lingyu123-su/MIDIFoundationModel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

1013 update

Browse Source
This commit is contained in:
FelixChan
2025-10-13 17:56:36 +08:00
parent d077e3210e
commit d6b68ef90b
17 changed files with 815 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files

167
Amadeus/symbolic_encoding/data_utils.py
View File

@ -530,6 +530,62 @@ class Melody(SymbolicMusicDataset):
test_names.extend(song_dict[song_name])
return train_names, valid_names, test_names, shuffled_tune_names
class msmidi(SymbolicMusicDataset):
def __init__(self, vocab, encoding_scheme, num_features, debug, aug_type, input_length, first_pred_feature, caption_path=None,
for_evaluation: bool = False):
super().__init__(vocab, encoding_scheme, num_features, debug, aug_type, input_length, first_pred_feature, caption_path,
for_evaluation=for_evaluation)
def _load_tune_in_idx(self) -> Tuple[Dict[str, np.ndarray], Dict[str, int], List[str]]:
'''
Irregular tunes are removed from the dataset for better generation quality
It includes tunes that are not quantized properly, mostly theay are expressive performance data
'''
print("preprocessed tune_in_idx data is being loaded")
tune_in_idx_list = sorted(list(Path(f"dataset/represented_data/tuneidx/tuneidx_{self.__class__.__name__}/{self.encoding_scheme}{self.num_features}").rglob("*.npz")))
if self.debug:
tune_in_idx_list = tune_in_idx_list[:5000]
tune_in_idx_dict = OrderedDict()
len_tunes = OrderedDict()
file_name_list = []
with open("metadata/LakhClean_irregular_tunes.json", "r") as f:
irregular_tunes = json.load(f)
for tune_in_idx_file in tqdm(tune_in_idx_list, total=len(tune_in_idx_list)):
if tune_in_idx_file.stem in irregular_tunes:
continue
tune_in_idx = np.load(tune_in_idx_file)['arr_0']
tune_in_idx_dict[tune_in_idx_file.stem] = tune_in_idx
len_tunes[tune_in_idx_file.stem] = len(tune_in_idx)
file_name_list.append(tune_in_idx_file.stem)
print(f"number of loaded tunes: {len(tune_in_idx_dict)}")
return tune_in_idx_dict, len_tunes, file_name_list
def _get_split_list_from_tune_in_idx(self, ratio, seed):
'''
As Lakh dataset contains multiple versions of the same song, we split the dataset based on the song name
'''
shuffled_tune_names = list(self.tune_in_idx.keys())
song_names_without_version = [re.sub(r"\.\d+$", "", song) for song in shuffled_tune_names]
song_dict = {}
for song, orig_song in zip(song_names_without_version, shuffled_tune_names):
if song not in song_dict:
song_dict[song] = []
song_dict[song].append(orig_song)
unique_song_names = list(song_dict.keys())
random.seed(seed)
random.shuffle(unique_song_names)
num_train = int(len(unique_song_names)*ratio)
num_valid = int(len(unique_song_names)*(1-ratio)/2)
train_names = []
valid_names = []
test_names = []
for song_name in unique_song_names[:num_train]:
train_names.extend(song_dict[song_name])
for song_name in unique_song_names[num_train:num_train+num_valid]:
valid_names.extend(song_dict[song_name])
for song_name in unique_song_names[num_train+num_valid:]:
test_names.extend(song_dict[song_name])
return train_names, valid_names, test_names, shuffled_tune_names
class IrishMan(SymbolicMusicDataset):
def __init__(self, vocab, encoding_scheme, num_features, debug, aug_type, input_length, first_pred_feature, caption_path=None,
@ -648,62 +704,62 @@ class ariamidi(SymbolicMusicDataset):
test_names.extend(song_dict[song_name])
return train_names, valid_names, test_names, shuffled_tune_names
class gigamidi(SymbolicMusicDataset):
def __init__(self, vocab, encoding_scheme, num_features, debug, aug_type, input_length, first_pred_feature, caption_path=None,
for_evaluation: bool = False):
super().__init__(vocab, encoding_scheme, num_features, debug, aug_type, input_length, first_pred_feature, caption_path,
for_evaluation=for_evaluation)
# class gigamidi(SymbolicMusicDataset):
# def __init__(self, vocab, encoding_scheme, num_features, debug, aug_type, input_length, first_pred_feature, caption_path=None,
# for_evaluation: bool = False):
# super().__init__(vocab, encoding_scheme, num_features, debug, aug_type, input_length, first_pred_feature, caption_path,
# for_evaluation=for_evaluation)
def _load_tune_in_idx(self) -> Tuple[Dict[str, np.ndarray], Dict[str, int], List[str]]:
'''
Irregular tunes are removed from the dataset for better generation quality
It includes tunes that are not quantized properly, mostly theay are expressive performance data
'''
print("preprocessed tune_in_idx data is being loaded")
tune_in_idx_list = sorted(list(Path(f"dataset/represented_data/tuneidx/tuneidx_{self.__class__.__name__}/{self.encoding_scheme}{self.num_features}").rglob("*.npz")))
if self.debug:
tune_in_idx_list = tune_in_idx_list[:5000]
