def init_test():
timer = Timer()
seed = 43 # random, non-fine tuned seed
tf.random.set_seed(seed)
np.random.seed(seed)
random.seed(seed)
config = Config()
base_folder = config.base_data_folder
song_folders = create_song_list(config.dataset.beat_maps_folder)
total = len(song_folders)
print(f'Found {total} folders')
config.dataset.storage_folder = base_folder / 'new_datasets'
# config.dataset.storage_folder = base_folder / 'test_datasets' # only 100 songs, for testing
# config.audio_processing.use_cache = False
# generate_datasets(song_folders, config, prefix)
train, val, test = load_datasets(config)
timer('Loaded datasets', 0)
# Ensure this song is excluded from the training data for hand tasting
train.drop(index='133b', inplace=True, errors='ignore')
train.drop(index='Daddy - PSY', inplace=True, errors='ignore')
# dataset_stats(train)
manager = multiprocessing.Manager()
return_list = manager.list()
return base_folder, return_list, test, timer, train, val
def velocities_from_config(config: Config):
song_folders = create_song_list(config.dataset.beat_maps_folder)
df = songs2dataset(song_folders, config)
print(df.head(2))
df_vec = get_vec_df(df)
generated_velocities_ = compute_multiple_velocities(df_vec)
return generated_velocities_
def main():
timer = Timer()
seed = 43 # random, non-fine tuned seed
np.random.seed(seed)
random.seed(seed)
config = Config()
base_folder = config.base_data_folder
# To generate full dataset
song_folders = create_song_list(config.dataset.beat_maps_folder)
# config.dataset.storage_folder = base_folder / 'old_datasets'
config.dataset.storage_folder = base_folder / 'new_datasets'
# To generate test dataset
# song_folders = create_song_list(config.dataset.beat_maps_folder)[:100]
# config.dataset.storage_folder = base_folder / 'test_datasets'
config.audio_processing.use_cache = False # The audio features need to be computed the first time
config.use_multiprocessing = True # since TF is not imported
total = len(song_folders)
print(f'Found {total} folders')
generate_datasets(song_folders, config)
# Test loading datasets
train, val, test = load_datasets(config)
timer('Loaded datasets', 5)
def main():
timer = Timer()
seed = 43 # random, non-fine tuned seed
tf.random.set_seed(seed)
np.random.seed(seed)
random.seed(seed)
config = Config()
base_folder = config.base_data_folder
# Use full dataset
song_folders = create_song_list(config.dataset.beat_maps_folder)
# config.dataset.storage_folder = base_folder / 'old_datasets'
config.dataset.storage_folder = base_folder / 'new_datasets'
# Use test set
# song_folders = create_song_list(config.dataset.beat_maps_folder)[:100]
# config.dataset.storage_folder = base_folder / 'test_datasets'
config.audio_processing.use_cache = True
total = len(song_folders)
print(f'Found {total} folders')
generate_datasets(song_folders, config)
train, val, test = load_datasets(config)
timer('Loaded datasets', 5)
# Ensure this song is excluded from the training data for hand tasting
train.drop(index='133b', inplace=True, errors='ignore')
train.drop(index='Daddy - PSY', inplace=True, errors='ignore')
dataset_stats(train)
train_seq = BeatmapSequence(df=train, is_train=True, config=config)
val_seq = BeatmapSequence(df=val, is_train=False, config=config)
test_seq = BeatmapSequence(df=test, is_train=False, config=config)
timer('Generated sequences', 5)
# del train, val, test # delete the data if experiencing RAM problems
gc.collect()
# keras.mixed_precision.experimental.set_policy('mixed_float16') # Undefined behavior with advanced models
model_path = base_folder / 'temp'
model_path.mkdir(parents=True, exist_ok=True)
train = True
if train:
model = get_architecture_fn(config)(train_seq, False, config)
model.summary()
callbacks = create_callbacks(train_seq, config)
model.fit(
train_seq,
validation_data=val_seq,
callbacks=callbacks,
epochs=400,
verbose=2,
workers=10,
max_queue_size=16,
use_multiprocessing=False,
)
timer('Trained model', 5)
model.evaluate(test_seq)
timer('Evaluated model', 5)
save_model(model, model_path, train_seq, config)
timer('Saved model', 5)
stateful_model = keras.models.load_model(model_path /
'stateful_model.keras',
custom_objects={
'Perplexity': Perplexity,
'mish': tfa.activations.mish
})
stateful_model.summary()
timer('Loaded stateful model', 5)
