better preprocessing norm

custom_nodes/audio_nodes.py

@@ -203,7 +203,7 @@ class PreviewAudio:
             output_path = increment_filename_no_overwrite(output_path)
         
         input_audio = get_audio(audio)
-        print(save_input_audio(output_path,input_audio,to_int16=True,to_stereo=save_channels==2))
+        print(save_input_audio(output_path,input_audio,to_int16=save_format!="mp3",to_stereo=save_channels==2))
 
         tempdir = os.path.join(temp_path,"preview")
         os.makedirs(tempdir, exist_ok=True)

custom_nodes/rvc_nodes.py

@@ -235,8 +235,7 @@ class RVCProcessDatasetNode:
                 "n_threads": ("INT", {"default": get_optimal_threads(), "min": 1, "max": multiprocessing.cpu_count()}),
                 "period": ("FLOAT", {"default": 3., "min": 1., "max": 10., "step": .1}),
                 "overlap": ("FLOAT",{"default": .3, "min": .1, "max": 1., "step": .1}),
-                "max_volume": ("FLOAT",{"default": 1., "min": .1, "max": 1., "step": .05}),
-                "alpha": ("FLOAT",{"default": .75, "min": .05, "max": 1., "step": .05}),
+                "max_volume": ("FLOAT",{"default": .99, "min": .1, "max": 1., "step": .01}),
                 "mute_ratio": ("FLOAT",{"default": .0, "min": .0, "max": .5, "step": .01}),
             }
         }
@@ -248,13 +247,13 @@ class RVCProcessDatasetNode:
 
     CATEGORY = CATEGORY
 
-    def process(self, model_name: str, dataset: str, hubert_model, pitch_extraction_params={}, sr="40k", n_threads=1, period=3., overlap=.3, max_volume=1., alpha=.75, mute_ratio=.0):
+    def process(self, model_name: str, dataset: str, hubert_model, pitch_extraction_params={}, sr="40k", n_threads=1, period=3., overlap=.3, max_volume=1., mute_ratio=.0):
         
         assert model_name, "Please provide a model name!"
         assert dataset, "Please upload a dataset!"
         
         f0_method = pitch_extraction_params.get("f0_method", "")
-        cached_params = [model_name, dataset, period, overlap, max_volume, alpha, mute_ratio, sr, f0_method]
+        cached_params = [model_name, dataset, period, overlap, max_volume, mute_ratio, sr, f0_method]
         crepe_hop_length = pitch_extraction_params.get("crepe_hop_length",160)
 
         if "crepe" in f0_method: cached_params.append(crepe_hop_length)
@@ -272,7 +271,7 @@ class RVCProcessDatasetNode:
                 files = extract_zip_without_structure(os.path.join(dataset_path,dataset),dataset_dir)
                 assert len(files), "Failed to extract zip file..."
             
-            print(preprocess_trainset(dataset_dir,SR_MAP[sr],n_threads,model_log_dir,period,overlap,max_volume,alpha))
+            print(preprocess_trainset(dataset_dir,SR_MAP[sr],n_threads,model_log_dir,period,overlap,max_volume))
             
             print(extract_features_trainset(hubert_model(), model_log_dir,n_p=n_threads,f0method=f0_method,device=device,if_f0=bool(f0_method),version="v2",crepe_hop_length=crepe_hop_length))
 

