cleanup/rewrite

data.py

@@ -0,0 +1,65 @@
+import librosa
+import numpy as np
+from pathlib import Path
+from multiprocessing import Pool, cpu_count
+
+SAMPLE_RATE = 22050
+
+def process_file(file_path):
+    """
+    Load 10 second chunks single song.
+    """
+    y, sr = librosa.load(file_path, mono=True, sr=SAMPLE_RATE)
+    size = int(SAMPLE_RATE * 10)
+    sample_len = len(y)
+
+    file_chunks = []
+    for start_pos in range(0, sample_len, size):
+        end = start_pos + size
+        if end <= sample_len:
+            chunk = y[start_pos:end]
+            file_chunks.append(chunk)
+    return file_chunks
+
+def load():
+    """
+    Load 10 second chunks of songs.
+    """
+    audio = []
+    files = list(Path("./data/").glob("*.mp3"))
+    with Pool(cpu_count()) as pool:
+        chunk_list = pool.map(process_file, files)
+    for l in chunk_list:
+        audio.extend(l)
+    return audio
+
+def audio_split(audio):
+    """
+    Split 10 seconds of audio to 2 5 second clips
+    """
+    size = int(SAMPLE_RATE*5)
+    x = audio[:size]
+    y = audio[size:size*2]
+
+    x = librosa.feature.melspectrogram(y=x, sr=SAMPLE_RATE)
+    y = librosa.feature.melspectrogram(y=y, sr=SAMPLE_RATE)
+
+    ma,mi = x.max(), x.min()
+    x = (x - mi) / (ma - mi)
+            
+    ma,mi = y.max(), y.min()
+    y = (y - mi) / (ma - mi)
+    
+    return x,y
+
+def detaset(chunks):
+    """
+    convert 10 second chunks to dataset
+    """
+    x,y=[],[]
+    with Pool(cpu_count()) as pool:
+        audio_list = pool.map(audio_split,chunks)
+    for (ax,ay) in audio_list:
+        x.append(ax)
+        y.append(ay)
+    return x,y