def network(self):
""" Assemble Critic network to predict q-values
"""
state = Input((self.env_dim))
x = Dense(32, activation='elu')(state)
x = Dense(16, activation='elu')(x)
out = Dense(1, activation='linear',
kernel_initializer=RandomUniform())(x)
return Model(state, out)
def create_network(self, S, G, num_A, dropout, l2reg):
h = concatenate([S, G])
for l in self.layers:
h = Dense(l, activation="relu", kernel_initializer=he_normal())(h)
Q_values = Dense(num_A,
activation='linear',
kernel_initializer=RandomUniform(minval=-3e-4,
maxval=3e-4))(h)
return Q_values
def model(self, input: Any) -> Layer:
size = len(self.alphabet())
initializer = RandomUniform(minval=-0.5, maxval=0.5)
embedding = TimeDistributed(
Embedding(size, 50, embeddings_initializer=initializer))(input)
embedding = SpatialDropout1D(self.__droput)(embedding)
output = TimeDistributed(Bidirectional(CuDNNLSTM(100)))(embedding)
output = SpatialDropout1D(self.__droput)(output)
return output
def __init__(self, env):
np.random.seed(123)
env.seed(123)
assert len(env.action_space.shape) == 1
nb_actions = env.action_space.shape[0]
# Next, we build a very simple model.
self.actor = Sequential()
self.actor.add(Flatten(input_shape=(1, ) +
env.observation_space.shape))
self.actor.add(Dense(16))
self.actor.add(Activation('relu'))
self.actor.add(Dense(16))
self.actor.add(Activation('relu'))
self.actor.add(Dense(16))
self.actor.add(Activation('relu'))
self.actor.add(
Dense(nb_actions,
activation='tanh',
kernel_initializer=RandomUniform()))
self.actor.add(Lambda(lambda x: x * 60.0))
print(self.actor.summary())
action_input = Input(shape=(nb_actions, ), name='action_input')
observation_input = Input(shape=(1, ) + env.observation_space.shape,
name='observation_input')
flattened_observation = Flatten()(observation_input)
x = Concatenate()([action_input, flattened_observation])
x = Dense(32)(x)
x = Activation('relu')(x)
x = Dense(32)(x)
x = Activation('relu')(x)
x = Dense(32)(x)
x = Activation('relu')(x)
x = Dense(1)(x)
x = Activation('linear')(x)
critic = Model(inputs=[action_input, observation_input], outputs=x)
print(critic.summary())
# Finally, we configure and compile our agent. You can use every built-in Keras optimizer and
# even the metrics!
memory = SequentialMemory(limit=100000, window_length=1)
random_process = OrnsteinUhlenbeckProcess(size=nb_actions,
theta=.15,
mu=0.,
sigma=.3)
self.agent = DDPGAgent(nb_actions=nb_actions,
actor=self.actor,
critic=critic,
critic_action_input=action_input,
memory=memory,
nb_steps_warmup_critic=100,
nb_steps_warmup_actor=100,
random_process=random_process,
gamma=.99,
target_model_update=1e-3)
self.agent.compile(Adam(lr=.001, clipnorm=1.), metrics=['mae'])
def build(self, input_shape):
initializer_uniform = RandomUniform(minval=0, maxval=1)
constraint_min_max = min_max_norm(min_value=0.0, max_value=1.0)
self.b = self.add_weight(name='b',
shape=(input_shape[-1], ),
initializer=initializer_uniform,
constraint=constraint_min_max,
trainable=True)
super(ANDNoisy, self).build(input_shape)
def __init__(self, output_dim, initializer=None, betas=1.0, **kwargs):
self.output_dim = output_dim
self.init_betas = betas
if not initializer:
self.initializer = RandomUniform(0.0, 1.0)
#self.initializer = Orthogonal()
else:
self.initializer = initializer
super(RBFLayer, self).__init__(**kwargs)
def _add_rnn_layer(self, rnn, return_sequences, x):
