def main():
os.environ["CUDA_VISIBLE_DEVICES"] = ''
print("Num GPUs Available: ",
len(tf.config.experimental.list_physical_devices('GPU')))
raw_maestro = waveform_dataset.get_waveform_dataset(MAESTRO_PATH)
generator = ls_conditional_wave_spec_gan.Generator()
checkpoint_path = '_results/conditioning/LSC_WaveSpecGAN_HR_8192/training_checkpoints/ckpt-2'
checkpoint = tf.train.Checkpoint(generator=generator)
checkpoint.restore(checkpoint_path).expect_partial()
seed = raw_maestro[SEED_INDEX]
sequence = [np.reshape(seed, (GENERATION_LENGTH))]
for i in range(N_GENERATIONS):
z_in = tf.random.uniform((1, Z_DIM), -1, 1)
seed_in = sequence[i][CONDITIONING_START_INDEX:]
seed_in = np.expand_dims(seed_in, 0)
gen = generator(seed_in, z_in)
sequence.append(np.squeeze(gen))
audio = np.reshape(sequence, (-1))
sf.write('babble.wav', audio, SAMPLE_RATE)
def main():
os.environ["CUDA_VISIBLE_DEVICES"] = '2'
print("Num GPUs Available: ",
len(tf.config.experimental.list_physical_devices('GPU')))
raw_maestro = waveform_dataset.get_waveform_dataset(MAESTRO_PATH)
raw_maestro_conditioning = waveform_dataset.get_waveform_dataset(
MAESTRO_MIDI_PATH).astype(np.float32)
generator = conditional_wave_spec_gan.Generator()
discriminator = conditional_wave_spec_gan.WaveformDiscriminator(
input_shape=WAVEFORM_SHAPE, weighting=CRITIC_WEIGHTINGS[0])
spec_discriminator = conditional_wave_spec_gan.SpectogramDiscriminator(
input_shape=MAGNITUDE_IMAGE_SHAPE, weighting=CRITIC_WEIGHTINGS[1])
generator_optimizer = tf.keras.optimizers.Adam(1e-4,
beta_1=0.5,
beta_2=0.9)
discriminator_optimizer = tf.keras.optimizers.Adam(1e-4,
beta_1=0.5,
beta_2=0.9)
get_waveform = lambda waveform: waveform
save_examples = lambda epoch, real, generated:\
save_helper.save_wav_data(
epoch, real, generated, SAMPLING_RATE, RESULT_DIR, get_waveform
)
wave_gan_model = conditional_wgan.ConditionalWGAN(
raw_maestro,
raw_maestro_conditioning,
generator, [discriminator, spec_discriminator],
Z_DIM,
generator_optimizer,
discriminator_optimizer,
discriminator_training_ratio=D_UPDATES_PER_G,
batch_size=BATCH_SIZE,
epochs=EPOCHS,
checkpoint_dir=CHECKPOINT_DIR,
fn_save_examples=save_examples,
fn_get_discriminator_input_representations=
_get_discriminator_input_representations)
wave_gan_model.train()
def main():
raw_maestro = waveform_dataset.get_waveform_dataset(MAESTRO_PATH)
raw_maestro_conditioning = waveform_dataset.get_waveform_dataset(
MAESTRO_CONDITIONING_PATH)
generator = ls_conditional_wave_spec_gan.Generator()
discriminator = ls_conditional_wave_spec_gan.WaveformDiscriminator(
input_shape=WAVEFORM_SHAPE, weighting=CRITIC_WEIGHTINGS[0])
spec_discriminator = ls_conditional_wave_spec_gan.SpectogramDiscriminator(
input_shape=MAGNITUDE_IMAGE_SHAPE, weighting=CRITIC_WEIGHTINGS[1])
generator_optimizer = tf.keras.optimizers.Adam(1e-4,
beta_1=0.5,
beta_2=0.9)
discriminator_optimizer = tf.keras.optimizers.Adam(1e-4,
beta_1=0.5,
beta_2=0.9)
get_waveform = lambda waveform: waveform
save_examples = lambda epoch, real, generated:\
save_helper.save_wav_data(
epoch, real, generated, SAMPLE_RATE, RESULT_DIR, get_waveform
)
wave_gan_model = conditional_wgan.ConditionalWGAN(
raw_maestro,
raw_maestro_conditioning,
generator, [discriminator, spec_discriminator],
Z_DIM,
generator_optimizer,
discriminator_optimizer,
discriminator_training_ratio=D_UPDATES_PER_G,
batch_size=BATCH_SIZE,
epochs=EPOCHS,
checkpoint_dir=CHECKPOINT_DIR,
fn_save_examples=save_examples,
fn_get_discriminator_input_representations=
_get_discriminator_input_representations)
wave_gan_model.train()
def main():
os.environ["CUDA_VISIBLE_DEVICES"] = '1'
print("Num GPUs Available: ",
len(tf.config.experimental.list_physical_devices('GPU')))
raw_maestro = waveform_dataset.get_stft_dataset(
MAESTRO_PATH, frame_length=FFT_FRAME_LENGTH,
frame_step=FFT_FRAME_STEP).astype(np.float32)
raw_maestro_conditioning = waveform_dataset.get_waveform_dataset(
