def __init__(
self,
dir_list,
width=28,
height=28,
channel=1,
initializer=None,
batch_size=40,
learning_rate=1e-03,
ctx=mx.gpu(),
discriminative_model=None,
generative_model=None,
):
'''
Init.
If you are not satisfied with this simple default setting,
delegate `discriminative_model` and `generative_model` designed by yourself.
Args:
dir_list: `list` of `str` of path to image files.
width: `int` of image width.
height: `int` of image height.
channel: `int` of image channel.
initializer: is-a `mxnet.initializer` for parameters of model.
If `None`, it is drawing from the Xavier distribution.
batch_size: `int` of batch size.
learning_rate: `float` of learning rate.
ctx: `mx.gpu()` or `mx.cpu()`.
discriminative_model: is-a `accelbrainbase.observabledata._mxnet.adversarialmodel.discriminative_model.DiscriminativeModel`.
generative_model: is-a `accelbrainbase.observabledata._mxnet.adversarialmodel.generative_model.GenerativeModel`.
'''
image_extractor = ImageExtractor(
width=width,
height=height,
channel=channel,
ctx=ctx
)
unlabeled_image_iterator = UnlabeledImageIterator(
image_extractor=image_extractor,
dir_list=dir_list,
batch_size=batch_size,
norm_mode="z_score",
scale=1.0,
noiseable_data=GaussNoise(sigma=1e-03, mu=0.0),
)
true_sampler = TrueSampler()
true_sampler.iteratorable_data = unlabeled_image_iterator
condition_sampler = ConditionSampler()
condition_sampler.true_sampler = true_sampler
computable_loss = L2NormLoss()
if discriminative_model is None:
output_nn = NeuralNetworks(
computable_loss=computable_loss,
initializer=initializer,
learning_rate=learning_rate,
learning_attenuate_rate=1.0,
attenuate_epoch=50,
units_list=[100, 1],
dropout_rate_list=[0.5, 0.0],
optimizer_name="SGD",
activation_list=["relu", "sigmoid"],
hidden_batch_norm_list=[BatchNorm(), None],
ctx=ctx,
hybridize_flag=True,
regularizatable_data_list=[],
scale=1.0,
output_no_bias_flag=True,
all_no_bias_flag=True,
not_init_flag=False,
)
d_model = ConvolutionalNeuralNetworks(
computable_loss=computable_loss,
initializer=initializer,
learning_rate=learning_rate,
learning_attenuate_rate=1.0,
attenuate_epoch=50,
hidden_units_list=[
Conv2D(
channels=16,
kernel_size=6,
strides=(2, 2),
padding=(1, 1),
),
Conv2D(
channels=32,
kernel_size=3,
strides=(2, 2),
padding=(1, 1),
),
],
input_nn=None,
input_result_height=None,
input_result_width=None,
input_result_channel=None,
output_nn=output_nn,
hidden_dropout_rate_list=[0.5, 0.5],
hidden_batch_norm_list=[BatchNorm(), BatchNorm()],
optimizer_name="SGD",
hidden_activation_list=["relu", "relu"],
hidden_residual_flag=False,
hidden_dense_flag=False,
dense_axis=1,
ctx=ctx,
hybridize_flag=True,
regularizatable_data_list=[],
scale=1.0,
)
discriminative_model = DiscriminativeModel(
model=d_model,
initializer=None,
learning_rate=learning_rate,
optimizer_name="SGD",
hybridize_flag=True,
scale=1.0,
ctx=ctx,
)
else:
if isinstance(discriminative_model, DiscriminativeModel) is False:
raise TypeError("The type of `discriminative_model` must be `DiscriminativeModel`.")
