From 84860767791fcb9e383563747d229ec9182b28d6 Mon Sep 17 00:00:00 2001
From: wataru <wataru@fdev.local.com>
Date: Thu, 5 Jan 2023 18:56:02 +0900
Subject: [PATCH] cpu

---
 client/demo/src/index.tsx                     |  2 +-
 client/demo/src/options_microphone.tsx        |  5 +---
 .../src/voice-changer-worklet-processor.ts    |  2 +-
 server/voice_changer/VoiceChanger.py          | 27 +++++++++++++++----
 4 files changed, 25 insertions(+), 11 deletions(-)

diff --git a/client/demo/src/index.tsx b/client/demo/src/index.tsx
index 687ef912..d4a427be 100644
--- a/client/demo/src/index.tsx
+++ b/client/demo/src/index.tsx
@@ -149,7 +149,7 @@ const App = () => {
             <>
                 <div className="body-row split-3-1-1-1-4 left-padding-1 highlight">
                     <div className="body-item-title">monitor:</div>
-                    <div className="body-item-text">vol(db):{volume.toFixed(4)}</div>
+                    <div className="body-item-text">vol(rms):{volume.toFixed(4)}</div>
                     <div className="body-item-text">buf(ms):{bufferingTime}</div>
                     <div className="body-item-text">res(ms):{responseTime}</div>
                     <div className="body-item-text"></div>
diff --git a/client/demo/src/options_microphone.tsx b/client/demo/src/options_microphone.tsx
index 1963ffd7..e297f653 100644
--- a/client/demo/src/options_microphone.tsx
+++ b/client/demo/src/options_microphone.tsx
@@ -133,12 +133,9 @@ export const useMicrophoneOptions = (audioContext?: AudioContext): MicrophoneOpt
     }, [inputAudioDeviceInfo, audioInput])
 
     const audioMediaInputRow = useMemo(() => {
-        console.log("GEN:audioMediaInputRow1")
         if (audioInput != "file") {
-            console.log("GEN:audioMediaInputRow2")
             return <></>
         }
-        console.log("GEN:audioMediaInputRow3")
 
         const onFileLoadClicked = async () => {
             const url = await fileSelectorAsDataURL("")
@@ -185,7 +182,7 @@ export const useMicrophoneOptions = (audioContext?: AudioContext): MicrophoneOpt
             </div>
         )
     }, [audioInput, audioOutput])
-    console.log("GEN:audioMediaInputRow3")
+
     useEffect(() => {
         if (!audioContext) {
             return
diff --git a/client/lib/worklet/src/voice-changer-worklet-processor.ts b/client/lib/worklet/src/voice-changer-worklet-processor.ts
index c7d660d7..3de70845 100644
--- a/client/lib/worklet/src/voice-changer-worklet-processor.ts
+++ b/client/lib/worklet/src/voice-changer-worklet-processor.ts
@@ -58,7 +58,7 @@ class VoiceChangerWorkletProcessor extends AudioWorkletProcessor {
         }
 
         if (this.playBuffer.length === 0) {
-            console.log("[worklet] no play buffer")
+            // console.log("[worklet] no play buffer")
             return true
         }
 
diff --git a/server/voice_changer/VoiceChanger.py b/server/voice_changer/VoiceChanger.py
index 390770c6..3567c2c8 100755
--- a/server/voice_changer/VoiceChanger.py
+++ b/server/voice_changer/VoiceChanger.py
@@ -112,8 +112,25 @@ class VoiceChanger():
                     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()
+                    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)
+
+                    if self.prev_strength.device != torch.device('cpu'):
+                        print(f"prev_strength move from {self.prev_strength.device} to cpu")
+                        self.prev_strength = self.prev_strength.cpu()
+                    if self.cur_strength.device != torch.device('cpu'):
+                        print(f"cur_strength move from {self.cur_strength.device} to cpu")
+                        self.cur_strength = self.cur_strength.cpu()
+
+                    if hasattr(self, 'prev_audio1') == True and self.prev_audio1.device == torch.device('cpu'):
+                        prev = self.prev_audio1[-1*unpackedData.shape[0]:]
+                        cur  = audio1[-2*unpackedData.shape[0]:-1*unpackedData.shape[0]]
+                        result = prev * self.prev_strength + cur * self.cur_strength
+                    else:
+                        cur = audio1[-2*unpackedData.shape[0]:-1*unpackedData.shape[0]]
+                        result = cur
+
+                    self.prev_audio1 = audio1
+                    result = result.cpu().float().numpy()
             # elif self.mps_enabled == True: # MPS doesnt support aten::weight_norm_interface, and PYTORCH_ENABLE_MPS_FALLBACK=1 cause a big dely.
             #         x, x_lengths, spec, spec_lengths, y, y_lengths, sid_src = [
             #             x.to("mps") for x in data]
@@ -137,15 +154,15 @@ class VoiceChanger():
 
 
 
-                    if hasattr(self, 'prev_audio1') == True:
+                    if hasattr(self, 'prev_audio1') == True and self.prev_audio1.device == torch.device('cuda', gpu):
                         prev = self.prev_audio1[-1*unpackedData.shape[0]:]
                         cur  = audio1[-2*unpackedData.shape[0]:-1*unpackedData.shape[0]]
                         result = prev * self.prev_strength + cur * self.cur_strength
-                        # print("merging...", prev.shape, cur.shape)
+                        print("merging...", prev.shape, cur.shape)
                     else:
                         cur = audio1[-2*unpackedData.shape[0]:-1*unpackedData.shape[0]]
                         result = cur
-                        # print("no merging...", cur.shape)
+                        print("no merging...", cur.shape)
                     self.prev_audio1 = audio1
 
                     #print(result)