Merge remote-tracking branch 'origin/main'

configs/base.yaml

@@ -57,7 +57,7 @@ max_batch_size: 8
 max_batch_frames: 80000
 max_val_batch_size: 1
 max_val_batch_frames: 10000
-num_valid_plots: 10
+num_valid_plots: 100
 log_interval: 100
 num_sanity_val_steps: 1  # steps of validation at the beginning
 val_check_interval: 1000

configs/continuous.yaml

@@ -9,7 +9,8 @@ binarization_args:
   num_workers: 0
   merge_rest: true  # merge continuous rest notes
   merge_slur: true  # merge slurs with the same pitch
-  round_midi: true  # round midi value
+  round_midi: false  # round midi value
+  slur_tolerance: 0.5  # maximum allowed value of pitch change of a slur to be merged
 
 key_shift_factor: 8
 key_shift_range: [-12, 12]

configs/discrete.yaml

@@ -8,6 +8,8 @@ binary_data_dir: data/some_ds_quant_spk4_aug8/binary
 binarization_args:
   num_workers: 0
   shuffle: true
+  merge_rest: true  # merge continuous rest notes
+  merge_slur: true  # merge slurs with the same pitch
 use_bound_loss: true
 use_midi_loss: true
 
@@ -18,7 +20,8 @@ pe: rmvpe
 pe_ckpt: pretrained/rmvpe/model.pt
 
 # global constants
-midi_shift_range: [-6, 6]
+key_shift_range: [-12, 12]
+key_shift_factor: 8
 
 # neural networks
 units_dim: 80  # 768

configs/quant_tow_head_model.yaml

@@ -3,9 +3,6 @@ base_config:
 
 binary_data_dir: data/some_ds_quant_spk4_aug8/binary
 
-# global constants
-midi_shift_range: [-6, 6]
-
 # neural networks
 units_dim: 80  # 768
 midi_num_bins: 129  # rest = 128

configs/some_bound.yaml

@@ -16,7 +16,6 @@ pe_ckpt: pretrained/rmvpe/model.pt
 
 # global constants
 midi_prob_deviation: 0.5
-midi_shift_range: [-6, 6]
 rest_threshold: 0.1
 
 # neural networks

preprocessing/me_binarizer.py

@@ -41,6 +41,7 @@ def __init__(self, config: dict):
         self.lr = LengthRegulator().to(self.device)
         self.merge_rest = self.binarization_args['merge_rest']
         self.merge_slur = self.binarization_args['merge_slur']
+        self.slur_tolerance = self.binarization_args['slur_tolerance']
         self.round_midi = self.binarization_args['round_midi']
         self.key_shift_min, self.key_shift_max = self.config['key_shift_range']
 
@@ -82,7 +83,7 @@ def load_meta_data(self, raw_data_dir: pathlib.Path, ds_id):
 
                 if self.merge_slur:
                     # merge slurs with the same pitch
-                    note_seq, note_dur = merge_slurs(note_seq, note_dur, note_slur)
+                    note_seq, note_dur = merge_slurs(note_seq, note_dur, note_slur, tolerance=self.slur_tolerance)
 
                 if self.merge_rest:
                     # merge continuous rest notes

utils/binarizer_utils.py

@@ -1,22 +1,52 @@
 from typing import Tuple
 
+import librosa
 import numpy as np
 import parselmouth
 import torch
 
 from utils.pitch_utils import interp_f0
 
 
-def merge_slurs(note_seq: list, note_dur: list, note_slur: list) -> Tuple[list, list]:
-    # merge slurs with the same pitch
-    note_seq_merge_slur = [note_seq[0]]
+def merge_slurs(note_seq: list, note_dur: list, note_slur: list, tolerance=None) -> Tuple[list, list]:
+    """
+    merge slurs with the same pitch
+    """
+    note_midi = [librosa.note_to_midi(n, round_midi=False) if n != 'rest' else 'rest' for n in note_seq]
+    prev_min = prev_max = None
+    note_midi_merge_slur = [note_midi[0]]
     note_dur_merge_slur = [note_dur[0]]
+
+    def can_be_merged(midi):
+        if tolerance is None or midi == 'rest' or note_midi_merge_slur[-1] == 'rest':
+            return note_midi_merge_slur[-1] == midi
+        return (
+            abs(midi - note_midi_merge_slur[-1]) <= tolerance
+            and (prev_min is None or abs(midi - prev_min) <= tolerance)
+            and (prev_max is None or abs(midi - prev_max) <= tolerance)
+        )
+
+    def get_merged_midi(midi1, dur1, midi2, dur2):
+        if midi1 == midi2:
+            return midi1
+        return (midi1 * dur1 + midi2 * dur2) / (dur1 + dur2)
+
     for i in range(1, len(note_seq)):
-        if note_slur[i] and note_seq[i] == note_seq[i - 1]:
+        if note_slur[i] and can_be_merged(note_midi[i]):
+            # update min and max
+            prev_min = min(note_midi[i], note_midi_merge_slur[-1]) if prev_min is None else min(prev_min, note_midi[i])
+            prev_max = max(note_midi[i], note_midi_merge_slur[-1]) if prev_max is None else max(prev_max, note_midi[i])
+            note_midi_merge_slur[-1] = get_merged_midi(
+                note_midi_merge_slur[-1], note_dur_merge_slur[-1], note_midi[i], note_dur[i]
+            )
             note_dur_merge_slur[-1] += note_dur[i]
         else:
-            note_seq_merge_slur.append(note_seq[i])
+            note_midi_merge_slur.append(note_midi[i])
             note_dur_merge_slur.append(note_dur[i])
+            prev_min = prev_max = None
+    note_seq_merge_slur = [
+        librosa.midi_to_note(n, cents=True, unicode=False) if n != 'rest' else 'rest' for n in note_midi_merge_slur
+    ]
     return note_seq_merge_slur, note_dur_merge_slur