pca version

tasks/audio.py

@@ -18,23 +18,25 @@ router = APIRouter()
 
 DESCRIPTION = "Random Baseline"
 ROUTE = "/audio"
-def create_spec(dataset):
+
+def create_spec(dataset, target_sampling_rate=3000):
     spectograms = []
-    
+    audio_length = int(36000/(12000/target_sampling_rate))
     for d in dataset:  
-        audio_sample = d["audio"]["array"] if d["audio"]["sampling_rate"] == 12000 else librosa.resample(
-                d["audio"]["array"], 
-                orig_sr= d["audio"]["sampling_rate"], 
-                target_sr=12000
-            )
+        audio_sample = librosa.resample(
+            d["audio"]["array"], 
+            orig_sr= d["audio"]["sampling_rate"], 
+            target_sr=target_sampling_rate
+        )
+        
         if len(audio_sample) == 0:
             continue     
-        if len(audio_sample) < 36000:
-            padding_needed = 36000 - len(audio_sample)
+        if len(audio_sample) < audio_length:
+            padding_needed = audio_length - len(audio_sample)
             repeats = (padding_needed // len(audio_sample)) + 1
-            audio_sample = np.concatenate([audio_sample] + [audio_sample[:padding_needed]] * repeats)[:36000]
-        elif len(audio_sample) > 36000:
-            audio_sample = audio_sample[:36000]
+            audio_sample = np.concatenate([audio_sample] + [audio_sample[:padding_needed]] * repeats)[:audio_length]
+        elif len(audio_sample) > audio_length:
+            audio_sample = audio_sample[:audio_length]
         
         rms = np.sqrt(np.mean(np.square(audio_sample)))
         scalar = 10 ** (-20 / 20) / (rms + 1e-8)
@@ -43,13 +45,13 @@ def create_spec(dataset):
             y=audio_sample*scalar,
             sr=12000,
             n_fft=2048,
-            hop_length=512,
-            n_mels=32,
+            hop_length=1024,
+            n_mels=12,
             power=2.0,
         )
         mel_db = librosa.power_to_db(mel, ref=np.max)  
         mel_db_normalized = (mel_db - mel_db.mean()) / (mel_db.std() + 1e-8)
-        spectograms.append(np.float16(mel_db_normalized).T.flatten())
+        spectograms.append(mel_db_normalized.T.flatten())
         
     return np.stack(spectograms)
 
@@ -88,6 +90,11 @@ async def evaluate_audio(request: AudioEvaluationRequest):
     H = np.load("H.npy")
     W_test = np.dot(test_spec, H)
     model = joblib.load('model.joblib')
+    
+    test_spec = create_spec(test_dataset)
+    model = joblib.load("3000_40_100/model.joblib")
+    H = np.load("3000_40_100/H.npy")
+    W_test = np.dot(test_spec, H)
     #--------------------------------------------------------------------------------------------
     # YOUR MODEL INFERENCE CODE HERE
     # Update the code below to replace the random baseline by your model inference within the inference pass where the energy consumption and emissions are tracked.