examples/rvc-model-trainer.json

@@ -1,27 +1,27 @@
 {
-  "last_node_id": 15,
-  "last_link_id": 25,
+  "last_node_id": 17,
+  "last_link_id": 30,
   "nodes": [
     {
       "id": 3,
       "type": "LoadPitchExtractionParams",
       "pos": [
-        -397,
-        179
+        -387,
+        130
       ],
       "size": {
         "0": 315,
-        "1": 178
+        "1": 202
       },
       "flags": {},
-      "order": 2,
+      "order": 0,
       "mode": 0,
       "outputs": [
         {
           "name": "pitch_extraction_params",
           "type": "PITCH_EXTRACTION",
           "links": [
-            17
+            29
           ],
           "slot_index": 0,
           "shape": 3
@@ -36,7 +36,8 @@
         0.75,
         0,
         0.25,
-        0.25
+        0.25,
+        160
       ]
     },
     {
@@ -51,14 +52,14 @@
         "1": 58
       },
       "flags": {},
-      "order": 0,
+      "order": 1,
       "mode": 0,
       "outputs": [
         {
           "name": "hubert_model",
           "type": "HUBERT_MODEL",
           "links": [
-            16
+            28
           ],
           "slot_index": 0,
           "shape": 3
@@ -139,7 +140,7 @@
         "1": 146
       },
       "flags": {},
-      "order": 1,
+      "order": 2,
       "mode": 0,
       "outputs": [
         {
@@ -185,7 +186,7 @@
         "1": 150
       },
       "flags": {},
-      "order": 3,
+      "order": 4,
       "mode": 0,
       "inputs": [
         {
@@ -285,65 +286,12 @@
         true
       ]
     },
-    {
-      "id": 13,
-      "type": "RVCProcessDatasetNode",
-      "pos": [
-        -8,
-        176
-      ],
-      "size": {
-        "0": 337.6000061035156,
-        "1": 294
-      },
-      "flags": {},
-      "order": 4,
-      "mode": 0,
-      "inputs": [
-        {
-          "name": "hubert_model",
-          "type": "HUBERT_MODEL",
-          "link": 16
-        },
-        {
-          "name": "pitch_extraction_params",
-          "type": "PITCH_EXTRACTION",
-          "link": 17
-        }
-      ],
-      "outputs": [
-        {
-          "name": "rvc_dataset_pipe",
-          "type": "RVC_DATASET_PIPE",
-          "links": [
-            22
-          ],
-          "slot_index": 0,
-          "shape": 3
-        }
-      ],
-      "properties": {
-        "Node name for S&R": "RVCProcessDatasetNode"
-      },
-      "widgets_values": [
-        "Comfy-Sayano",
-        "Sayano.zip",
-        "40k",
-        1,
-        3,
-        0.3,
-        0.9500000000000001,
-        0.75,
-        0.01,
-        "image"
-      ]
-    },
     {
       "id": 15,
       "type": "RVCTrainModelNode",
       "pos": [
         374,
-        119
+        118
       ],
       "size": {
         "0": 506.4000244140625,
@@ -356,7 +304,7 @@
         {
           "name": "rvc_dataset_pipe",
           "type": "RVC_DATASET_PIPE",
-          "link": 22
+          "link": 30
         }
       ],
       "outputs": [
@@ -399,18 +347,70 @@
       },
       "widgets_values": [
         "0",
-        4,
-        300,
-        0,
-        "pretrained_v2/f0Ov2Super40kD.pth",
-        "pretrained_v2/f0Ov2Super40kG.pth",
-        true,
+        5,
+        100,
+        10,
+        "pretrained_v2/G_DMR-V1-32k-f0.pth",
+        "pretrained_v2/D_DMR-V1-32k-f0.pth",
         true,
         false,
+        false,
         true,
         true,
         true
       ]
+    },
+    {
+      "id": 17,
+      "type": "RVCProcessDatasetNode",
+      "pos": [
+        -8,
+        176
+      ],
+      "size": {
+        "0": 337.6000061035156,
+        "1": 270
+      },
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "hubert_model",
+          "type": "HUBERT_MODEL",
+          "link": 28
+        },
+        {
+          "name": "pitch_extraction_params",
+          "type": "PITCH_EXTRACTION",
+          "link": 29
+        }
+      ],
+      "outputs": [
+        {
+          "name": "rvc_dataset_pipe",
+          "type": "RVC_DATASET_PIPE",
+          "links": [
+            30
+          ],
+          "shape": 3,
+          "slot_index": 0
+        }
+      ],
+      "properties": {
+        "Node name for S&R": "RVCProcessDatasetNode"
+      },
+      "widgets_values": [
+        "Comfy-Mae",
+        "Mae.zip",
+        "32k",
+        12,
+        3,
+        0.5,
+        0.9500000000000001,
+        0.01,
+        "image"
+      ]
     }
   ],
   "links": [
@@ -446,30 +446,6 @@
       1,
       "AUDIO,VHS_AUDIO"
     ],
-    [
-      16,
-      5,
-      0,
-      13,
-      0,
-      "HUBERT_MODEL"
-    ],
-    [
-      17,
-      3,
-      0,
-      13,
-      1,
-      "PITCH_EXTRACTION"
-    ],
-    [
-      22,
-      13,
-      0,
-      15,
-      0,
-      "RVC_DATASET_PIPE"
-    ],
     [
       23,
       15,
@@ -493,6 +469,30 @@
       6,
       3,
       "PITCH_EXTRACTION"
+    ],
+    [
+      28,
+      5,
+      0,
+      17,
+      0,
+      "HUBERT_MODEL"
+    ],
+    [
+      29,
+      3,
+      0,
+      17,
+      1,
+      "PITCH_EXTRACTION"
+    ],
+    [
+      30,
+      17,
+      0,
+      15,
+      0,
+      "RVC_DATASET_PIPE"
     ]
   ],
   "groups": [
@@ -522,10 +522,10 @@
   "config": {},
   "extra": {
     "ds": {
-      "scale": 1.1,
+      "scale": 0.6830134553650705,
       "offset": [
-        652.4400126183286,
-        108.78821868462816
+        544.5635228085466,
+        563.3345057885277
       ]
     }
   },

