
|
||||
fea = self.bridge(x)
|
||||
fea = F.interpolate(fea, scale_factor=(1.875 if self.version=="v3"else 2), mode="nearest") ##BCT
|
||||
fea = F.interpolate(fea, scale_factor=(1.875 if self.version == "v3" else 2), mode="nearest") ##BCT
|
||||
####more wn paramter to learn mel
|
||||
fea, y_mask_ = self.wns1(fea, y_lengths1, ge)
|
||||
return fea, ge
|
||||
|
||||
@@ -28,18 +28,18 @@ def my_save(fea, path): #####fix issue: torch.save doesn't support chinese path
|
||||
from io import BytesIO
|
||||
|
||||
|
||||
def my_save2(fea, path,cfm_version):
|
||||
def my_save2(fea, path, cfm_version):
|
||||
bio = BytesIO()
|
||||
torch.save(fea, bio)
|
||||
bio.seek(0)
|
||||
data = bio.getvalue()
|
||||
byte=b"03" if cfm_version=="v3"else b"04"
|
||||
byte = b"03" if cfm_version == "v3" else b"04"
|
||||
data = byte + data[2:]
|
||||
with open(path, "wb") as f:
|
||||
f.write(data)
|
||||
|
||||
|
||||
def savee(ckpt, name, epoch, steps, hps, cfm_version=None,lora_rank=None):
|
||||
def savee(ckpt, name, epoch, steps, hps, cfm_version=None, lora_rank=None):
|
||||
try:
|
||||
opt = OrderedDict()
|
||||
opt["weight"] = {}
|
||||
@@ -51,7 +51,7 @@ def savee(ckpt, name, epoch, steps, hps, cfm_version=None,lora_rank=None):
|
||||
opt["info"] = "%sepoch_%siteration" % (epoch, steps)
|
||||
if lora_rank:
|
||||
opt["lora_rank"] = lora_rank
|
||||
my_save2(opt, "%s/%s.pth" % (hps.save_weight_dir, name),cfm_version)
|
||||
my_save2(opt, "%s/%s.pth" % (hps.save_weight_dir, name), cfm_version)
|
||||
else:
|
||||
my_save(opt, "%s/%s.pth" % (hps.save_weight_dir, name))
|
||||
return "Success."
|
||||
|
||||
@@ -31,7 +31,6 @@ from module.data_utils import (
|
||||
TextAudioSpeakerLoaderV3,
|
||||
TextAudioSpeakerCollateV4,
|
||||
TextAudioSpeakerLoaderV4,
|
||||
|
||||
)
|
||||
from module.models import (
|
||||
SynthesizerTrnV3 as SynthesizerTrn,
|
||||
@@ -88,8 +87,8 @@ def run(rank, n_gpus, hps):
|
||||
if torch.cuda.is_available():
|
||||
torch.cuda.set_device(rank)
|
||||
|
||||
TextAudioSpeakerLoader=TextAudioSpeakerLoaderV3 if hps.model.version=="v3"else TextAudioSpeakerLoaderV4
|
||||
TextAudioSpeakerCollate=TextAudioSpeakerCollateV3 if hps.model.version=="v3"else TextAudioSpeakerCollateV4
|
||||
TextAudioSpeakerLoader = TextAudioSpeakerLoaderV3 if hps.model.version == "v3" else TextAudioSpeakerLoaderV4
|
||||
TextAudioSpeakerCollate = TextAudioSpeakerCollateV3 if hps.model.version == "v3" else TextAudioSpeakerCollateV4
|
||||
train_dataset = TextAudioSpeakerLoader(hps.data) ########
|
||||
train_sampler = DistributedBucketSampler(
|
||||
train_dataset,
|
||||
@@ -365,7 +364,8 @@ def train_and_evaluate(rank, epoch, hps, nets, optims, schedulers, scaler, loade
|
||||
hps.name + "_e%s_s%s_l%s" % (epoch, global_step, lora_rank),
|
||||
epoch,
|
||||
global_step,
|
||||
hps,cfm_version=hps.model.version,
|
||||
hps,
|
||||
cfm_version=hps.model.version,
|
||||
lora_rank=lora_rank,
|
||||
),
|
||||
)
|
||||
|
||||
@@ -1,27 +1,28 @@
|
||||
# This code is modified from https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/paddlespeech/t2s/frontend/g2pw
|
||||
# This code is modified from https://github.com/GitYCC/g2pW
|
||||
|
||||
import warnings
|
||||
|
||||
warnings.filterwarnings("ignore")
|
||||
import json
|
||||
import os
|
||||
import warnings
|
||||
import zipfile
|
||||
from typing import Any, Dict, List, Tuple
|
||||
|
||||
import numpy as np
|
||||
import onnxruntime
|
||||
import requests
|
||||
|
||||
onnxruntime.set_default_logger_severity(3)
|
||||
import torch
|
||||
from opencc import OpenCC
|
||||
from pypinyin import Style, pinyin
|
||||
from transformers import AutoTokenizer
|
||||
from transformers.models.auto.tokenization_auto import AutoTokenizer
|
||||
|
||||
from ..zh_normalization.char_convert import tranditional_to_simplified
|
||||
from .dataset import get_char_phoneme_labels, get_phoneme_labels, prepare_onnx_input
|
||||
from .utils import load_config
|
||||
|
||||
onnxruntime.set_default_logger_severity(3)
|
||||
onnxruntime.preload_dlls()
|
||||
warnings.filterwarnings("ignore")
|
||||
|
||||
model_version = "1.1"
|
||||
|
||||
|
||||
@@ -87,7 +88,7 @@ class G2PWOnnxConverter:
|
||||
sess_options = onnxruntime.SessionOptions()
|
||||
sess_options.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
|
||||
sess_options.execution_mode = onnxruntime.ExecutionMode.ORT_SEQUENTIAL
|
||||
sess_options.intra_op_num_threads = 2
|
||||
sess_options.intra_op_num_threads = 2 if torch.cuda.is_available() else 0
|
||||
try:
|
||||
self.session_g2pW = onnxruntime.InferenceSession(
|
||||
os.path.join(uncompress_path, "g2pW.onnx"),
|
||||
|
||||
@@ -16,7 +16,7 @@ logging.getLogger("matplotlib").setLevel(logging.ERROR)
|
||||
|
||||
MATPLOTLIB_FLAG = False
|
||||
|
||||
logging.basicConfig(stream=sys.stdout, level=logging.ERROR)
|
||||
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
|
||||
logger = logging
|
||||
|
||||
|
||||
@@ -309,13 +309,13 @@ def check_git_hash(model_dir):
|
||||
def get_logger(model_dir, filename="train.log"):
|
||||
global logger
|
||||
logger = logging.getLogger(os.path.basename(model_dir))
|
||||
logger.setLevel(logging.ERROR)
|
||||
logger.setLevel(logging.INFO)
|
||||
|
||||
formatter = logging.Formatter("%(asctime)s\t%(name)s\t%(levelname)s\t%(message)s")
|
||||
if not os.path.exists(model_dir):
|
||||
os.makedirs(model_dir)
|
||||
h = logging.FileHandler(os.path.join(model_dir, filename))
|
||||
h.setLevel(logging.ERROR)
|
||||
h.setLevel(logging.INFO)
|
||||
h.setFormatter(formatter)
|
||||
logger.addHandler(h)
|
||||
return logger
|
||||
|
||||
Reference in New Issue
Block a user