def generate(hparams):
input_vocab_size = 128 + 128 + 128 + 128
target_vocab_size = 128 + 128 + 128 + 128
dataset = tf.data.Dataset.list_files(
'/home/big/datasets/maestro-v2.0.0/**/*.midi')
dataset_single = pro.pipeline([
pro.midi(),
pro.frame(hparams['frame_size'], hparams['frame_size'], True),
pro.unbatch(),
])(dataset).skip(16000).as_numpy_iterator()
transformer = Transformer(input_vocab_size=input_vocab_size,
target_vocab_size=target_vocab_size,
pe_input=input_vocab_size,
pe_target=target_vocab_size,
hparams=hparams)
trainer = Trainer(dataset, hparams)
ckpt = tf.train.Checkpoint(step=trainer.step,
transformer=transformer,
optimizer=transformer.optimizer)
trainer.init_checkpoint(ckpt)
return generate_from_model(hparams, transformer, dataset_single)
def maestro_from_files(root_path, frame_size):
dataset = tf.data.Dataset.list_files(os.path.join(root_path, '**/*.wav'))
dataset = pro.pipeline([
pro.read_file(),
pro.decode_wav(desired_channels=1),
pro.map_transform(lambda x: x[0]),
pro.reshape([-1]),
pro.frame(frame_size, frame_size),
pro.unbatch()
])(dataset)
return dataset
def start(hparams):
dataset = tf.data.Dataset.list_files('dataset/*.midi')
dataset = pro.pipeline([
pro.midi(),
pro.prefetch(),
pro.frame(hparams['frame_size'], hparams['frame_size'], True),
pro.unbatch(),
# pro.map_transform(tf_serialize_example)
])(dataset)
def generator():
for features in dataset:
yield serialize_example(features)
serialized_features_dataset = tf.data.Dataset.from_generator(
generator, output_types=tf.string, output_shapes=())
filename = 'midi_dataset.tfrecord'
writer = tf.data.experimental.TFRecordWriter(filename)
writer.write(serialized_features_dataset)
def start(hparams):
# Load nsynth dataset
# dataset = tfds.load('nsynth/gansynth_subset', split='train', shuffle_files=True)
dataset = tf.data.Dataset.list_files(
'/home/big/datasets/maestro-v2.0.0/**/*.wav')
dataset = pro.pipeline([
pro.wav(),
# pro.resample(16000, hparams['sample_rate'], tf.float32),
pro.normalize(),
pro.frame(hparams['window_samples'], hparams['window_samples']),
pro.unbatch(),
pro.set_channels(1),
pro.dupe(),
pro.shuffle(hparams['buffer_size']),
pro.batch(hparams['batch_size']),
pro.prefetch()
])(dataset)
vae = VAE(hparams)
vae.vae.summary()
trainer = Trainer(dataset, hparams)
ckpt = tf.train.Checkpoint(
step=trainer.step,
encoder=vae.encoder,
decoder=vae.decoder,
vae=vae.vae,
)
trainer.on_epoch_start = on_epoch_start
trainer.on_step = on_step
trainer.init_tensorboard()
trainer.init_checkpoint(ckpt)
trainer.set_train_step(vae.train_step)
trainer.run()
def start(hparams):
gc.collect()
dataset = tf.data.Dataset.list_files(
'/home/big/datasets/maestro-v2.0.0/**/*.midi')
dataset_single = pro.pipeline([
pro.midi(),
pro.frame(hparams['frame_size'] * 2, hparams['frame_hop_len'], True),
pro.unbatch(),
])(dataset)
def _reshape(inp, tar):
inp = tf.reshape(inp, [hparams['frame_size']])
tar = tf.reshape(tar, [hparams['frame_size']])
return inp, tar
dataset = pro.pipeline([
pro.split(2),
#pro.batch(2, True),
# pro.dupe(),
pro.map_transform(_reshape),
pro.cache(),
pro.shuffle(hparams['buffer_size']),
pro.batch(hparams['batch_size'], True),
pro.prefetch(),
])(dataset_single)
dataset_single = pro.shuffle(hparams['buffer_size'] // 4)(dataset_single)
dataset_single = dataset_single.as_numpy_iterator()
transformer = Transformer(input_vocab_size=input_vocab_size,
target_vocab_size=target_vocab_size,
pe_input=input_vocab_size,
pe_target=target_vocab_size,
hparams=hparams)
# pop_size = 10
# generations = 100
# mutation_rate = 0.2
# pool.populate(pop_size, 1)
# for generation in range(generations):
# pool.evaluate(evaluate(dataset, hparams))
# print(f"--- GENERATION: {generation} ---")
# print("BEST:", pool.best, pool.fitness)
# pool.select(pop_size)
# pool.populate(pop_size, mutation_rate)
#
image_save_step = hparams[
'image_save_step'] if 'image_save_step' in hparams else 2000
def generate_image(step, tsw):
print("Generating sample...")
encoded, seed = generate_from_model(hparams, transformer,
dataset_single)
print("Generating sample done.")
print("Decoding midi...")
decoded_seed = pro.decode_midi()(seed)
decoded = pro.decode_midi()(encoded)
print("Decoding midi done.")
print("Saving midi...")
with open(f'gen_transformer_{step}.midi', 'wb') as f:
M.write_midi(f, decoded)
with open(f'prior_transformer_{step}.midi', 'wb') as f:
M.write_midi(f, decoded_seed)
print("Saving midi done.")
print("Plotting midi...")
plt.title('Prior')
M.display_midi(decoded_seed)
image_seed = util.get_plot_image()
plt.clf()
plt.title('Generated')
M.display_midi(decoded)
image = util.get_plot_image()
plt.clf()
image_conc = tf.concat([image_seed, image], axis=1)
print("Plotting done.")
with tsw.as_default():
tf.summary.image(f'image', image_conc, step=step)
print("Complete.")
# This runs at every step in the training (for each batch in dataset)
def on_step(epoch, step, stats, tsw):
loss, tar_real, predictions = stats
train_loss(loss)
train_accuracy(tar_real, predictions)
if step % 100 == 0:
print(
f"Epoch: {epoch}, Step: {step}, Loss: {train_loss.result()}, Accuracy: {train_accuracy.result()}"
)
if step % image_save_step == 0:
generate_image(step, tsw)
with tsw.as_default():
tf.summary.scalar('loss', train_loss.result(), step=step)
trainer = Trainer(dataset, hparams)
ckpt = tf.train.Checkpoint(step=trainer.step,
transformer=transformer,
optimizer=transformer.optimizer)
trainer.init_checkpoint(ckpt)
trainer.init_tensorboard()
trainer.set_train_step(transformer.train_step)
trainer.on_epoch_start = on_epoch_start
trainer.on_step = on_step
trainer.on_epoch_complete = on_epoch_complete
#generate_image(trainer.step.numpy(), trainer.train_summary_writer)
trainer.run()