def make_rand_songs(write_dir, rand_vecs):
for i in range(rand_vecs.shape[0]):
x_rand = rand_vecs[i:i + 1]
y_song = func([x_rand, 0])[0]
midi.samples_to_midi(
[y_song[0, meas] for meas in range(y_song[0].ndim)],
write_dir + 'rand' + str(i) + '.mid',
thresh=0.25)
def make_rand_songs(write_dir, rand_vecs, thresh):
for i in range(rand_vecs.shape[0]):
x_rand = torch.from_numpy(rand_vecs[i:i + 1]).to(device).float()
with torch.no_grad():
y_song = model.decoder_forward(x_rand).cpu().numpy()
midi.samples_to_midi(y_song[0],
write_dir + 'rand' + str(i) + '.mid', 16,
thresh)
fs.midi_to_audio(write_dir + 'rand' + str(i) + '.mid',
output_dir + 'sample' + str(i) + '.wav')
else:
rand_ix = np.array([rand_ix], dtype=np.int64)
x = enc_model.predict(rand_ix, batch_size=1)[0]
rand_ix = (rand_ix + 1) % model.layers[0].input_dim
if use_pca:
cur_params = np.dot(x - means, evecs.T) / evals
else:
cur_params = (x - means) / stds
needs_update = True
audio_reset = True
if event.key == pygame.K_g:
audio_pause = True
audio_reset = True
midi.samples_to_midi(cur_notes, 'live.mid', 16, note_thresh)
save_audio = ''
while True:
save_audio += audio_callback(None, 1024, None, None)[0]
if audio_time == 0:
break
wave_output = wave.open('live.wav', 'w')
wave_output.setparams((1, 2, sample_rate, 0, 'NONE', 'not compressed'))
wave_output.writeframes(save_audio)
wave_output.close()
audio_pause = False
if event.key == pygame.K_ESCAPE:
running = False
break
if event.key == pygame.K_SPACE:
audio_pause = not audio_pause
def make_rand_songs(write_dir, rand_vecs):
for i in range(rand_vecs.shape[0]):
x_rand = rand_vecs[i:i + 1]
y_song = func([x_rand, 0])[0]
midi.samples_to_midi(y_song[0], write_dir + 'rand' + str(i) + '.mid',
16, 0.25)
epsilon = K.random_normal(shape=K.shape(z_mean), mean=0.0, stddev=VAE_B1)
return z_mean + K.exp(z_log_sigma_sq * 0.5) * epsilon
def vae_loss(x, x_decoded_mean):
xent_loss = binary_crossentropy(x, x_decoded_mean)
kl_loss = VAE_B2 * K.mean(
1 + z_log_sigma_sq - K.square(z_mean) - K.exp(z_log_sigma_sq),
axis=None)
return xent_loss - kl_loss
test_ix = 0
y_test_song = np.copy(y_train[test_ix:test_ix + 1])
x_test_song = np.copy(x_train[test_ix:test_ix + 1])
midi.samples_to_midi(y_test_song[0], 'gt.mid', 16)
###################################
# Create Model
###################################
if CONTINUE_TRAIN or PLAY_ONLY:
print("Loading Model...")
model = load_model('model.h5', custom_objects=custom_objects)
else:
print("Building Model...")
if USE_EMBEDDING:
x_in = Input(shape=x_shape[1:])
print((None, ) + x_shape[1:])
x = Embedding(x_train.shape[0], PARAM_SIZE, input_length=1)(x_in)
x = Flatten(name='pre_encoder')(x)
def make_rand_songs(write_dir, rand_vecs): for i in xrange(rand_vecs.shape[0]): x_rand = rand_vecs[i:i+1] y_song = func([x_rand, 0])[0] midi.samples_to_midi(y_song[0], write_dir + 'rand' + str(i) + '.mid', ticks_per_sample = 0.125, thresh=0.20)
epsilon = K.random_normal(shape=K.shape(z_mean), mean=0.0, stddev=VAE_B1)
return z_mean + K.exp(z_log_sigma_sq * 0.5) * epsilon
def vae_loss(x, x_decoded_mean):
xent_loss = binary_crossentropy(x, x_decoded_mean)
kl_loss = VAE_B2 * K.mean(
1 + z_log_sigma_sq - K.square(z_mean) - K.exp(z_log_sigma_sq),
axis=None)
return xent_loss - kl_loss
test_ix = 120
y_test_song = np.copy(y_train[test_ix:test_ix + 1])
x_test_song = np.copy(x_train[test_ix:test_ix + 1])
midi.samples_to_midi(y_test_song[0], 'gt.mid', 16)
###################################
# Create Model
###################################
if CONTINUE_TRAIN or PLAY_ONLY:
print("Loading Model...")
model = load_model('model.h5', custom_objects=custom_objects)
else:
print("Building Model...")
if USE_EMBEDDING: # Embedding is used when
x_in = Input(shape=x_shape[1:])
print((None, ) + x_shape[1:])
x = Embedding(x_train.shape[0], PARAM_SIZE, input_length=1)(x_in)
x = Flatten(name='pre_encoder')(x)