if self.params["rnn_bidirectional"][x] == False:
self.cnn = rnn(
units=self.params["rnn_units"][x],
dropout=self.params["rnn_dropout_input"][x],
recurrent_dropout=self.params["rnn_dropout_recurrent"][x],
kernel_initializer=RandomUniform(),
kernel_constraint=max_norm(self.params["kernel_constraint"]),
return_sequences=return_sequences)(self.cnn)
else:
self.cnn = Bidirectional(
rnn(units=self.params["rnn_units"][x],
dropout=self.params["rnn_dropout_input"][x],
recurrent_dropout=self.params["rnn_dropout_recurrent"][x],
kernel_initializer=RandomUniform(),
kernel_constraint=max_norm(
self.params["kernel_constraint"]),
return_sequences=return_sequences))(self.cnn)
def get_fI(STATE_SIZE, ACTION_COUNT):
model = Sequential(name='f_I')
model.add(
Dense(2048,
activation="relu",
input_shape=(STATE_SIZE * 2, ),
kernel_initializer=RandomUniform(-0.1, 0.1)))
model.add(Dense(ACTION_COUNT))
return model
def get_fM(STATE_SIZE, ACTION_COUNT, AGENT_HISTORY_LENGTH):
state = Input(shape=(STATE_SIZE, ), name='current_state')
actions = Input(shape=(ACTION_COUNT * AGENT_HISTORY_LENGTH, ),
name='action_performed')
x = Dense(2048, kernel_initializer=RandomUniform(-0.1, 0.1))(actions)
x = Multiply()([x, state])
next_state = Dense(2048, name='next_state')(x)
model = Model(inputs=[state, actions], outputs=[next_state], name='f_M') #
return model
def _model(
self,
*,
embedding_dim=2,
embedding_max_size=10000,
forecaster_features=1,
forecaster_hidden_units=(8, 8),
lr=0.1
):
# Embedder
embedding_initialize = RandomUniform(minval=-1, maxval=1)
customer_in = Input((1,))
embedding_out = Embedding(input_dim=embedding_max_size,
output_dim=embedding_dim, input_length=1,
embeddings_initializer=embedding_initialize
)(customer_in)
embedding_out = Flatten()(embedding_out)
embedding_model = Model(
customer_in,
embedding_out,
name='Embedder'
)
# Forecaster
features_in = Input((forecaster_features,))
embedding_in = Input((embedding_dim,))
forecaster_output = Concatenate()([features_in, embedding_in])
# append final output
forecaster_dense_units = list(forecaster_hidden_units) + [1]
for idx, units in enumerate(forecaster_dense_units):
forecaster_output = Dense(
units=units,
activation='relu',
name='forecaster_dense_{}'.format(idx)
)(forecaster_output)
forecaster_model = Model(
[features_in, embedding_in],
forecaster_output,
name='Forecaster'
)
# Combined model
combined_output = forecaster_model(
[features_in, embedding_model(customer_in)]
)
combined_model = Model(
[features_in, customer_in],
combined_output,
name='Combined'
)
optimizer = Adam(lr=lr)
combined_model.compile(optimizer=optimizer, loss='mse')
return {
'forecaster': forecaster_model,
'embedder': embedding_model,
'combined': combined_model
}
def create_model(self, critic_lr, channels, kernels, strides, activations,
linear_units, units, low_dimensional):
initializer = RandomUniform(-3e-4, 3e-4)
self.obs = tf.placeholder(tf.float32, (None, ) + self.obs_shape)
self.action = tf.placeholder(tf.float32, (None, ) + self.action_shape)
obs = Input(tensor=self.obs)
action = Input(tensor=self.action)
i = 0
x = obs
if low_dimensional:
for u in units:
x = Dense(u,
activation='relu',
kernel_initializer=initializer,
bias_initializer=initializer)(x)
y = Dense(units[-1],
activation='relu',
kernel_initializer=initializer,
bias_initializer=initializer)(action)
x = Add()([x, y])
else:
for c, k, s, a in zip(channels, kernels, strides, activations):