MAESTRO_MIDI_PATH).astype(np.float32)
generator = midi_conditional_spec_gan.Generator(channels=2)
discriminator = midi_conditional_spec_gan.Discriminator(
input_shape=STFT_IMAGE_SHAPE)
spec_discriminator = midi_conditional_spec_gan.Discriminator(
input_shape=MAGNITUDE_IMAGE_SHAPE)
generator_optimizer = tf.keras.optimizers.Adam(1e-4,
beta_1=0.5,
beta_2=0.9)
discriminator_optimizer = tf.keras.optimizers.Adam(1e-4,
beta_1=0.5,
beta_2=0.9)
get_waveform = lambda stft:\
spectral.stft_2_waveform(
stft, FFT_FRAME_LENGTH, FFT_FRAME_STEP
)[0]
save_examples = lambda epoch, real, generated:\
save_helper.save_wav_data(
epoch, real, generated, SAMPLING_RATE, RESULT_DIR, get_waveform
)
stft_gan_model = conditional_wgan.ConditionalWGAN(
raw_maestro,
raw_maestro_conditioning,
generator, [discriminator, spec_discriminator],
Z_DIM,
generator_optimizer,
discriminator_optimizer,
discriminator_training_ratio=D_UPDATES_PER_G,
batch_size=BATCH_SIZE,
epochs=EPOCHS,
checkpoint_dir=CHECKPOINT_DIR,
fn_save_examples=save_examples,
fn_get_discriminator_input_representations=
_get_discriminator_input_representations)
stft_gan_model.train()
def main():
os.environ["CUDA_VISIBLE_DEVICES"] = ''
print("Num GPUs Available: ",
len(tf.config.experimental.list_physical_devices('GPU')))
raw_maestro = waveform_dataset.get_waveform_dataset(MAESTRO_PATH)
for model in MODELS:
n_basis_functions = MODELS[model]['num_basis_functions']
basis_function_length = MODELS[model]['basis_function_legnth']
encoder = learned_basis_function.Encoder(basis_function_length,
n_basis_functions)
decoder = learned_basis_function.Decoder(basis_function_length)
checkpoint = tf.train.Checkpoint(encoder=encoder, decoder=decoder)
checkpoint.restore(MODELS[model]['checkpoint_path']).expect_partial()
spectogram_like = encoder(
raw_maestro[DECOMPOSITION_IDX:DECOMPOSITION_IDX + 1])
reconstructed = decoder(spectogram_like)
sf.write('{}_orig.wav'.format(model), raw_maestro[DECOMPOSITION_IDX],
SAMPLE_RATE)
sf.write('{}_reconstructed.wav'.format(model),
tf.squeeze(reconstructed), SAMPLE_RATE)
encoder_basis = tf.transpose(tf.squeeze(encoder.conv_layer.kernel))
decoder_basis = tf.transpose(
tf.squeeze(decoder.transpose_conv_layer.conv_2d.kernel))
sorted_encoder_basis, sorted_encoder_ids = sort_basis_functions(
encoder_basis)
sorted_decoder_basis, _ = sort_basis_functions(decoder_basis)
save_plot(sorted_encoder_basis, '{}_encoder.png'.format(model))
save_plot(sorted_decoder_basis, '{}_decoder.png'.format(model))
plot_basis_functions_frequency_domain(
sorted_encoder_basis, '{}_encoder_frequency.png'.format(model))
plot_basis_functions_frequency_domain(
sorted_decoder_basis, '{}_decoder_frequency.png'.format(model))
spectogram_like = tf.squeeze(spectogram_like)
spectogram_like = np.array(spectogram_like)[:, sorted_encoder_ids]
save_plot(spectogram_like, '{}_decomposition.png'.format(model))
def main():
os.environ['CUDA_VISIBLE_DEVICES'] = '1'
print('Num GPUs Available: ',
len(tf.config.experimental.list_physical_devices('GPU')))
raw_dataset = waveform_dataset.get_waveform_dataset(DATASET_PATH)
generator = wave_gan.Generator()
discriminator = wave_gan.Discriminator(input_shape=WAVEFORM_SHAPE)
generator_optimizer = tf.keras.optimizers.Adam(1e-4,
beta_1=0.5,
beta_2=0.9)
discriminator_optimizer = tf.keras.optimizers.Adam(1e-4,
beta_1=0.5,
beta_2=0.9)
get_waveform = lambda waveform: waveform
save_examples = lambda epoch, real, generated:\
save_helper.save_wav_data(
epoch, real, generated, SAMPLING_RATE, RESULT_DIR, get_waveform
)
wave_gan_model = wgan.WGAN(raw_dataset,
generator, [discriminator],
Z_DIM,
generator_optimizer,
discriminator_optimizer,
discriminator_training_ratio=D_UPDATES_PER_G,
batch_size=BATCH_SIZE,
epochs=EPOCHS,
checkpoint_dir=CHECKPOINT_DIR,
fn_save_examples=save_examples)
wave_gan_model.restore('ckpt-180', 1800)
wave_gan_model.train()