Modification:

- Timer update
  - Diffusion SVC Performance monitor

server/voice_changer/DiffusionSVC/inferencer/DiffusionSVCInferencer.py

@@ -107,18 +107,19 @@ class DiffusionSVCInferencer(Inferencer):
         silence_front: float,
         skip_diffusion: bool = True,
     ) -> torch.Tensor:
-        with Timer2("pre-process", False) as t:
+        use_timer = False
+        with Timer2("    Naive", use_timer) as t:
             gt_spec = self.naive_model_call(feats, pitch, volume, spk_id=sid, spk_mix_dict=None, aug_shift=0, spk_emb=None)
 
         # print("[    ----Timer::1: ]", t.secs)
 
-        with Timer2("pre-process", False) as t:
+        with Timer2("    Diffuser", use_timer) as t:
             if skip_diffusion == 0:
                 out_mel = self.__call__(feats, pitch, volume, spk_id=sid, spk_mix_dict=None, aug_shift=0, gt_spec=gt_spec, infer_speedup=infer_speedup, method="dpm-solver", k_step=k_step, use_tqdm=False, spk_emb=None)
                 gt_spec = out_mel
         # print("[    ----Timer::2: ]", t.secs)
 
-        with Timer2("pre-process", False) as t:  # NOQA
+        with Timer2("    Vocoder", use_timer) as t:  # NOQA
             if self.vocoder_onnx is None:
                 start_frame = int(silence_front * self.vocoder.vocoder_sample_rate / self.vocoder.vocoder_hop_size)
                 out_wav = self.mel2wav(gt_spec, pitch, start_frame=start_frame)

server/voice_changer/DiffusionSVC/pipeline/Pipeline.py

@@ -102,8 +102,9 @@ class Pipeline(object):
         protect=0.5,
         skip_diffusion=True,
     ):
+        use_timer = False
         # print("---------- pipe line --------------------")
-        with Timer2("pre-process", False) as t:
+        with Timer2("pre-process", use_timer) as t:
             audio_t = torch.from_numpy(audio).float().unsqueeze(0).to(self.device)
             audio16k = self.resamplerIn(audio_t)
             volume, mask = self.extract_volume_and_mask(audio16k, threshold=-60.0)
@@ -111,7 +112,7 @@ class Pipeline(object):
             n_frames = int(audio16k.size(-1) // self.hop_size + 1)
         # print("[Timer::1: ]", t.secs)
 
-        with Timer2("pre-process", False) as t:
+        with Timer2("extract pitch", use_timer) as t:
             # ピッチ検出
             try:
                 # pitch = self.pitchExtractor.extract(
@@ -141,7 +142,7 @@ class Pipeline(object):
             feats = feats.view(1, -1)
         # print("[Timer::2: ]", t.secs)
 
-        with Timer2("pre-process", False) as t:
+        with Timer2("extract feature", use_timer) as t:
             # embedding
             with autocast(enabled=self.isHalf):
                 try:
@@ -158,7 +159,7 @@ class Pipeline(object):
             feats = F.interpolate(feats.permute(0, 2, 1), size=int(n_frames), mode="nearest").permute(0, 2, 1)
         # print("[Timer::3: ]", t.secs)
 
-        with Timer2("pre-process", False) as t:
+        with Timer2("infer", use_timer) as t:
             # 推論実行
             try:
                 with torch.no_grad():
@@ -179,7 +180,7 @@ class Pipeline(object):
                     raise e
         # print("[Timer::4: ]", t.secs)
 
-        with Timer2("pre-process", False) as t:  # NOQA
+        with Timer2("post-process", use_timer) as t:  # NOQA
             feats_buffer = feats.squeeze(0).detach().cpu()
             if pitch is not None:
                 pitch_buffer = pitch.squeeze(0).detach().cpu()

server/voice_changer/utils/Timer.py

@@ -72,6 +72,7 @@ class Timer2(object):
         if self.enable is False:
             return self
         self.current = time.time()
+        self.start = time.time()
         return self
 
     def record(self, lapname: str):
@@ -88,9 +89,11 @@ class Timer2(object):
     def __exit__(self, *_):
         if self.enable is False:
             return
+        self.end = time.time()
+        self.elapsed = (self.end - self.start) * 1000
         title = self.key.split("_")[-1]
-        print(f"---- {title} ----")
+        print(f"---- {title}(elapsed:{round(self.elapsed,1)}ms) ----")
         for key, val in self.storedSecs[self.key].items():
             section = key.split("_")[-1]
             milisecAvr = sum(val) / len(val) * 1000
-            print(f"{section}: {milisecAvr} msec")
+            print(f"{section}: {round(milisecAvr,1)} msec, {val[-1]}")