def main():
print 'Generating the training model...'
training_model = get_training_model()
print 'Compiling the training model...'
optimizer = Adadelta(lr=0.01, epsilon=1e-6)
training_model.compile(loss=objective,
optimizer=optimizer,
metrics=['accuracy'])
print 'Retrieving the repository...'
pieces = repository_handler.load_repository(args.repository)
print 'Learning...'
train(training_model, pieces)
print 'Retrieving the weights...'
weights = training_model.get_weights()
print 'Generating the composition model...'
composition_model = get_composition_model()
print 'Compiling the composition model...'
composition_model.compile(loss=objective, optimizer=optimizer)
print 'Setting the weights...'
composition_model.set_weights(weights)
print 'Composing a piece...'
random_batch, _ = piece_handler.get_piece_batch(pieces, 5)
initial_note = random_batch[0][0].reshape((1, 78, 80))
piece = compose_piece(composition_model, initial_note)
piece_handler.save_piece(piece, args.piece)
def main():
print 'Retrieving the repository...'
pieces = repository_handler.load_repository(args.repository)
_, data = piece_handler.get_piece_batch(pieces, args.batch_size)
print 'Generating Probability Matrix..'
probability_matrix = get_probability_matrix(data)
print 'Making Sweet Jams...'
segment = generate_music(data[0, 0], probability_matrix,
args.segment_length)
print 'Saving Sweet Jams...'
piece_handler.save_piece(segment, args.file_path)
def train(deep_jammer, pieces, epochs, batch_size):
loss_history = []
for epoch in xrange(epochs):
loss = deep_jammer.update(*piece_handler.get_piece_batch(pieces, batch_size))
loss_history.append(loss)
if epoch % SUMMARY_THRESHOLD == 0:
display_summary(epoch, batch_size, loss)
if epoch % CHECKPOINT_THRESHOLD == 0:
print 'Epoch %d -> Checkpoint' % epoch
save_configuration(deep_jammer, epoch)
save_loss_history(loss_history)
piece = generate_piece(deep_jammer, pieces)
save_generated_piece(piece, epoch)
def main():
print 'Generating the composition model...'
composition_model = deep_jammer.get_composition_model()
print 'Compiling the composition model...'
optimizer = Adam(lr=1)
composition_model.compile(loss=objective, optimizer=optimizer, metrics=['accuracy'])
print 'Loading the weights...'
composition_model.load_weights('checkpoints/model-weights-40.h5')
print 'Generating the initial note of the piece...'
repository = repository_handler.load_repository('1-repository')
X_train, _ = piece_handler.get_piece_batch(repository, 5)
initial_note = X_train[0][0].reshape((1, 78, 80))
print 'Generating a piece...'
piece = compose_piece(composition_model, initial_note)
piece_handler.save_piece(piece, 'checkpoint-piece')
def train(model, pieces):
loss_history = []
for epoch in xrange(EPOCHS):
print 'Training on epoch %s/%s...' % (epoch, EPOCHS)
X, y = piece_handler.get_piece_batch(pieces, BATCH_SIZE)
loss, _ = model.train_on_batch([X, y], y)
print 'Loss =', loss
loss_history.append(loss)
if ARE_CHECKPOINTS_ENABLED and epoch % CHECKPOINT_THRESHOLD == 0:
filename = '%s/model-weights-%s.h5' % (CHECKPOINT_DIRECTORY, epoch)
model.save_weights(filename)
if ARE_ACCURACIES_SAVED and epoch % ACCURACIES_THRESHOLD == 0:
if IS_GPU_USED:
filename = '%s/model-accuracies.txt' % ACCURACIES_DIRECTORY
f = open(filename, 'w')
for new_loss in loss_history:
f.write('%s\n' % new_loss)
else:
import matplotlib.pyplot as plt
filename = '%s/model-accuracies.png' % ACCURACIES_DIRECTORY
figure = plt.figure()
plt.plot(loss_history)
figure.suptitle('Loss Analysis')
plt.xlabel('Iterations')
plt.ylabel('Loss')
figure.savefig(filename)