Use GPU processing for UVR5 and input audio loading and resampling where possible to improve inference efficiency and reduce CPU usage

train/dataset/extract_f0.py

@@ -8,9 +8,8 @@ import logging
 
 import numpy as np
 
-from infer.audio import load_audio
-from i18n.i18n import I18nAuto
-from tools.progress import should_report
+from i18n.i18n import I18nAuto
+from tools.progress import should_report
 
 
 i18n = I18nAuto()
@@ -44,6 +43,9 @@ elif mode in ("dml", "directml"):
 else:
     raise ValueError("Unsupported F0 extraction mode: %s" % mode)
 
+# CUDA_VISIBLE_DEVICES must be set before infer.audio imports torch/configs.
+from infer.audio import load_audio
+
 f = open("%s/extract_f0_feature.log" % exp_dir, "a", encoding="utf8")
 
 

train/dataset/slicer2.py

@@ -1,15 +1,15 @@
-import numpy as np
+import numpy as np
 
 
 # This function is obtained from librosa.
-def get_rms(
+def get_rms(
     y,
     frame_length=2048,
     hop_length=512,
     pad_mode="constant",
-):
-    padding = (int(frame_length // 2), int(frame_length // 2))
-    y = np.pad(y, padding, mode=pad_mode)
+):
+    padding = (int(frame_length // 2), int(frame_length // 2))
+    y = np.pad(y, padding, mode=pad_mode)
 
     axis = -1
     # put our new within-frame axis at the end for now
@@ -72,16 +72,16 @@ class Slicer:
             ]
 
     # @timeit
-    def slice(self, waveform):
-        if len(waveform.shape) > 1:
-            samples = waveform.mean(axis=0)
-        else:
-            samples = waveform
+    def slice(self, waveform):
+        if len(waveform.shape) > 1:
+            samples = waveform.mean(axis=0)
+        else:
+            samples = waveform
         if samples.shape[0] <= self.min_length:
             return [waveform]
-        rms_list = get_rms(
-            y=samples, frame_length=self.win_size, hop_length=self.hop_size
-        ).squeeze(0)
+        rms_list = get_rms(
+            y=samples, frame_length=self.win_size, hop_length=self.hop_size
+        ).squeeze(0)
         sil_tags = []
         silence_start = None
         clip_start = 0

train/preprocess.py

@@ -10,13 +10,13 @@ n_p = int(sys.argv[3])
 exp_dir = sys.argv[4]
 noparallel = sys.argv[5] == "True"
 per = float(sys.argv[6])
-import os
 import traceback
 
 import librosa
 import numpy as np
 from scipy.io import wavfile
 
+os.environ["RVC_AUDIO_FORCE_CPU"] = "1"
 from infer.audio import load_audio
 from train.dataset.slicer2 import Slicer
 from i18n.i18n import I18nAuto
@@ -73,13 +73,13 @@ class PreProcess:
             self.sr,
             tmp_audio.astype(np.float32),
         )
-        tmp_audio = librosa.resample(
+        audio_16k = librosa.resample(
             tmp_audio, orig_sr=self.sr, target_sr=16000
-        )  # , res_type="soxr_vhq"
+        ).astype(np.float32)
         wavfile.write(
             "%s/%s_%s.wav" % (self.wavs16k_dir, idx0, idx1),
             16000,
-            tmp_audio.astype(np.float32),
+            audio_16k,
         )
         return True
 
@@ -140,19 +140,24 @@ class PreProcess:
                 ("%s/%s" % (inp_root, name), idx, total)
                 for idx, name in enumerate(names)
             ]
-            println(i18n("[数据切分] 待处理：%s | 进程数：%s") % (total, n_p))
+            worker_count = max(n_p, 1)
+            worker_count = min(worker_count, max(total, 1))
+            println(
+                i18n("[数据切分] 待处理：%s | 进程数：%s")
+                % (total, worker_count)
+            )
             if noparallel:
-                for i in range(n_p):
-                    self.pipeline_mp(infos[i::n_p])
+                for i in range(worker_count):
+                    self.pipeline_mp(infos[i::worker_count])
             else:
                 ps = []
-                for i in range(n_p):
+                for i in range(worker_count):
                     p = multiprocessing.Process(
-                        target=self.pipeline_mp, args=(infos[i::n_p],)
+                        target=self.pipeline_mp, args=(infos[i::worker_count],)
                     )
                     ps.append(p)
                     p.start()
-                for i in range(n_p):
+                for i in range(worker_count):
                     ps[i].join()
         except Exception:
             println(i18n("[数据切分][失败] %s") % traceback.format_exc())