x = Conv2D(c,
k,
strides=(s, s),
activation=a,
kernel_initializer=initializer,
bias_initializer=initializer)(x)
x = Flatten()(x)
# this to output layer
x = Dense(linear_units,
activation='relu',
kernel_initializer=initializer,
bias_initializer=initializer)(x)
y = Dense(linear_units,
activation='relu',
kernel_initializer=initializer,
bias_initializer=initializer)(action)
x = Add()([x, y])
output = Dense(1,
kernel_initializer=initializer,
bias_initializer=initializer)(x)
self.output = output
self.model = Model(inputs=[obs, action], outputs=output)
self.gradients = tf.gradients(self.output, self.action)
self.q_targets = tf.placeholder(tf.float32, (None, 1))
self.loss = tf.reduce_mean(
tf.squared_difference(self.output, self.q_targets))
self.optimizer = tf.train.AdamOptimizer(critic_lr)
self.params = tf.trainable_variables(scope='critic')
grads = self.optimizer.compute_gradients(self.loss)
clipped_grads = [(tf.clip_by_value(grad, -1.0, 1.0), var)
for grad, var in grads]
self.optimize = self.optimizer.apply_gradients(clipped_grads)
def get_model(self):
word_embeddings = self.word_embeddings
case_embeddings = self.case_embeddings
char_idx = self.char_idx
label_idx = self.label_idx
words_input = Input(shape=(None, ), dtype='int32', name='words_input')
words = Embedding(input_dim=word_embeddings.shape[0],
output_dim=word_embeddings.shape[1],
weights=[word_embeddings],
trainable=False)(words_input)
casing_input = Input(shape=(None, ),
dtype='int32',
name='casing_input')
casing = Embedding(output_dim=case_embeddings.shape[1],
input_dim=case_embeddings.shape[0],
weights=[case_embeddings],
trainable=False)(casing_input)
character_input = Input(shape=(
None,
52,
), name='char_input')
embed_char_out = TimeDistributed(
Embedding(len(char_idx),
30,
embeddings_initializer=RandomUniform(minval=-0.5,
maxval=0.5)),
name='char_embedding')(character_input)
dropout = Dropout(0.5)(embed_char_out)
conv1d_out = TimeDistributed(
Conv1D(kernel_size=3,
filters=30,
padding='same',
activation='tanh',
strides=1))(dropout)
maxpool_out = TimeDistributed(MaxPooling1D(52))(conv1d_out)
char = TimeDistributed(Flatten())(maxpool_out)
char = Dropout(0.5)(char)
output = concatenate([words, casing, char])
output = Bidirectional(
LSTM(200,
return_sequences=True,
dropout=0.50,
recurrent_dropout=0.25))(output)
output = TimeDistributed(Dense(len(label_idx),
activation='softmax'))(output)
model = Model(inputs=[words_input, casing_input, character_input],
outputs=[output])
model.compile(loss='sparse_categorical_crossentropy',
optimizer='nadam')
model.summary()
return model
def best_learning_rate(train_x, train_y, test_x, test_y, lr, batch_size=256):
"""
Perform 10 random initializations of model, then compile and fit it.
Model is initialized randomly (random_uniform), compiled and fitted.
Operation is repeated 10 times. Then all histories are returned as a list.
Parameters
----------
train_x : np.array(float)
Training `x` set (training features).
train_y : np.array(int)
Training `y` set (training labels).
test_x : np.array(float)
Test `x` set (test features).
test_y : np.array(int)
Test `y` set (test labels).
lr : float
Learning rate.
batch_size : int, optional
Size of batch (amount of samples) for model fitting.
Returns
-------
history_set : list(keras.callbacks.History object)
History of loss, accuracy, validation loss and validation
accuracy during model fitting.
