includes converting lab file to midi file

Author: jayg996

README.md

@@ -12,6 +12,7 @@ This repository has the source codes for the paper "A Bi-Directional Transformer
 - librosa >= 0.6.3
 - pyyaml >= 3.13
 - mir_eval >= 0.5
+- pretty_midi >= 0.2.8
 
 ## File descriptions
   * `audio_dataset.py` : loads data and preprocesses label files to chord labels and mp3 files to constant-q transformation. 
@@ -33,7 +34,7 @@ $ python test.py --audio_dir audio_folder --save_dir save_folder --voca False
   * save_dir : a forder for saving recognition results (default: './test')
   * voca : False means major and minor label type, and True means large vocabulary label type (default: False)
 
-The resulting files are lab files of the form shown below.
+The resulting files are lab files of the form shown below and midi files.
 
   <img src="png/example.png">
 

png/example.png

@@ -1,4 +1,6 @@
 import os
+import mir_eval
+import pretty_midi as pm
 from utils import logger
 from btc_model import *
 from utils.mir_eval_modules import audio_file_to_features, idx2chord, idx2voca_chord, get_audio_paths
@@ -12,7 +14,7 @@
 
 # hyperparameters
 parser = argparse.ArgumentParser()
-parser.add_argument('--voca', default=False, type=lambda x: (str(x).lower() == 'true'))
+parser.add_argument('--voca', default=True, type=lambda x: (str(x).lower() == 'true'))
 parser.add_argument('--audio_dir', type=str, default='./test')
 parser.add_argument('--save_dir', type=str, default='./test')
 args = parser.parse_args()
@@ -75,12 +77,12 @@
                     continue
                 if prediction[i].item() != prev_chord:
                     lines.append(
-                        '%.6f %.6f %s\n' % (start_time, time_unit * (n_timestep * t + i), idx_to_chord[prev_chord]))
+                        '%.3f %.3f %s\n' % (start_time, time_unit * (n_timestep * t + i), idx_to_chord[prev_chord]))
                     start_time = time_unit * (n_timestep * t + i)
                     prev_chord = prediction[i].item()
                 if t == num_instance - 1 and i + num_pad == n_timestep:
                     if start_time != time_unit * (n_timestep * t + i):
-                        lines.append('%.6f %.6f %s\n' % (start_time, time_unit * (n_timestep * t + i), idx_to_chord[prev_chord]))
+                        lines.append('%.3f %.3f %s\n' % (start_time, time_unit * (n_timestep * t + i), idx_to_chord[prev_chord]))
                     break
 
     # lab file write
@@ -93,3 +95,36 @@
 
     logger.info("label file saved : %s" % save_path)
 
+    # lab file to midi file
+    
+
+    starts, ends, pitchs = list(), list(), list()
+
+    intervals, chords = mir_eval.io.load_labeled_intervals(save_path)
+    for p in range(12):
+        for i, (interval, chord) in enumerate(zip(intervals, chords)):
+            root_num, relative_bitmap, _ = mir_eval.chord.encode(chord)
+            tmp_label = mir_eval.chord.rotate_bitmap_to_root(relative_bitmap, root_num)[p]
+            if i == 0:
+                start_time = interval[0]
+                label = tmp_label
+                continue
+            if tmp_label != label:
+                if label == 1.0:
+                    starts.append(start_time), ends.append(interval[0]), pitchs.append(p + 48)
+                start_time = interval[0]
+                label = tmp_label
+            if i == (len(intervals) - 1): 
+                if label == 1.0:
+                    starts.append(start_time), ends.append(interval[1]), pitchs.append(p + 48)
+
+    midi = pm.PrettyMIDI()
+    instrument = pm.Instrument(program=0)
+
+    for start, end, pitch in zip(starts, ends, pitchs):
+        pm_note = pm.Note(velocity=120, pitch=pitch, start=start, end=end)
+        instrument.notes.append(pm_note)
+
+    midi.instruments.append(instrument)
+    midi.write(save_path.replace('.lab', '.midi'))    
+