lib/audio.py

@@ -44,7 +44,7 @@ def load_audio(file, sr, **kwargs):
 
     return remix_audio((np.frombuffer(out, np.float32).flatten(), sr),**kwargs)
 
-def remix_audio(input_audio,target_sr=None,norm=False,to_int16=False,resample=False,axis=0,merge_type=None,**kwargs):
+def remix_audio(input_audio,target_sr=None,norm=False,to_int16=False,resample=False,axis=0,merge_type=None,max_volume=.99,**kwargs):
     audio = np.array(input_audio[0],dtype="float32")
     if target_sr is None: target_sr=input_audio[1]
 
@@ -57,10 +57,10 @@ def remix_audio(input_audio,target_sr=None,norm=False,to_int16=False,resample=Fa
         audio=merge_func(audio,axis=axis)
     if norm: audio = librosa.util.normalize(audio,axis=axis)
 
-    audio_max = np.abs(audio).max()/.99
+    audio_max = np.abs(audio).max()/max_volume
     if audio_max > 1: audio = audio / audio_max
         
-    if to_int16: audio = np.clip(audio * MAX_INT16, a_min=-MAX_INT16+1, a_max=MAX_INT16-1).astype("int16")
+    if to_int16: audio = np.clip(audio * MAX_INT16, a_min=1-MAX_INT16, a_max=MAX_INT16-1).astype("int16")
     print(f"after remix: shape={audio.shape}, max={audio.max()}, min={audio.min()}, mean={audio.mean()}, sr={target_sr}")
 
     return audio, target_sr
@@ -71,16 +71,14 @@ def load_input_audio(fname,sr=None,**kwargs):
     print(f"loading sound {fname=} {audio.ndim=} {audio.max()=} {audio.min()=} {audio.dtype=} {sr=}")
     return audio, sr
    
-def save_input_audio(fname,input_audio,sr=None,to_int16=False,to_stereo=False):
+def save_input_audio(fname,input_audio,sr=None,to_int16=False,to_stereo=False,max_volume=.99):
     print(f"saving sound to {fname}")
     os.makedirs(os.path.dirname(fname),exist_ok=True)
     audio=np.array(input_audio[0],dtype="float32")
 
-    if to_int16:
-        audio_max = np.abs(audio).max()/.99
-        if audio_max > 1: audio = audio / audio_max
-        audio = np.clip(audio * MAX_INT16, a_min=-MAX_INT16+1, a_max=MAX_INT16-1)
-
+    audio_max = np.abs(audio).max()/max_volume
+    if audio_max > 1: audio = audio / audio_max
+    if to_int16: audio = np.clip(audio * MAX_INT16, a_min=1-MAX_INT16, a_max=MAX_INT16-1)
     if to_stereo and audio.ndim<2: audio=np.stack([audio,audio],axis=-1)
     
     try:

preprocessing_utils.py

@@ -1,19 +1,17 @@
 import sys, os, multiprocessing
 from threading import Thread
 import numpy as np, os, traceback
-from .lib.model_utils import load_hubert
 from .lib.slicer2 import Slicer
-import librosa, traceback
+import traceback
 from scipy.io import wavfile
-from .lib.audio import load_audio
 from .pitch_extraction import FeatureExtractor
-from .lib.audio import load_input_audio
+from .lib.audio import load_input_audio, remix_audio
 from .lib.utils import gc_collect
 from .config import config
 import torch
 
 class Preprocess:
-    def __init__(self, sr, exp_dir, noparallel=True, period=3.0, overlap=.3, alpha=.75, max_volume=.95):
+    def __init__(self, sr, exp_dir, noparallel=True, period=3.0, overlap=.3, max_volume=.95):
         self.slicer = Slicer(
             sr=sr,
             threshold=-42,
@@ -26,11 +24,10 @@ class Preprocess:
         self.per = period
         self.overlap = overlap
         self.tail = self.per + self.overlap
-        self.max = max_volume
-        self.alpha = alpha
+        self.max_volume = max_volume
         self.exp_dir = exp_dir
-        self.gt_wavs_dir = "%s/0_gt_wavs" % exp_dir
-        self.wavs16k_dir = "%s/1_16k_wavs" % exp_dir
+        self.gt_wavs_dir = os.path.join(exp_dir,"0_gt_wavs")
+        self.wavs16k_dir = os.path.join(exp_dir,"1_16k_wavs")
         self.noparallel = noparallel
         os.makedirs(self.exp_dir, exist_ok=True)
         os.makedirs(self.gt_wavs_dir, exist_ok=True)
@@ -45,26 +42,9 @@ class Preprocess:
         # mutex.release()
 
     def norm_write(self, tmp_audio, idx0, idx1):
-        tmp_max = np.abs(tmp_audio).max()
-        if tmp_max > 2.5:
-            print("%s-%s-%s-filtered" % (idx0, idx1, tmp_max))
-            return
-        tmp_audio = (tmp_audio / tmp_max * (self.max * self.alpha)) + (
-            1 - self.alpha
-        ) * tmp_audio
-        wavfile.write(
-            "%s/%s_%s.wav" % (self.gt_wavs_dir, idx0, idx1),
-            self.sr,
-            tmp_audio.astype(np.float32),
-        )
-        tmp_audio = librosa.resample(
-            tmp_audio, orig_sr=self.sr, target_sr=16000
-        )  # , res_type="soxr_vhq"
-        wavfile.write(
-            "%s/%s_%s.wav" % (self.wavs16k_dir, idx0, idx1),
-            16000,
-            tmp_audio.astype(np.float32),
-        )
+        wavfile.write(os.path.join(self.gt_wavs_dir, f"{idx0}_{idx1}.wav"),self.sr,tmp_audio.astype(np.float32))
+        remixed_audio = remix_audio((tmp_audio, self.sr), target_sr=16000, norm=True,max_volume=self.max_volume)
+        wavfile.write(os.path.join(self.wavs16k_dir, f"{idx0}_{idx1}.wav"),16000,remixed_audio[0].astype(np.float32))
 
     def pipeline(self, path, idx0):
         try:
@@ -208,9 +188,9 @@ class FeatureInput(FeatureExtractor):
                 except:
                     self.printt("f0fail-%s-%s-%s" % (idx, inp_path, traceback.format_exc()))
 
-def preprocess_trainset(inp_root, sr, n_p, exp_dir, period=3.0, overlap=.3, max_volume=1., alpha=.75):
+def preprocess_trainset(inp_root, sr, n_p, exp_dir, period=3.0, overlap=.3, max_volume=1.):
     try:
-        pp = Preprocess(sr, exp_dir, period=period, overlap=overlap, max_volume=max_volume, alpha=alpha)
+        pp = Preprocess(sr, exp_dir, period=period, overlap=overlap, max_volume=max_volume)
         pp.println("start preprocess")
         pp.println(sys.argv)
         pp.pipeline_mp_inp_dir(inp_root, n_p)