"""
optimizer = TFOptimizer(tf.train.GradientDescentOptimizer(lr))
history_set = []
for i in range(10):
model = Sequential()
initializer = RandomUniform(minval=-1.0, maxval=1.0, seed=None)
model.add(
Dense(1,
kernel_initializer=initializer,
bias_initializer=initializer,
input_dim=train_x.shape[1],
activation='sigmoid'))
model.compile(loss='binary_crossentropy',
optimizer=optimizer,
metrics=['accuracy'])
history = model.fit(train_x,
train_y,
epochs=1000,
validation_data=(test_x, test_y),
batch_size=batch_size,
verbose=0)
history_set.append(history)
return history_set
def __init__(self,
learning_rate=None,
vocab_size=None,
embedding_size=None,
rnn_output_size=None,
dropout_rate=None,
bidirectional_rnn=None,
rnn_type=None,
rnn_layers=None,
l1_reg=None,
l2_reg=None,
initializer=None,
word_vector_init=None):
"""
If an arg is None, it will get its value from config.active_config.
"""
self._learning_rate = learning_rate or active_config().learning_rate
self._vocab_size = vocab_size or active_config().vocab_size
self._embedding_size = embedding_size or active_config().embedding_size
self._rnn_output_size = (rnn_output_size
or active_config().rnn_output_size)
self._dropout_rate = dropout_rate or active_config().dropout_rate
self._rnn_type = rnn_type or active_config().rnn_type
self._rnn_layers = rnn_layers or active_config().rnn_layers
self._word_vector_init = (word_vector_init
or active_config().word_vector_init)
self._initializer = initializer or active_config().initializer
if self._initializer == 'vinyals_uniform':
self._initializer = RandomUniform(-0.08, 0.08)
if bidirectional_rnn is None:
self._bidirectional_rnn = active_config().bidirectional_rnn
else:
self._bidirectional_rnn = bidirectional_rnn
l1_reg = l1_reg or active_config().l1_reg
l2_reg = l2_reg or active_config().l2_reg
self._regularizer = l1_l2(l1_reg, l2_reg)
self._keras_model = None
if self._vocab_size is None:
raise ValueError('config.active_config().vocab_size cannot be '
'None! You should check your config or you can '
'explicitly pass the vocab_size argument.')
if self._rnn_type not in ('lstm', 'gru'):
raise ValueError('rnn_type must be either "lstm" or "gru"!')
if self._rnn_layers < 1:
raise ValueError('rnn_layers must be >= 1!')
if self._word_vector_init is not None and self._embedding_size != 300:
raise ValueError('If word_vector_init is not None, embedding_size '
'must be 300')
def test_name_entity_recognition(self):
K.clear_session()
words_input = Input(shape=(None, ), dtype='int32', name='words_input')
words = Embedding(input_dim=10,
output_dim=20,
weights=None,
trainable=False)(words_input)
casing_input = Input(shape=(None, ),
dtype='int32',
name='casing_input')
casing = Embedding(output_dim=20,
input_dim=12,
weights=None,
trainable=False)(casing_input)
character_input = Input(shape=(
None,
52,
), name='char_input')
embed_char_out = TimeDistributed(
Embedding(26,
20,
embeddings_initializer=RandomUniform(minval=-0.5,
maxval=0.5)),
name='char_embedding')(character_input)
dropout = Dropout(0.5)(embed_char_out)
conv1d_out = TimeDistributed(
Conv1D(kernel_size=3,
filters=30,
padding='same',
activation='tanh',
strides=1))(dropout)
maxpool_out = TimeDistributed(MaxPooling1D(52))(conv1d_out)
char = TimeDistributed(Flatten())(maxpool_out)
char = Dropout(0.5)(char)
output = concatenate([words, casing, char])
output = Bidirectional(
LSTM(200,
return_sequences=True,
dropout=0.50,
