Merge pull request #73 from w-okada/dev

update

demo/serverFastAPI.py

@@ -13,14 +13,19 @@ import torch
 import numpy as np
 from scipy.io.wavfile import write, read
 
-sys.path.append("mod")
+# sys.path.append("mod")
-sys.path.append("mod/text")
+# sys.path.append("mod/text")
+
+sys.path.append("/MMVC_Trainer")
+sys.path.append("/MMVC_Trainer/text")
 
 import utils
+import commons 
 from data_utils import TextAudioSpeakerLoader, TextAudioSpeakerCollate
 from models import SynthesizerTrn
 from text.symbols import symbols
-
+from mel_processing import spectrogram_torch
+from text import text_to_sequence, cleaned_text_to_sequence
 
 class VoiceChanger():
     def __init__(self, config, model):
@@ -48,16 +53,52 @@ class VoiceChanger():
         try:
             if gpu<0 or self.gpu_num==0 :
                 with torch.no_grad():
-                    dataset = TextAudioSpeakerLoader("dummy.txt", self.hps.data, no_use_textfile=True)
+                    # dataset = TextAudioSpeakerLoader("dummy.txt", self.hps.data, no_use_textfile=True)
-                    data = dataset.get_audio_text_speaker_pair([ unpackedData, srcId, "a"])
+                    # data = dataset.get_audio_text_speaker_pair([ unpackedData, srcId, "a"])
+
+                    text_norm = text_to_sequence("a", self.hps.data.text_cleaners)
+                    text_norm = commons.intersperse(text_norm, 0)
+                    text_norm = torch.LongTensor(text_norm)
+
+                    audio = torch.FloatTensor(unpackedData.astype(np.float32))
+                    audio_norm = audio /self.hps.data.max_wav_value
+                    audio_norm = audio_norm.unsqueeze(0)
+
+
+                    spec = spectrogram_torch(audio_norm, self.hps.data.filter_length,
+                            self.hps.data.sampling_rate, self.hps.data.hop_length, self.hps.data.win_length,
+                            center=False)
+                    spec = torch.squeeze(spec, 0)
+                    sid = torch.LongTensor([int(srcId)])
+                    
+                    data =  (text_norm, spec, audio_norm, sid)
+
                     data = TextAudioSpeakerCollate()([data])
                     x, x_lengths, spec, spec_lengths, y, y_lengths, sid_src = [x.cpu() for x in data]
                     sid_tgt1 = torch.LongTensor([dstId]).cpu()
                     audio1 = (self.net_g.cpu().voice_conversion(spec, spec_lengths, sid_src=sid_src, sid_tgt=sid_tgt1)[0][0,0].data * self.hps.data.max_wav_value).cpu().float().numpy()
             else:
                 with torch.no_grad():
-                    dataset = TextAudioSpeakerLoader("dummy.txt", self.hps.data, no_use_textfile=True)
+                    # dataset = TextAudioSpeakerLoader("dummy.txt", self.hps.data, no_use_textfile=True)
-                    data = dataset.get_audio_text_speaker_pair([ unpackedData, srcId, "a"])
+                    # data = dataset.get_audio_text_speaker_pair([ unpackedData, srcId, "a"])
+
+                    text_norm = text_to_sequence("a", self.hps.data.text_cleaners)
+                    text_norm = commons.intersperse(text_norm, 0)
+                    text_norm = torch.LongTensor(text_norm)
+
+                    audio = torch.FloatTensor(unpackedData.astype(np.float32))
+                    audio_norm = audio /self.hps.data.max_wav_value
+                    audio_norm = audio_norm.unsqueeze(0)
+
+
+                    spec = spectrogram_torch(audio_norm, self.hps.data.filter_length,
+                            self.hps.data.sampling_rate, self.hps.data.hop_length, self.hps.data.win_length,
+                            center=False)
+                    spec = torch.squeeze(spec, 0)
+                    sid = torch.LongTensor([int(srcId)])
+                    
+                    data =  (text_norm, spec, audio_norm, sid)
+
                     data = TextAudioSpeakerCollate()([data])
                     x, x_lengths, spec, spec_lengths, y, y_lengths, sid_src = [x.cuda(gpu) for x in data]
                     sid_tgt1 = torch.LongTensor([dstId]).cuda(gpu)

demo/serverSIO.py

@@ -9,12 +9,17 @@ import torch
 import numpy as np
 from scipy.io.wavfile import write
 
-sys.path.append("mod")
+sys.path.append("/MMVC_Trainer")
-sys.path.append("mod/text")
+sys.path.append("/MMVC_Trainer/text")
+
+
 import utils
+import commons 
 from data_utils import TextAudioSpeakerLoader, TextAudioSpeakerCollate
 from models import SynthesizerTrn
 from text.symbols import symbols
+from mel_processing import spectrogram_torch
+from text import text_to_sequence, cleaned_text_to_sequence
 
 class MyCustomNamespace(socketio.Namespace): 
     def __init__(self, namespace, config, model):
@@ -50,17 +55,49 @@ class MyCustomNamespace(socketio.Namespace):
 
