optimize: some training optimizations (#95)

* optimzie(train&uvr5): rm sf & simp. AudioPre

* fix(audio): too many mallocs

* feat(audio): load_audio support stereo

* fix(audio): float32 wav saving

* fix(train): missing ckpt var

tools/cmd/infer-pm-index256.py

@@ -18,7 +18,6 @@ from time import time as ttime
 # import pyworld
 import librosa
 import numpy as np
-import soundfile as sf
 import torch.nn.functional as F
 from fairseq import checkpoint_utils
 
@@ -33,6 +32,7 @@ from scipy.io import wavfile
 # from models import SynthesizerTrn256NSFsim as SynthesizerTrn256#hifigan_nsf
 # from models import SynthesizerTrn256NSFsimFlow as SynthesizerTrn256#hifigan_nsf
 
+from infer.lib.audio import load_audio
 
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 model_path = r"E:\codes\py39\vits_vc_gpu_train\assets\hubert\hubert_base.pt"  #
@@ -132,7 +132,7 @@ for idx, name in enumerate(
 ):  ##
     wav_path = "todo-songs/%s" % name  #
     f0_up_key = -2  #
-    audio, sampling_rate = sf.read(wav_path)
+    audio, sampling_rate = load_audio(wav_path)
     if len(audio.shape) > 1:
         audio = librosa.to_mono(audio.transpose(1, 0))
     if sampling_rate != 16000:

tools/cmd/onnx/infer.py

@@ -1,8 +1,9 @@
-import soundfile
 import librosa
 
 from rvc.onnx import RVC
 
+from infer.lib.audio import save_audio
+
 hop_size = 512
 sampling_rate = 40000  # 采样率
 f0_up_key = 0  # 升降调
@@ -19,4 +20,4 @@ wav, sr = librosa.load(wav_path, sr=sampling_rate)
 
 audio = model.infer(wav, sr, sampling_rate, sid, f0_method, f0_up_key)
 
-soundfile.write(out_path, audio, sampling_rate)
+save_audio(out_path, audio, sampling_rate)