# Use generated action placement
input_folder = base_folder / 'evaluation_dataset' / 'beat_sage'
input_folder = base_folder / 'evaluation_dataset' / 'oxai_deepsaber_expert'
output_folder = base_folder / 'testing' / 'generated_songs'
dirs = [x for x in input_folder.glob('*/') if x.is_dir()]
# Use human action placement from test set
# input_folder = base_folder / 'human_beatmaps' / 'new_dataformat'
# dirs = list(x for x in test.index.to_frame()["name"].unique()[:13])
for song_code in dirs:
beatmap_folder = input_folder / song_code
print(f'Working on {beatmap_folder.name}')
generate_complete_beatmaps(beatmap_folder, output_folder,
stateful_model, config)
timer('Generated beatmaps', 5)
def mainly_vec():
base_folder, return_list, test, timer, train, val = init_test()
prefix = 'vec'
test_name = 'information_comparison'
print('Running information comparison')
csv_file = base_folder / 'temp' / f'{prefix}{test_name}.csv'
dataset = DatasetConfig()
ALL = [
dataset.categorical + dataset.audio + dataset.regression,
]
# Common configuration
config = Config()
config.training.model_type = ModelType.CUSTOM
config.training.cnn_repetition = 0
config.training.lstm_repetition = 2
config.training.dense_repetition = 0
config.training.model_size = 512
config.training.dropout = 0.4
config.training.initial_learning_rate = 1e-2
config.training.batch_size = 128
config.training.label_smoothing = 0.5
config.training.mixup_alpha = 0.5
config.training.l2_regularization = 0
hp = None
config.dataset.beat_maps_folder = config.dataset.beat_maps_folder.parent / 'test_new_dataformat'
config.dataset.storage_folder = base_folder / 'test_new_datasets'
song_folders = create_song_list(config.dataset.beat_maps_folder)
# First generate all data using all of the audio features
config.audio_processing.use_temp_derrivatives = True
config.audio_processing.time_shift = -0.4
dataset_stats(train)
for repetition in range(7):
config.training.x_groups = ALL
config.training.y_groups = [
['word_vec'],
]
configuration_name = f'X: {config.training.x_groups}\nY: {config.training.y_groups}'
eval_config(csv_file, timer, return_list, train, val, test, config,
test_name, configuration_name, hp)
config.training.x_groups = [
['prev_word_vec'],
]
config.training.y_groups = [
['word_id'],
]
configuration_name = f'X: {config.training.x_groups}\nY: {config.training.y_groups}'
eval_config(csv_file, timer, return_list, train, val, test, config,
test_name, configuration_name, hp)
config.training.x_groups = [
['prev_word_vec'],
['mfcc'],
]
config.training.y_groups = [
['word_id'],
]
configuration_name = f"X: {[['prev_word_id'], ['MFCC', 'dMFCC']]}\nY: {config.training.y_groups}"
eval_config(csv_file, timer, return_list, train, val, test, config,
test_name, configuration_name, hp)
config.training.x_groups = [
['prev_word_vec'],
['prev', 'next'],
]
config.training.y_groups = [
['word_id'],
]
configuration_name = f"X: {config.training.x_groups}\nY: {config.training.y_groups}"
eval_config(csv_file, timer, return_list, train, val, test, config,
test_name, configuration_name, hp)
config.training.x_groups = [
['prev_word_vec'],
['part'],
]
config.training.y_groups = [
['word_id'],
]
configuration_name = f"X: {config.training.x_groups}\nY: {config.training.y_groups}"
eval_config(csv_file, timer, return_list, train, val, test, config,
test_name, configuration_name, hp)
config.training.x_groups = [
['prev_word_vec'],
] + ALL
config.training.y_groups = [
['word_id'],
]
configuration_name = f'X: {config.training.x_groups}\nY: {config.training.y_groups}'
eval_config(csv_file, timer, return_list, train, val, test, config,
test_name, configuration_name, hp)
config.training.x_groups = [
['prev_word_vec', 'prev_word_id'],
] + ALL
config.training.y_groups = [
['word_id'],
]
configuration_name = f'X: {config.training.x_groups}\nY: {config.training.y_groups}'
eval_config(csv_file, timer, return_list, train, val, test, config,
test_name, configuration_name, hp)
config.training.x_groups = [
['prev_word_vec', 'prev_word_id'],
] + ALL
config.training.y_groups = [
['word_vec', 'word_id'],
]
configuration_name = f'X: {config.training.x_groups}\nY: {config.training.y_groups}'
eval_config(csv_file, timer, return_list, train, val, test, config,
test_name, configuration_name, hp)
pass