         if gpu<0 or self.gpu_num==0 :
             with torch.no_grad():
-                dataset = TextAudioSpeakerLoader("dummy.txt", self.hps.data, no_use_textfile=True)
+                
-                data = dataset.get_audio_text_speaker_pair([ unpackedData, srcId, "a"])
+                text_norm = text_to_sequence("a", self.hps.data.text_cleaners)
+                text_norm = commons.intersperse(text_norm, 0)
+                text_norm = torch.LongTensor(text_norm)
+
+                audio = torch.FloatTensor(unpackedData.astype(np.float32))
+                audio_norm = audio /self.hps.data.max_wav_value
+                audio_norm = audio_norm.unsqueeze(0)
+
+
+                spec = spectrogram_torch(audio_norm, self.hps.data.filter_length,
+                        self.hps.data.sampling_rate, self.hps.data.hop_length, self.hps.data.win_length,
+                        center=False)
+                spec = torch.squeeze(spec, 0)
+                sid = torch.LongTensor([int(srcId)])
+                
+                data =  (text_norm, spec, audio_norm, sid)
+
                 data = TextAudioSpeakerCollate()([data])
                 x, x_lengths, spec, spec_lengths, y, y_lengths, sid_src = [x.cpu() for x in data]
                 sid_tgt1 = torch.LongTensor([dstId]).cpu()
                 audio1 = (self.net_g.cpu().voice_conversion(spec, spec_lengths, sid_src=sid_src, sid_tgt=sid_tgt1)[0][0,0].data * self.hps.data.max_wav_value).cpu().float().numpy()
         else:
             with torch.no_grad():
-                dataset = TextAudioSpeakerLoader("dummy.txt", self.hps.data, no_use_textfile=True)
+
-                data = dataset.get_audio_text_speaker_pair([ unpackedData, srcId, "a"])
+                text_norm = text_to_sequence("a", self.hps.data.text_cleaners)
+                text_norm = commons.intersperse(text_norm, 0)
+                text_norm = torch.LongTensor(text_norm)
+
+                audio = torch.FloatTensor(unpackedData.astype(np.float32))
+                audio_norm = audio /self.hps.data.max_wav_value
+                audio_norm = audio_norm.unsqueeze(0)
+
+
+                spec = spectrogram_torch(audio_norm, self.hps.data.filter_length,
+                        self.hps.data.sampling_rate, self.hps.data.hop_length, self.hps.data.win_length,
+                        center=False)
+                spec = torch.squeeze(spec, 0)
+                sid = torch.LongTensor([int(srcId)])
+                
+                data =  (text_norm, spec, audio_norm, sid)
                 data = TextAudioSpeakerCollate()([data])
+
                 x, x_lengths, spec, spec_lengths, y, y_lengths, sid_src = [x.cuda(gpu) for x in data]
                 sid_tgt1 = torch.LongTensor([dstId]).cuda(gpu)
                 audio1 = (self.net_g.cuda(gpu).voice_conversion(spec, spec_lengths, sid_src=sid_src, sid_tgt=sid_tgt1)[0][0,0].data * self.hps.data.max_wav_value).cpu().float().numpy()

frontend/dist/index.html

@@ -1 +1,13 @@
-<!doctype html><html lang="ja" style="width:100%;height:100%;overflow:hidden"><head><meta charset="utf-8"/><title>voice recorder</title><script defer="defer" src="index.js"></script></head><body style="width:100%;height:100%;margin:0"><div id="app" style="width:100%;height:100%"></div><noscript><strong>javascriptを有効にしてください</strong></noscript></body></html>
+<!DOCTYPE html>
+<html lang="ja" style="width: 100%; height: 100%; overflow: hidden">
+    <head>
+        <meta charset="utf-8" />
+        <title>voice recorder</title>
+    <script defer src="index.js"></script></head>
+    <body style="width: 100%; height: 100%; margin: 0px">
+        <div id="app" style="width: 100%; height: 100%"></div>
+        <noscript>
+            <strong>javascriptを有効にしてください</strong>
+        </noscript>
+    </body>
+</html>

start2.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 set -eu
 
-DOCKER_IMAGE=dannadori/voice-changer:20221003_002318
+DOCKER_IMAGE=dannadori/voice-changer:20221028_191234
 #DOCKER_IMAGE=voice-changer
 
 

trainer/Dockerfile

@@ -1,4 +1,4 @@
-FROM dannadori/voice-changer-internal:20221002_193031 as front
+FROM dannadori/voice-changer-internal:20221028_190940 as front
 FROM debian:bullseye-slim as base
 
 ARG DEBIAN_FRONTEND=noninteractive
@@ -8,7 +8,7 @@ RUN apt-get install -y python3-pip git
 RUN apt-get install -y espeak
 RUN apt-get install -y cmake
 
-RUN git clone --depth 1 https://github.com/isletennos/MMVC_Trainer.git -b v1.3.1.0
+RUN git clone --depth 1 https://github.com/isletennos/MMVC_Trainer.git -b v1.3.1.3
 
 RUN pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu113
 
@@ -49,18 +49,22 @@ COPY fine_model/D_180000.pth /MMVC_Trainer/fine_model/D_180000.pth
 
 ### Copy from base
 COPY --from=base --chmod=777 /usr/local/lib/python3.9/dist-packages /usr/local/lib/python3.9/dist-packages
-COPY --from=base --chmod=777 /MMVC_Trainer /MMVC_Trainer
+
+### Copy from frontend
+##### MMVC Trainer
+COPY --from=front --chmod=777 /MMVC_Trainer /MMVC_Trainer
 RUN chmod 0777 /MMVC_Trainer 
 
 WORKDIR /MMVC_Trainer
 ADD /setup.sh  /MMVC_Trainer/
 ADD /exec.sh  /MMVC_Trainer/
 
-### Copy from frontend
+##### Voice changer Internal
 COPY --from=front --chmod=777 /voice-changer-internal/frontend/dist /voice-changer-internal/frontend/dist
 COPY --from=front --chmod=777 /voice-changer-internal/voice-change-service /voice-changer-internal/voice-change-service
 RUN chmod 0777 /voice-changer-internal/voice-change-service 
 
+##### Soft VC
 COPY --from=front /hubert /hubert
 COPY --from=front /acoustic-model /acoustic-model
 COPY --from=front /hifigan /hifigan