recurrent_dropout=0.25))(output)
output = TimeDistributed(Dense(35, activation='softmax'))(output)
keras_model = Model(
inputs=[words_input, casing_input, character_input],
outputs=[output])
data1 = np.random.rand(2, 6).astype(np.float32)
data2 = np.random.rand(2, 6).astype(np.float32)
data3 = np.random.rand(2, 6, 52).astype(np.float32)
expected = keras_model.predict([data1, data2, data3])
onnx_model = keras2onnx.convert_keras(keras_model, keras_model.name)
self.assertTrue(
run_keras_and_ort(onnx_model.graph.name,
onnx_model,
keras_model, [data1, data2, data3],
expected,
self.model_files,
compare_perf=True))
def Minecraft_DDPG(window_length, grayscale, width, height, nb_actions):
assert width == 32 and height == 32, 'Model accepts 32x32 input size'
if grayscale:
channels = 1
else:
channels = 3
if K.image_data_format() == 'channels_last':
observation_shape = (32, 32, window_length * channels)
else:
observation_shape = (window_length * channels, 32, 32)
# Build actor and critic networks
inputs = Input(shape=observation_shape)
x = Conv2D(32, (4, 4), strides=(2, 2), activation='relu')(inputs)
x = Conv2D(64, (3, 3), strides=(2, 2), activation='relu')(x)
x = Flatten()(x)
x = Dense(256, activation='relu')(x)
x = Dense(256, activation='relu')(x)
x = Dense(nb_actions,
activation='tanh',
kernel_initializer=RandomUniform(-3e-4, 3e-4))(x)
actor = Model(inputs=inputs, outputs=x)
print(actor.summary())
# critic network has 2 inputs, one action input and one observation input.
action_input = Input(shape=(nb_actions, ), name='action_input')
observation_input = Input(shape=observation_shape,
name='observation_input')
x = Conv2D(32, (4, 4), strides=(2, 2),
activation='relu')(observation_input)
x = Conv2D(64, (3, 3), strides=(2, 2), activation='relu')(x)
x = Flatten()(x)
x = Dense(256, activation='relu')(x)
x = concatenate([x, action_input
]) # Actions are not included until the 2nd dense layer
x = Dense(256, activation='relu')(x)
x = Dense(1,
activation='linear',
kernel_initializer=RandomUniform(-3e-4, 3e-4))(x)
critic = Model(inputs=[action_input, observation_input], outputs=x)
print(critic.summary())
return actor, critic, action_input
def create_critic_network(self, S, V):
if self.network == '0':
L1 = concatenate([multiply([subtract([S, G]), M]), S])
L2 = Dense(400,
activation="relu",
kernel_initializer=lecun_uniform(),
kernel_regularizer=l2(0.01))(L1)
L3 = Dense(300,
activation="relu",
kernel_initializer=lecun_uniform(),
kernel_regularizer=l2(0.01))(L2)
Q_values = Dense(self.env.action_dim,
activation='linear',
kernel_initializer=RandomUniform(minval=-3e-4,
maxval=3e-4),
kernel_regularizer=l2(0.01),
bias_initializer=RandomUniform(minval=-3e-4,
maxval=3e-4))(L3)
else:
L1 = Dense(200,
activation="relu",
kernel_initializer=lecun_uniform(),
kernel_regularizer=l2(0.01))
L2 = Dense(300,
activation="relu",
kernel_initializer=lecun_uniform(),
kernel_regularizer=l2(0.01))
i1 = multiply([subtract([S, G]), M])
i2 = S
h1 = L1(i1)
h2 = L1(i2)
h3 = concatenate([h1, h2])
h4 = L2(h3)
Q_values = Dense(self.env.action_dim,
activation='linear',
kernel_initializer=RandomUniform(minval=-3e-4,
maxval=3e-4),
kernel_regularizer=l2(0.01),
bias_initializer=RandomUniform(minval=-3e-4,
maxval=3e-4))(h4)
return Q_values
def generator_model(noise_dim, feature_dim):
noise_input = keras.layers.Input(shape=(noise_dim, ))
eeg_input = keras.layers.Input(shape=(feature_dim, ))
