Fixing split methods for empty files (no tracks and/or no notes) (#177)

* fixing split methods for empty files (no tracks and/or no notes)

* fix

miditok/utils/split_utils.py

@@ -8,14 +8,16 @@
 from typing import TYPE_CHECKING, Any
 from warnings import warn
 
-from symusic import Score, TextMeta
+from symusic import Score, TextMeta, TimeSignature
+from symusic.core import TimeSignatureTickList
 from tqdm import tqdm
 
 from miditok.constants import (
     MAX_NUM_FILES_NUM_TOKENS_PER_NOTE,
     MIDI_FILES_EXTENSIONS,
     SCORE_LOADING_EXCEPTION,
     SUPPORTED_MUSIC_FILE_EXTENSIONS,
+    TIME_SIGNATURE,
 )
 
 from .utils import (
@@ -112,7 +114,10 @@ def split_files_for_training(
         except SCORE_LOADING_EXCEPTION:
             continue
 
-        # Separate track first if needed
+        # First preprocess time signatures to avoid cases where they might cause errors
+        _preprocess_time_signatures(scores[0], tokenizer)
+
+        # Separate track if needed
         tracks_separated = False
         if not tokenizer.one_token_stream and len(scores[0].tracks) > 1:
             scores = split_score_per_tracks(scores[0])
@@ -308,6 +313,9 @@ def get_average_num_tokens_per_note(
                 if (num_notes := track.note_num()) > 0:
                     num_tokens_per_note.append(len(seq) / num_notes)
 
+    if len(num_tokens_per_note) == 0:
+        msg = "All the music files provided are empty and contain no note."
+        raise ValueError(msg)
     return sum(num_tokens_per_note) / len(num_tokens_per_note)
 
 
@@ -381,3 +389,20 @@ def split_tokens_files_to_subsequences(
                 new_tok = deepcopy(tokens)
                 new_tok["ids"] = subseq
                 json.dump(tokens, outfile)
+
+
+def _preprocess_time_signatures(score: Score, tokenizer: MusicTokenizer) -> None:
+    """
+    Make sure a Score contains time signature valid according to a tokenizer.
+
+    :param score: ``symusic.Score`` to preprocess the time signature.
+    :param tokenizer: :class:`miditok.MusicTokenizer`.
+    """
+    if tokenizer.config.use_time_signatures:
+        tokenizer._filter_unsupported_time_signatures(score.time_signatures)
+        if len(score.time_signatures) == 0 or score.time_signatures[0].time != 0:
+            score.time_signatures.insert(0, TimeSignature(0, *TIME_SIGNATURE))
+    else:
+        score.time_signatures = TimeSignatureTickList(
+            [TimeSignature(0, *TIME_SIGNATURE)]
+        )

miditok/utils/utils.py

@@ -780,7 +780,7 @@ def get_num_notes_per_bar(
     :param tracks_indep: whether to process each track independently or all together.
     :return: the number of notes within each bar.
     """
-    if len(score.tracks) == 0:
+    if score.end() == 0:
         return [] if tracks_indep else [0]
 
     # Get bar and note times
@@ -789,13 +789,16 @@ def get_num_notes_per_bar(
         bar_ticks.append(score.end())
     tracks_times = [track.notes.numpy()["time"] for track in score.tracks]
     if not tracks_indep:
-        tracks_times = [np.concatenate(tracks_times)]
-        tracks_times[-1].sort()
-        num_notes_per_bar = []
+        if len(tracks_times) > 0:
+            tracks_times = [np.concatenate(tracks_times)]
+            tracks_times[-1].sort()
+        num_notes_per_bar = np.zeros(len(bar_ticks) - 1, dtype=np.int32)
     else:
-        num_notes_per_bar = [[] for _ in range(len(bar_ticks) - 1)]
+        num_notes_per_bar = np.zeros(
+            (len(bar_ticks) - 1, max(len(score.tracks), 1)), dtype=np.int32
+        )
 
-    for notes_times in tracks_times:
+    for ti, notes_times in enumerate(tracks_times):
         current_note_time_idx = previous_note_time_idx = 0
         current_bar_tick_idx = 0
         while current_bar_tick_idx < len(bar_ticks) - 1:
@@ -807,14 +810,14 @@ def get_num_notes_per_bar(
                 current_note_time_idx += 1
             num_notes = current_note_time_idx - previous_note_time_idx
             if tracks_indep and len(score.tracks) > 1:
-                num_notes_per_bar[current_bar_tick_idx].append(num_notes)
+                num_notes_per_bar[current_bar_tick_idx, ti] = num_notes
             else:
-                num_notes_per_bar.append(num_notes)
+                num_notes_per_bar[current_bar_tick_idx] = num_notes
 
             current_bar_tick_idx += 1
             previous_note_time_idx = current_note_time_idx
 
-    return num_notes_per_bar
+    return num_notes_per_bar.tolist()
 
 
 def get_score_ticks_per_beat(score: Score) -> np.ndarray: