def __init__(self, checkpoint, is_training=False):
"""Initialize the MajMinPopMusicTransformer model, setting the default
input dimension, memory length, number of encoder layers, and other
hyperparameters.
Args:
checkpoint: the filepath for the pre-trained model.
is_training: boolean indicating if model should be used in
train or evaluation mode.
"""
PopMusicTransformer.__init__(self, checkpoint, is_training)
# tick step-size; used to get tempo information in input.
self.DEFAULT_RESOLUTION = 480
def main(num_samples, num_bars, temperature, prompt_dir, output_dir):
# declare model
model = PopMusicTransformer(checkpoint='REMI-tempo-chord-checkpoint',
is_training=False)
from_scratch_path = output_dir + 'from_scratch/'
with_prompt_path = output_dir + 'with_prompt/'
if not os.path.exists(from_scratch_path):
os.makedirs(from_scratch_path)
if not os.path.exists(with_prompt_path):
os.makedirs(with_prompt_path)
for i in range(num_samples):
# generate from scratch
model.generate(n_target_bar=num_bars,
temperature=temperature,
output_path=from_scratch_path + '{}.midi'.format(i),
prompt=None)
# generate continuation
model.generate(n_target_bar=num_bars,
temperature=temperature,
output_path=with_prompt_path + '{}.midi'.format(i),
prompt=prompt_dir + f'{i:03}' + '.midi')
# close model
model.close()
def main():
# declare model
model = PopMusicTransformer(
checkpoint='my-piano',
is_training=False)
for i in range(0,10):
# generate continuation
model.generate(
n_target_bar=30,
temperature=1.2,
topk=5,
output_path='./result/continuation'+str(i)+'.midi',
prompt='./data/evaluation/00'+str(i)+'.mid')
# generate from scratch
model.generate(
n_target_bar=30,
temperature=1.2,
topk=5,
output_path='./result/from_scratch'+str(i)+'.midi',
prompt=None)
# close model
model.close()
def main():
# declare model
model = PopMusicTransformer(checkpoint='Chinese-piano-checkpoint',
is_training=True)
# prepare data
midi_paths = glob('data/train/*.mid') # you need to revise it
training_data = model.prepare_data(midi_paths=midi_paths)
# check output checkpoint folder
output_checkpoint_folder = 'my-piano' # your decision
if not os.path.exists(output_checkpoint_folder):
os.mkdir(output_checkpoint_folder)
# finetune
model.finetune(training_data=training_data,
output_checkpoint_folder=output_checkpoint_folder)
####################################
# after finetuning, please choose which checkpoint you want to try
# and change the checkpoint names you choose into "model"
# and copy the "dictionary.pkl" into the your output_checkpoint_folder
# ***** the same as the content format in "REMI-tempo-checkpoint" *****
# and then, you can use "main.py" to generate your own music!
# (do not forget to revise the checkpoint path to your own in "main.py")
####################################
# close
model.close()
def main():
# declare model
model = PopMusicTransformer(checkpoint='REMI-tempo-checkpoint',
is_training=True)
# prepare data
midi_paths = glob('finetune_midis/*.midi') # you need to revise it
training_data = model.prepare_data(midi_paths=midi_paths)
# check output checkpoint folder
####################################
# if you use "REMI-tempo-chord-checkpoint" for the pre-trained checkpoint
# please name your output folder as something with "chord"
# for example: my-love-chord, cute-doggy-chord, ...
# if use "REMI-tempo-checkpoint"
# for example: my-love, cute-doggy, ...
####################################
output_checkpoint_folder = 'REMI-finetune' # your decision
if not os.path.exists(output_checkpoint_folder):
os.mkdir(output_checkpoint_folder)
# finetune
model.finetune(training_data=training_data,
output_checkpoint_folder=output_checkpoint_folder)
####################################
# after finetuning, please choose which checkpoint you want to try
# and change the checkpoint names you choose into "model"
# and copy the "dictionary.pkl" into the your output_checkpoint_folder
# ***** the same as the content format in "REMI-tempo-checkpoint" *****
# and then, you can use "main.py" to generate your own music!
# (do not forget to revise the checkpoint path to your own in "main.py")
####################################
# close
model.close()
def main():
# declare model
model = PopMusicTransformer(checkpoint='REMI-tempo-checkpoint',
is_training=False)
# generate from scratch
model.generate(n_target_bar=16,
temperature=1.2,
output_path='./result/from_scratch.midi',
prompt=None)
# generate continuation
model.generate(n_target_bar=16,
temperature=1.2,
output_path='./result/continuation.midi',
prompt='./data/evaluation/000.midi')
# close model
model.close()
def main():
# declare model
model = PopMusicTransformer(is_training=False)
# generate from scratch
for i in range(100):
model.generate(n_target_bar=50,
temperature=1.2,
output_path='./result/from_scratch' + str(i) + '.midi',
prompt="evaluation/" + str(i).zfill(3) + ".midi")
model.close()
def main():
# declare model
model = PopMusicTransformer(checkpoint='REMI-tempo-checkpoint',
is_training=True,
use_chords=True,
train_from_scratch=False)
# prepare data
midi_paths = glob('midi/**/*.midi', recursive=True) + glob(
'midi/**/*.mid', recursive=True) # you need to revise it
if os.path.exists(
"233.data"
): # Revise the training data save location, as data preprocessing could take hours or even days on commercial hardware.
with open("233.data", "rb") as file:
training_data = pickle.load(file)
else:
training_data = model.prepare_data(midi_paths=midi_paths)
with open("233.data", "wb") as file:
pickle.dump(training_data, file,
protocol=4) # For large preprocessed files
# check output checkpoint folder
####################################
# Restrictions on the folder name is removed.
####################################
output_checkpoint_folder = 'REMI-chord-finetune' # your decision
if not os.path.exists(output_checkpoint_folder):
os.mkdir(output_checkpoint_folder)
# finetune
model.finetune(training_data=training_data,
output_checkpoint_folder=output_checkpoint_folder)
####################################
# after finetuning, please choose which checkpoint you want to try
# and change the checkpoint names you choose into "model"
# and copy the "dictionary.pkl" into the your output_checkpoint_folder
# ***** the same as the content format in "REMI-tempo-checkpoint" *****
# and then, you can use "main.py" to generate your own music!
# (do not forget to revise the checkpoint path to your own in "main.py")
####################################
# close
model.close()
def main():
# declare model
model = PopMusicTransformer(
checkpoint='kukkik-finetuned-checkpoint',
is_training=False)
# generate from scratch
model.generate(
n_target_bar=16,
temperature=1.2,
topk=5,
output_path='./result/from_scratch_finetuned.midi',
prompt=None)
# generate continuation
model.generate(
n_target_bar=16,
temperature=1.2,
topk=5,
output_path='./result/continuation_finetuned.midi',
prompt='./data/evaluation/012.midi')
# close model
model.close()
def load_model():
return PopMusicTransformer(checkpoint='REMI-tempo-checkpoint',
is_training=False)