x = MoGLayer(kernel_initializer=RandomUniform(minval=-0.2, maxval=0.2),
bias_initializer=RandomUniform(minval=-1.0, maxval=1.0),
kernel_regularizer=l2(0.01))(noise_input)
x = keras.layers.concatenate([x, eeg_input])
x = Dense(1024, activation="tanh")(x)
x = BatchNormalization(momentum=0.8)(x)
x = Dense(128 * 7 * 7, activation="tanh")(x)
x = Reshape((7, 7, 128))(x)
x = UpSampling2D()(x)
x = BatchNormalization(momentum=0.8)(x)
x = Conv2D(64, kernel_size=5, padding="same", activation="tanh")(x)
x = UpSampling2D()(x)
x = Conv2D(1, kernel_size=3, padding="same")(x)
output = Activation("tanh")(x)
return Model(inputs=[noise_input, eeg_input], outputs=[output])
def network(self):
state = Input((self.state_dim,))
action = Input((self.action_dim,))
x = Dense(800, activation='relu')(state)
#x = concatenate([Flatten()(x), action])
x = concatenate([x, action])
x = Dense(500, activation='relu')(x)
out = Dense(1, activation='linear', kernel_initializer=RandomUniform())(x)
return Model([state, action], out)
def network(self):
""" Assemble Critic network to predict q-values
"""
state = Input(shape=[self.env_dim])
action = Input(shape=[self.act_dim])
x = Dense(256, activation='relu')(state)
x = concatenate([x, action])
x = Dense(256, activation='relu')(x)
out = Dense(1, activation='linear', kernel_initializer=RandomUniform())(x)
return Model([state, action], out)
def build(self, shape_input):
M = shape_input[1] - 1
self.I = k_back.eye(M)
init_mu = RandomUniform(minval=0.01, maxval=10)
init_pfd = RandomUniform(minval=0.01, maxval=10)
self.mu = self.add_weight('mu',
shape=(M, 1),
initializer=init_mu,
constraint=NonNeg())
data_p = self.add_weight('data_p',
shape=(M, M - 1),
initializer=init_pfd,
constraint=NonNeg())
data_p_scaled = data_p / k_back.sum(data_p, axis=1, keepdims=True)
self.P = k_back.reshape(
k_back.flatten(data_p_scaled)[None, :] @ k_back.one_hot(
[j for j in range(M * M) if j % (M + 1) != 0], M * M), (M, M))
self.odot = (self.P - self.I) * self.mu
self.is_built = True
def create_critic_network(self, S, G=None, M=None):
input = concatenate([multiply([subtract([S, G]), M]), S])
L1 = Dense(400,
activation="relu",
kernel_initializer=lecun_uniform(),
kernel_regularizer=l2(0.01))
L1out = L1(input)
L2 = Dense(300,
activation="relu",
kernel_initializer=lecun_uniform(),
kernel_regularizer=l2(0.01))
L2out = L2(L1out)
L3 = Dense(self.env.action_dim,
activation='linear',
kernel_initializer=RandomUniform(minval=-3e-4, maxval=3e-4),
kernel_regularizer=l2(0.01),
bias_initializer=RandomUniform(minval=-3e-4, maxval=3e-4))
Q_values = L3(L2out)
return [L1, L2, L3], Q_values
def network(self):
state = Input((self.env_dim, ))
action = Input((self.act_dim, ))
x = Dense(128, activation='relu')(state)
x = concatenate([x, action])
x = Dense(256, activation='relu')(x)
out = Dense(self.env_dim,
activation='linear',
kernel_initializer=RandomUniform())(x)
return Model([state, action], out)
def build(self):
phi = Input(name='Xs', shape=(21, ))
# Learning to Rank
out_ = Dense(1,
kernel_initializer=RandomUniform(minval=-0.014,
maxval=0.014),
bias_initializer='zeros')(phi)
model = Model(inputs=[phi], outputs=[out_])
return model
def __init__(self,
learning_rate=None,
vocab_size=None,
embedding_size=None,
rnn_output_size=None,
dropout_rate=None,
bidirectional_rnn=None,
rnn_type=None,
rnn_layers=None,
l1_reg=None,
l2_reg=None,
initializer=None,
word_vector_init=None):