tune_in_idx_dict = OrderedDict()
len_tunes = OrderedDict()
file_name_list = []
with open("metadata/LakhClean_irregular_tunes.json", "r") as f:
irregular_tunes = json.load(f)
for tune_in_idx_file in tqdm(tune_in_idx_list, total=len(tune_in_idx_list)):
if tune_in_idx_file.stem in irregular_tunes:
continue
tune_in_idx = np.load(tune_in_idx_file)['arr_0']
tune_in_idx_dict[tune_in_idx_file.stem] = tune_in_idx
len_tunes[tune_in_idx_file.stem] = len(tune_in_idx)
file_name_list.append(tune_in_idx_file.stem)
print(f"number of loaded tunes: {len(tune_in_idx_dict)}")
return tune_in_idx_dict, len_tunes, file_name_list
# def _load_tune_in_idx(self) -> Tuple[Dict[str, np.ndarray], Dict[str, int], List[str]]:
# '''
# Irregular tunes are removed from the dataset for better generation quality
# It includes tunes that are not quantized properly, mostly theay are expressive performance data
# '''
# print("preprocessed tune_in_idx data is being loaded")
# tune_in_idx_list = sorted(list(Path(f"dataset/represented_data/tuneidx/tuneidx_{self.__class__.__name__}/{self.encoding_scheme}{self.num_features}").rglob("*.npz")))
# if self.debug:
# tune_in_idx_list = tune_in_idx_list[:5000]
# tune_in_idx_dict = OrderedDict()
# len_tunes = OrderedDict()
# file_name_list = []
# with open("metadata/LakhClean_irregular_tunes.json", "r") as f:
# irregular_tunes = json.load(f)
# for tune_in_idx_file in tqdm(tune_in_idx_list, total=len(tune_in_idx_list)):
# if tune_in_idx_file.stem in irregular_tunes:
# continue
# tune_in_idx = np.load(tune_in_idx_file)['arr_0']
# tune_in_idx_dict[tune_in_idx_file.stem] = tune_in_idx
# len_tunes[tune_in_idx_file.stem] = len(tune_in_idx)
# file_name_list.append(tune_in_idx_file.stem)
# print(f"number of loaded tunes: {len(tune_in_idx_dict)}")
# return tune_in_idx_dict, len_tunes, file_name_list
def _get_split_list_from_tune_in_idx(self, ratio, seed):
'''
As Lakh dataset contains multiple versions of the same song, we split the dataset based on the song name
'''
shuffled_tune_names = list(self.tune_in_idx.keys())
song_names_without_version = [re.sub(r"\.\d+$", "", song) for song in shuffled_tune_names]
song_dict = {}
for song, orig_song in zip(song_names_without_version, shuffled_tune_names):
if song not in song_dict:
song_dict[song] = []
song_dict[song].append(orig_song)
unique_song_names = list(song_dict.keys())
random.seed(seed)
random.shuffle(unique_song_names)
num_train = int(len(unique_song_names)*ratio)
num_valid = int(len(unique_song_names)*(1-ratio)/2)
train_names = []
valid_names = []
test_names = []
for song_name in unique_song_names[:num_train]:
train_names.extend(song_dict[song_name])
for song_name in unique_song_names[num_train:num_train+num_valid]:
valid_names.extend(song_dict[song_name])
for song_name in unique_song_names[num_train+num_valid:]:
test_names.extend(song_dict[song_name])
return train_names, valid_names, test_names, shuffled_tune_names
# def _get_split_list_from_tune_in_idx(self, ratio, seed):
# '''
# As Lakh dataset contains multiple versions of the same song, we split the dataset based on the song name
# '''
# shuffled_tune_names = list(self.tune_in_idx.keys())
# song_names_without_version = [re.sub(r"\.\d+$", "", song) for song in shuffled_tune_names]
# song_dict = {}
# for song, orig_song in zip(song_names_without_version, shuffled_tune_names):
# if song not in song_dict:
# song_dict[song] = []
# song_dict[song].append(orig_song)
# unique_song_names = list(song_dict.keys())
# random.seed(seed)
# random.shuffle(unique_song_names)
# num_train = int(len(unique_song_names)*ratio)
# num_valid = int(len(unique_song_names)*(1-ratio)/2)
# train_names = []
# valid_names = []
# test_names = []
# for song_name in unique_song_names[:num_train]:
# train_names.extend(song_dict[song_name])
# for song_name in unique_song_names[num_train:num_train+num_valid]:
# valid_names.extend(song_dict[song_name])
# for song_name in unique_song_names[num_train+num_valid:]:
# test_names.extend(song_dict[song_name])
# return train_names, valid_names, test_names, shuffled_tune_names
class ariamidi(SymbolicMusicDataset):
def __init__(self, vocab, encoding_scheme, num_features, debug, aug_type, input_length, first_pred_feature, caption_path=None,
@ -788,6 +844,9 @@ class gigamidi(SymbolicMusicDataset):
for tune_in_idx_file in tqdm(tune_in_idx_list, total=len(tune_in_idx_list)):
if tune_in_idx_file.stem in irregular_tunes:
continue
if "drums-only" in tune_in_idx_file.stem:
print(f"skipping {tune_in_idx_file.stem} as it is a drums-only file")
continue
tune_in_idx = np.load(tune_in_idx_file)['arr_0']
tune_in_idx_dict[tune_in_idx_file.stem] = tune_in_idx
len_tunes[tune_in_idx_file.stem] = len(tune_in_idx)