if generative_model is None:
g_model = ConvolutionalNeuralNetworks(
computable_loss=computable_loss,
initializer=initializer,
learning_rate=learning_rate,
learning_attenuate_rate=1.0,
attenuate_epoch=50,
hidden_units_list=[
Conv2DTranspose(
channels=16,
kernel_size=6,
strides=(1, 1),
padding=(1, 1),
),
Conv2DTranspose(
channels=1,
kernel_size=3,
strides=(1, 1),
padding=(1, 1),
),
],
input_nn=None,
input_result_height=None,
input_result_width=None,
input_result_channel=None,
output_nn=None,
hidden_dropout_rate_list=[0.5, 0.0],
hidden_batch_norm_list=[BatchNorm(), None],
optimizer_name="SGD",
hidden_activation_list=["relu", "identity"],
hidden_residual_flag=False,
hidden_dense_flag=False,
dense_axis=1,
ctx=ctx,
hybridize_flag=True,
regularizatable_data_list=[],
scale=1.0,
)
generative_model = GenerativeModel(
noise_sampler=UniformNoiseSampler(
low=-1e-05,
high=1e-05,
batch_size=batch_size,
seq_len=0,
channel=channel,
height=height,
width=width,
ctx=ctx
),
model=g_model,
initializer=None,
condition_sampler=condition_sampler,
conditonal_dim=1,
learning_rate=learning_rate,
optimizer_name="SGD",
hybridize_flag=True,
scale=1.0,
ctx=ctx,
)
else:
if isinstance(generative_model, GenerativeModel) is False:
raise TypeError("The type of `generative_model` must be `GenerativeModel`.")
GAN = GANController(
true_sampler=true_sampler,
generative_model=generative_model,
discriminative_model=discriminative_model,
generator_loss=GeneratorLoss(weight=1.0),
discriminator_loss=DiscriminatorLoss(weight=1.0),
feature_matching_loss=L2NormLoss(weight=1.0),
optimizer_name="SGD",
learning_rate=learning_rate,
learning_attenuate_rate=1.0,
attenuate_epoch=50,
hybridize_flag=True,
scale=1.0,
ctx=ctx,
initializer=initializer,
)
self.GAN = GAN
def __init__(
self,
midi_path_list,
batch_size=20,
seq_len=8,
time_fraction=1.0,
learning_rate=1e-10,
learning_attenuate_rate=0.1,
attenuate_epoch=50,
generative_model=None,
discriminative_model=None,
ctx=mx.gpu(),
initializer=None,
):
'''
Init.
Args:
midi_path_list: `list` of paths to MIDI files.
batch_size: Batch size.
seq_len: The length of sequence that LSTM networks will observe.
time_fraction: Time fraction or time resolution (seconds).
learning_rate: Learning rate in `Generator` and `Discriminator`.
learning_attenuate_rate: Attenuate the `learning_rate` by a factor of this value every `attenuate_epoch`.
attenuate_epoch: Attenuate the `learning_rate` by a factor of `learning_attenuate_rate` every `attenuate_epoch`.
true_sampler: is-a `TrueSampler`.
noise_sampler: is-a `NoiseSampler`.
generative_model: is-a `GenerativeModel`.
discriminative_model: is-a `DiscriminativeModel`.
ctx: `mx.cpu()` or `mx.gpu()`.
initializer: is-a `mxnet.initializer` for parameters of model. If `None`, it is drawing from the Xavier distribution.
'''
computable_loss = mx.gluon.loss.SoftmaxCrossEntropyLoss(
sparse_label=False)
self.__midi_controller = MidiController()
self.__midi_df_list = [
self.__midi_controller.extract(midi_path)
for midi_path in midi_path_list
]
bar_gram = BarGram(midi_df_list=self.__midi_df_list,
time_fraction=time_fraction)
self.__bar_gram = bar_gram
dim = self.__bar_gram.dim
c_true_sampler = ConditionalBarGramTrueSampler(
bar_gram=bar_gram,
midi_df_list=self.__midi_df_list,
batch_size=batch_size,
seq_len=seq_len,
time_fraction=time_fraction)
true_sampler = BarGramTrueSampler(bar_gram=bar_gram,
midi_df_list=self.__midi_df_list,
batch_size=batch_size,
seq_len=seq_len,
time_fraction=time_fraction)
if generative_model is None:
condition_sampler = ConditionSampler()
condition_sampler.true_sampler = true_sampler
c_model = ConvolutionalNeuralNetworks(
computable_loss=computable_loss,
initializer=initializer,
learning_rate=learning_rate,
learning_attenuate_rate=1.0,
attenuate_epoch=50,
hidden_units_list=[
Conv2D(
channels=16,
kernel_size=6,
strides=(1, 1),
padding=(1, 1),
),
Conv2D(