self.keras_model = None
self.learning_rate = learning_rate or base_configuration["params"][
"learning_rate"]
self.vocab_size = vocab_size or base_configuration["params"][
"vocab_size"]
self.embedding_size = embedding_size or base_configuration["params"][
"embedding_size"]
self.rnn_output_size = rnn_output_size or base_configuration["params"][
"rnn_output_size"]
self.dropout_rate = dropout_rate or base_configuration["params"][
"dropout_rate"]
self.rnn_type = rnn_type or base_configuration["params"]["rnn_type"]
self.rnn_layers = rnn_layers or base_configuration["params"][
"rnn_layers"]
self.word_vector_init = word_vector_init or base_configuration[
"params"]["word_vector_init"]
self.initializer = initializer or base_configuration["params"][
"initializer"]
if self.initializer == 'vinyals_uniform':
self.initializer = RandomUniform(-0.08, 0.08)
self.bidirectional_rnn = bidirectional_rnn or base_configuration[
"params"]["bidirectional_rnn"]
self.regularizer = l1_l2() # (l1_reg, l2_reg)
if self.vocab_size is None:
raise ValueError(
'config.active_config().vocab_size cannot be None! You should check your config or you can'
' explicitly pass the vocab_size argument.')
if self.rnn_type not in ('lstm', 'gru'):
raise ValueError('rnn_type must be either "lstm" or "gru"!')
if self.rnn_layers < 1:
raise ValueError('rnn_layers must be >= 1!')
if self.word_vector_init is not None and self.embedding_size != 300:
raise ValueError(
'If word_vector_init is not None, embedding_size must be 300')
def main():
units = 512
rng_units = 128
z_k = 10
pz_regularizer = BalanceRegularizer(1e-2)
iwgan_weight = 1e-1
initializer = RandomUniform(minval=-0.05, maxval=0.05)
((x, y), (xt, yt)) = mnist.load_data()
x = np.float32(x) / 255.
x = np.reshape(x, (x.shape[0], -1))
input_units = 28 * 28
classifier = MLP(input_units=input_units,
hidden_units=units,
output_units=z_k,
hidden_depth=2,
hidden_activation=leaky_relu,
initializer=initializer,
output_activation=softmax_nd)
generator = MLP(input_units=units,
hidden_units=units,
output_units=input_units,
hidden_depth=2,
hidden_activation=leaky_relu,
initializer=initializer,
output_activation=T.nnet.sigmoid)
discriminator = MLP(input_units=units,
hidden_units=units,
output_units=1,
hidden_depth=2,
initializer=initializer,
hidden_activation=leaky_relu)
model = FCGAN(
z_k=z_k,
classifier=classifier,
generator=generator,
discriminator=discriminator,
optd=Adam(1e-3),
optg=Adam(1e-3),
input_units=input_units,
rng_units=rng_units,
units=units,
activation=leaky_relu,
pz_regularizer=pz_regularizer,
initializer=initializer,
iwgan_weight=iwgan_weight
)
model.train(
x=x,
output_path='output/mnist_clustering',
epochs=500,
batches=512,
discriminator_batches=5,
batch_size=128
)
def __init__(self, config):
self.config = config
recurrent_unit = self.config.recurrent_unit.lower()
get_custom_objects().update({'EncoderSlice': EncoderSlice, 'DecoderSlice': DecoderSlice})
initial_weights = RandomUniform(minval=-0.08, maxval=0.08, seed=config.seed)
stacked_input = Input(shape=(None,))
# encoder_input = Lambda(lambda x: x[:, config.input_split_index:])(stacked_input)
encoder_input = EncoderSlice(config.input_split_index)(stacked_input)
encoder_embedding = Embedding(config.source_vocab_size, config.embedding_dim,
weights=[config.source_embedding_map],
trainable=False)
encoder_embedded = encoder_embedding(encoder_input)
if recurrent_unit == 'lstm':
