Update utils/mir_eval_modules.py

utils/mir_eval_modules.py

@@ -26,64 +26,64 @@ def idx2voca_chord():
         idx2voca_chord[i] = chord
     return idx2voca_chord
 
-def audio_file_to_features(audio_file, config):
-    original_wav, sr = librosa.load(audio_file, sr=config.mp3['song_hz'], mono=True)
-
-    # Total duration in seconds
-    total_duration = librosa.get_duration(y=original_wav, sr=sr)
-
-    # Define segment length in seconds (you can still use config.mp3['inst_len'] for the segment size)
-    segment_len_sec = config.mp3.get('inst_len', 10)  # default to 10s if not defined
-
-    currunt_sec_hz = 0
-    feature = None
-
-    while currunt_sec_hz + sr * segment_len_sec < len(original_wav):
-        start_idx = int(currunt_sec_hz)
-        end_idx = int(currunt_sec_hz + sr * segment_len_sec)
-        tmp = librosa.cqt(
-            original_wav[start_idx:end_idx], sr=sr,
-            n_bins=config.feature['n_bins'],
-            bins_per_octave=config.feature['bins_per_octave'],
-            hop_length=config.feature['hop_length']
-        )
-        feature = tmp if feature is None else np.concatenate((feature, tmp), axis=1)
-        currunt_sec_hz = end_idx
-
-    # Handle the final chunk (remaining audio)
-    tmp = librosa.cqt(
-        original_wav[currunt_sec_hz:], sr=sr,
-        n_bins=config.feature['n_bins'],
-        bins_per_octave=config.feature['bins_per_octave'],
-        hop_length=config.feature['hop_length']
-    )
-    feature = np.concatenate((feature, tmp), axis=1)
-    feature = np.log(np.abs(feature) + 1e-6)
+# def audio_file_to_features(audio_file, config):
+#     original_wav, sr = librosa.load(audio_file, sr=config.mp3['song_hz'], mono=True)
 
-    feature_per_second = segment_len_sec / config.model['timestep']
-    song_length_second = len(original_wav) / sr
+#     # Total duration in seconds
+#     total_duration = librosa.get_duration(y=original_wav, sr=sr)
 
-    return feature, feature_per_second, song_length_second
+#     # Define segment length in seconds (you can still use config.mp3['inst_len'] for the segment size)
+#     segment_len_sec = config.mp3.get('inst_len', 10)  # default to 10s if not defined
 
-# def audio_file_to_features(audio_file, config):
-#     original_wav, sr = librosa.load(audio_file, sr=config.mp3['song_hz'], mono=True)
 #     currunt_sec_hz = 0
-#     while len(original_wav) > currunt_sec_hz + config.mp3['song_hz'] * config.mp3['inst_len']:
+#     feature = None
+
+#     while currunt_sec_hz + sr * segment_len_sec < len(original_wav):
 #         start_idx = int(currunt_sec_hz)
-#         end_idx = int(currunt_sec_hz + config.mp3['song_hz'] * config.mp3['inst_len'])
-#         tmp = librosa.cqt(original_wav[start_idx:end_idx], sr=sr, n_bins=config.feature['n_bins'], bins_per_octave=config.feature['bins_per_octave'], hop_length=config.feature['hop_length'])
-#         if start_idx == 0:
-#             feature = tmp
-#         else:
-#             feature = np.concatenate((feature, tmp), axis=1)
+#         end_idx = int(currunt_sec_hz + sr * segment_len_sec)
+#         tmp = librosa.cqt(
+#             original_wav[start_idx:end_idx], sr=sr,
+#             n_bins=config.feature['n_bins'],
+#             bins_per_octave=config.feature['bins_per_octave'],
+#             hop_length=config.feature['hop_length']
+#         )
+#         feature = tmp if feature is None else np.concatenate((feature, tmp), axis=1)
 #         currunt_sec_hz = end_idx
-#     tmp = librosa.cqt(original_wav[currunt_sec_hz:], sr=sr, n_bins=config.feature['n_bins'], bins_per_octave=config.feature['bins_per_octave'], hop_length=config.feature['hop_length'])
+
+#     # Handle the final chunk (remaining audio)
+#     tmp = librosa.cqt(
+#         original_wav[currunt_sec_hz:], sr=sr,
+#         n_bins=config.feature['n_bins'],
+#         bins_per_octave=config.feature['bins_per_octave'],
+#         hop_length=config.feature['hop_length']
+#     )
 #     feature = np.concatenate((feature, tmp), axis=1)
 #     feature = np.log(np.abs(feature) + 1e-6)
-#     feature_per_second = config.mp3['inst_len'] / config.model['timestep']
-#     song_length_second = len(original_wav)/config.mp3['song_hz']
+
+#     feature_per_second = segment_len_sec / config.model['timestep']
+#     song_length_second = len(original_wav) / sr
+
 #     return feature, feature_per_second, song_length_second
 
+def audio_file_to_features(audio_file, config):
+    original_wav, sr = librosa.load(audio_file, sr=config.mp3['song_hz'], mono=True)
+    currunt_sec_hz = 0
+    while len(original_wav) > currunt_sec_hz + config.mp3['song_hz'] * config.mp3['inst_len']:
+        start_idx = int(currunt_sec_hz)
+        end_idx = int(currunt_sec_hz + config.mp3['song_hz'] * config.mp3['inst_len'])
+        tmp = librosa.cqt(original_wav[start_idx:end_idx], sr=sr, n_bins=config.feature['n_bins'], bins_per_octave=config.feature['bins_per_octave'], hop_length=config.feature['hop_length'])
+        if start_idx == 0:
+            feature = tmp
+        else:
+            feature = np.concatenate((feature, tmp), axis=1)
+        currunt_sec_hz = end_idx
+    tmp = librosa.cqt(original_wav[currunt_sec_hz:], sr=sr, n_bins=config.feature['n_bins'], bins_per_octave=config.feature['bins_per_octave'], hop_length=config.feature['hop_length'])
+    feature = np.concatenate((feature, tmp), axis=1)
+    feature = np.log(np.abs(feature) + 1e-6)
+    feature_per_second = config.mp3['inst_len'] / config.model['timestep']
+    song_length_second = len(original_wav)/config.mp3['song_hz']
+    return feature, feature_per_second, song_length_second
+
 # Audio files with format of wav and mp3
 def get_audio_paths(audio_dir):
     return [os.path.join(root, fname) for (root, dir_names, file_names) in os.walk(audio_dir, followlinks=True)