channels=len(true_sampler.program_list),
kernel_size=6,
strides=(1, 1),
padding=(1, 1),
),
],
input_nn=None,
input_result_height=None,
input_result_width=None,
input_result_channel=None,
output_nn=None,
hidden_dropout_rate_list=[0.5, 0.0],
hidden_batch_norm_list=[BatchNorm(), None],
optimizer_name="SGD",
hidden_activation_list=["relu", "identity"],
hidden_residual_flag=False,
hidden_dense_flag=False,
dense_axis=1,
ctx=ctx,
hybridize_flag=True,
regularizatable_data_list=[],
scale=1.0,
)
condition_sampler.model = c_model
g_model = ConvolutionalNeuralNetworks(
computable_loss=computable_loss,
initializer=initializer,
learning_rate=learning_rate,
learning_attenuate_rate=1.0,
attenuate_epoch=50,
hidden_units_list=[
Conv2DTranspose(
channels=16,
kernel_size=6,
strides=(1, 1),
padding=(1, 1),
),
Conv2DTranspose(
channels=len(true_sampler.program_list),
kernel_size=6,
strides=(1, 1),
padding=(1, 1),
),
],
input_nn=None,
input_result_height=None,
input_result_width=None,
input_result_channel=None,
output_nn=None,
hidden_dropout_rate_list=[0.5, 0.0],
hidden_batch_norm_list=[BatchNorm(), None],
optimizer_name="SGD",
hidden_activation_list=["relu", "identity"],
hidden_residual_flag=False,
hidden_dense_flag=False,
dense_axis=1,
ctx=ctx,
hybridize_flag=True,
regularizatable_data_list=[],
scale=1.0,
)
generative_model = GenerativeModel(
noise_sampler=None,
model=g_model,
initializer=None,
condition_sampler=condition_sampler,
conditonal_dim=1,
learning_rate=learning_rate,
optimizer_name="SGD",
hybridize_flag=True,
scale=1.0,
ctx=ctx,
)
else:
if isinstance(generative_model, GenerativeModel) is False:
raise TypeError(
"The type of `generative_model` must be `GenerativeModel`."
)
if discriminative_model is None:
output_nn = NeuralNetworks(
computable_loss=computable_loss,
initializer=initializer,
learning_rate=learning_rate,
learning_attenuate_rate=1.0,
attenuate_epoch=50,
units_list=[100, 1],
dropout_rate_list=[0.5, 0.0],
optimizer_name="SGD",
activation_list=["relu", "sigmoid"],
hidden_batch_norm_list=[BatchNorm(), None],
ctx=ctx,
hybridize_flag=True,
regularizatable_data_list=[],
scale=1.0,
output_no_bias_flag=True,
all_no_bias_flag=True,
not_init_flag=False,
)
d_model = ConvolutionalNeuralNetworks(
computable_loss=computable_loss,
initializer=initializer,
learning_rate=learning_rate,
learning_attenuate_rate=1.0,
attenuate_epoch=50,
hidden_units_list=[
Conv2D(
channels=16,
kernel_size=6,
strides=(2, 2),
padding=(1, 1),
),
Conv2D(
channels=32,
kernel_size=3,
strides=(2, 2),
padding=(1, 1),
),
],
input_nn=None,
input_result_height=None,
input_result_width=None,
input_result_channel=None,
output_nn=output_nn,
hidden_dropout_rate_list=[0.5, 0.5],
hidden_batch_norm_list=[BatchNorm(), BatchNorm()],
optimizer_name="SGD",
hidden_activation_list=["relu", "relu"],
hidden_residual_flag=False,
hidden_dense_flag=False,
dense_axis=1,
ctx=ctx,
hybridize_flag=True,
regularizatable_data_list=[],
scale=1.0,
)
discriminative_model = DiscriminativeModel(
model=d_model,
initializer=None,
learning_rate=learning_rate,
optimizer_name="SGD",
hybridize_flag=True,
scale=1.0,
ctx=ctx,
)
else:
if isinstance(discriminative_model, DiscriminativeModel) is False:
raise TypeError(
"The type of `discriminative_model` must be `DiscriminativeModel`."
)
GAN = GANController(
true_sampler=c_true_sampler,
generative_model=generative_model,
discriminative_model=discriminative_model,
generator_loss=GeneratorLoss(weight=1.0),
discriminator_loss=DiscriminatorLoss(weight=1.0),
feature_matching_loss=L2NormLoss(weight=1.0),
optimizer_name="SGD",
learning_rate=learning_rate,
learning_attenuate_rate=1.0,
attenuate_epoch=50,
hybridize_flag=True,
scale=1.0,
ctx=ctx,
initializer=initializer,
)
self.__true_sampler = true_sampler
self.__generative_model = generative_model
self.__discriminative_model = discriminative_model
self.__GAN = GAN
self.__time_fraction = time_fraction