Add pretrained model

.gitignore

@@ -3,4 +3,29 @@ __pycache__
 .idea
 data
 runs
+output
 
+.vscode
+
+# Mac
+.DS_Store
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST

README.md

@@ -55,6 +55,12 @@ In order to test on the Maestro dataset's test split instead of the MAPS databas
 python evaluate.py runs/model/model-100000.pt Maestro test
 ```
 
+You can download a pretrained model [here](https://drive.google.com/file/d/1Mj2Em07Lvl3mvDQCCxOYHjPiB-S0WGT1/view?usp=sharing) and run `transcribe.py` to transcribe piano audio files:
+
+```bash
+python transcribe.py model-500000.pt <path to audio files> --save-path output/
+```
+
 ## Implementation Details
 
 This implementation contains a few of the additional improvements on the model that were reported in the Maestro paper, including:
@@ -71,5 +77,3 @@ Meanwhile, this implementation does not include the following features:
 * Harmonically decaying weights on the frame loss
 
 Despite these, this implementation is able to achieve a comparable performance to what is reported on the Maestro paper as the performance without data augmentation.
-
-

onsets_and_frames/midi.py

@@ -2,9 +2,10 @@
 import sys
 
 import mido
+import pretty_midi
 import numpy as np
+import collections
 from joblib import Parallel, delayed
-from mido import Message, MidiFile, MidiTrack
 from mir_eval.util import hz_to_midi
 from tqdm import tqdm
 
@@ -60,28 +61,32 @@ def save_midi(path, pitches, intervals, velocities):
     intervals: list of (onset_index, offset_index)
     velocities: list of velocity values
     """
-    file = MidiFile()
-    track = MidiTrack()
-    file.tracks.append(track)
-    ticks_per_second = file.ticks_per_beat * 2.0
+    file = pretty_midi.PrettyMIDI()
+    piano_program = pretty_midi.instrument_name_to_program('Acoustic Grand Piano')
+    piano = pretty_midi.Instrument(program=piano_program)
 
-    events = []
+    # Remove overlapping intervals (end time should be smaller of equal start time of next note on the same pitch)
+    intervals_dict = collections.defaultdict(list)
     for i in range(len(pitches)):
-        events.append(dict(type='on', pitch=pitches[i], time=intervals[i][0], velocity=velocities[i]))
-        events.append(dict(type='off', pitch=pitches[i], time=intervals[i][1], velocity=velocities[i]))
-    events.sort(key=lambda row: row['time'])
-
-    last_tick = 0
-    for event in events:
-        current_tick = int(event['time'] * ticks_per_second)
-        velocity = int(event['velocity'] * 127)
-        if velocity > 127:
-            velocity = 127
-        pitch = int(round(hz_to_midi(event['pitch'])))
-        track.append(Message('note_' + event['type'], note=pitch, velocity=velocity, time=current_tick - last_tick))
-        last_tick = current_tick
-
-    file.save(path)
+        pitch = int(round(hz_to_midi(pitches[i])))
+        intervals_dict[pitch].append((intervals[i], i))
+    for pitch in intervals_dict:
+        interval_list = intervals_dict[pitch]
+        interval_list.sort(key=lambda x: x[0][0])
+        for i in range(len(interval_list) - 1):
+            # assert interval_list[i][1] <= interval_list[i+1][0], f'End time should be smaller of equal start time of next note on the same pitch. It was {interval_list[i][1]}, {interval_list[i+1][0]} for pitch {key}'
+            interval_list[i][0][1] = min(interval_list[i][0][1], interval_list[i+1][0][0])
+
+    for pitch in intervals_dict:
+        interval_list = intervals_dict[pitch]
+        for interval,i in interval_list:
+            pitch = int(round(hz_to_midi(pitches[i])))
+            velocity = int(127*min(velocities[i], 1))
+            note = pretty_midi.Note(velocity=velocity, pitch=pitch, start=interval[0], end=interval[1])
+            piano.notes.append(note)
+
+    file.instruments.append(piano)
+    file.write(path)
 
 
 if __name__ == '__main__':