update timer

server/voice_changer/RVC/pipeline/Pipeline.py

@@ -18,7 +18,7 @@ from voice_changer.RVC.inferencer.OnnxRVCInferencer import OnnxRVCInferencer
 from voice_changer.RVC.inferencer.OnnxRVCInferencerNono import OnnxRVCInferencerNono
 
 from voice_changer.RVC.pitchExtractor.PitchExtractor import PitchExtractor
-from voice_changer.utils.Timer import Timer
+from voice_changer.utils.Timer import Timer2
 
 logger = VoiceChangaerLogger.get_instance().getLogger()
 
@@ -153,7 +153,7 @@ class Pipeline(object):
         # print(f"pipeline exec input, audio:{audio.shape}, pitchf:{pitchf.shape}, feature:{feature.shape}")
         # print(f"pipeline exec input, silence_front:{silence_front}, out_size:{out_size}")
 
-        with Timer("main-process", False) as t:  # NOQA
+        with Timer2("Pipeline-Exec", False) as t:  # NOQA
             # 16000のサンプリングレートで入ってきている。以降この世界は16000で処理。
             search_index = self.index is not None and self.big_npy is not None and index_rate != 0
             # self.t_pad = self.sr * repeat  # 1秒
@@ -180,27 +180,14 @@ class Pipeline(object):
             assert feats.dim() == 1, feats.dim()
             feats = feats.view(1, -1)
 
+            t.record("pre-process")
             # ピッチ検出
-            with Timer("main-process", True) as t:
+            pitch, pitchf = self.extractPitch(audio_pad, if_f0, pitchf, f0_up_key, silence_front)
-                pitch, pitchf = self.extractPitch(audio_pad, if_f0, pitchf, f0_up_key, silence_front)
+            t.record("extract-pitch")
-            print(f"[Perform(Pit)] {t.secs}")
 
             # embedding
-            with Timer("main-process", True) as t:
+            feats = self.extractFeatures(feats, embOutputLayer, useFinalProj)
-                # with autocast(enabled=self.isHalf):
+            t.record("extract-feats")
-                #     try:
-                #         feats = self.embedder.extractFeatures(feats, embOutputLayer, useFinalProj)
-                #         if torch.isnan(feats).all():
-                #             raise DeviceCannotSupportHalfPrecisionException()
-                #     except RuntimeError as e:
-                #         if "HALF" in e.__str__().upper():
-                #             raise HalfPrecisionChangingException()
-                #         elif "same device" in e.__str__():
-                #             raise DeviceChangingException()
-                #         else:
-                #             raise e
-                feats = self.extractFeatures(feats, embOutputLayer, useFinalProj)
-            print(f"[Perform(Emb)] {t.secs}")
 
             # Index - feature抽出
             # if self.index is not None and self.feature is not None and index_rate != 0:
@@ -268,10 +255,10 @@ class Pipeline(object):
                 pitchf = pitchf[:, -feats_len:]
             p_len = torch.tensor([feats_len], device=self.device).long()
 
+            t.record("mid-precess")
             # 推論実行
-            with Timer("main-process", True) as t:
+            audio1 = self.infer(feats, p_len, pitch, pitchf, sid, out_size)
-                audio1 = self.infer(feats, p_len, pitch, pitchf, sid, out_size)
+            t.record("infer")
-            print(f"[Perform(Inf)] {t.secs}")
 
             feats_buffer = feats.squeeze(0).detach().cpu()
             if pitchf is not None:
@@ -290,6 +277,7 @@ class Pipeline(object):
                 audio1 = audio1[offset:end]
 
             del sid
+            t.record("post-process")
             # torch.cuda.empty_cache()
         # print("EXEC AVERAGE:", t.avrSecs)
         return audio1, pitchf_buffer, feats_buffer

server/voice_changer/utils/Timer.py

@@ -41,3 +41,56 @@ class Timer(object):
         self.storedSecs[self.key].append(self.secs)
         self.storedSecs[self.key] = self.storedSecs[self.key][-self.maxStores:]
         self.avrSecs = sum(self.storedSecs[self.key]) / len(self.storedSecs[self.key])
+
+
+class Timer2(object):
+    storedSecs = {}  # Class variable
+
+    def __init__(self, title: str, enalbe: bool = True):
+        self.title = title
+        self.enable = enalbe
+        self.secs = 0
+        self.msecs = 0
+        self.avrSecs = 0
+
+        if self.enable is False:
+            return
+
+        self.maxStores = 10
+
+        current_frame = inspect.currentframe()
+        caller_frame = inspect.getouterframes(current_frame, 2)
+        frame = caller_frame[1]
+        filename = frame.filename
+        line_number = frame.lineno
+        self.key = f"{filename}_{line_number}_{title}"
+        if self.key not in self.storedSecs:
+            self.storedSecs[self.key] = {}
+
+    def __enter__(self):
+        if self.enable is False:
+            return self
+        self.current = time.time()
+        return self
+
+    def record(self, lapname: str):
+        if self.enable is False:
+            return
+        self.lapkey = f"{self.key}_{lapname}"
+        prev = self.current
+        self.current = time.time()
+        if self.lapkey not in self.storedSecs[self.key]:
+            self.storedSecs[self.key][self.lapkey] = []
+        self.storedSecs[self.key][self.lapkey].append(self.current - prev)
+        self.storedSecs[self.key][self.lapkey] = self.storedSecs[self.key][self.lapkey][-self.maxStores:]
+
+    def __exit__(self, *_):
+        if self.enable is False:
+            return
+        title = self.key.split("_")[-1]
+        print(f"---- {title} ----")
+        for key, val in self.storedSecs[self.key].items():
+            section = key.split("_")[-1]
+            milisecAvr = sum(val) / len(val) * 1000
+            print(f"{section}: {milisecAvr} msec")
+