encoder = LSTM(self.config.hidden_dim, return_state=True, return_sequences=True,
recurrent_initializer=initial_weights)(encoder_embedded)
for i in range(1, self.config.num_encoder_layers):
encoder = LSTM(self.config.hidden_dim, return_state=True, return_sequences=True)(encoder)
_, state_h, state_c = encoder
encoder_states = [state_h, state_c]
else:
encoder = GRU(self.config.hidden_dim, return_state=True, return_sequences=True,
recurrent_initializer=initial_weights)(encoder_embedded)
for i in range(1, self.config.num_encoder_layers):
encoder = GRU(self.config.hidden_dim, return_state=True, return_sequences=True)(encoder)
_, state_h = encoder
encoder_states = [state_h]
# decoder_input = Lambda(lambda x: x[:, config.input_split_index:])(stacked_input)
decoder_input = DecoderSlice(config.input_split_index)(stacked_input)
decoder_embedding = Embedding(config.target_vocab_size, config.embedding_dim,
weights=[config.target_embedding_map],
trainable=False)
decoder_embedded = decoder_embedding(decoder_input)
if recurrent_unit.lower() == 'lstm':
decoder = LSTM(self.config.hidden_dim, return_state=True, return_sequences=True)(decoder_embedded, initial_state=encoder_states)
for i in range(1, self.config.num_decoder_layers):
decoder = LSTM(self.config.hidden_dim, return_state=True, return_sequences=True)(decoder)
decoder_output, decoder_state = decoder[0], decoder[1:]
else:
decoder = GRU(self.config.hidden_dim, return_state=True, return_sequences=True)(decoder_embedded, initial_state=encoder_states)
for i in range(1, self.config.num_decoder_layers):
decoder = GRU(self.config.hidden_dim, return_state=True, return_sequences=True)(decoder)
decoder_output, decoder_state = decoder[0], decoder[1]
decoder_dense = Dense(config.target_vocab_size, activation='softmax')
decoder_output = decoder_dense(decoder_output)
self.model = Model(stacked_input, decoder_output)
optimizer = Adam(lr=config.lr, clipnorm=25.)
self.model.compile(optimizer=optimizer, loss='categorical_crossentropy', metrics=['acc'])
print(self.model.summary())
def decoder(self, params):
print('genetic_params:', params)
NUM_NET_1 = 1 # for LR
NUM_NET_2 = 8 # for the rest network params
NUM_TIME = 3
NUM_DELAY_TYPE = 3
NUM_DELAY = 4
# netowr_params
BATCH_SIZE = [16, 32, 64, 128]
SEQ_LEN = [16, 32, 64, 128]
STATE_SIZE = [16, 32, 64, 128]
LR = list(np.logspace(-3, -6, 16))
DR = [0.99, 0.98, 0.97, 0.96]
PKEEP = [0.9, 0.8, 0.7, 0.6]
ACTIVATION = ["relu", "tanh", "sigmoid", "softsign"]
INIT = [zeros(), TruncatedNormal(), Orthogonal(), RandomUniform()]
net_name = ['lr', 'batch_size', 'seq_len', 'state_size', 'dr', 'pkeep', 'optimizer', 'activation_f', 'initializer']
network_params = {}
network_params['lr'] = LR[BitArray(params[0: NUM_NET_1 * 4]).uint]
for i in range(NUM_NET_2):
name = net_name[i + 1]
network_params[name] = BitArray(params[4 + i * 2: 4 + i * 2 + 2]).uint
network_params['batch_size'] = BATCH_SIZE[network_params['batch_size']]
network_params['seq_len'] = SEQ_LEN[network_params['seq_len']]
network_params['state_size'] = STATE_SIZE[network_params['state_size']]
network_params['dr'] = DR[network_params['dr']]
network_params['pkeep'] = PKEEP[network_params['pkeep']]
network_params['activation_f'] = ACTIVATION[network_params['activation_f']]
network_params['initializer'] = INIT[network_params['initializer']]
# timeseries_params
timeseries_params = {}
TIME_STEP_DAYS = [7, 14, 30, 60]
TIME_STEP_WEEKS = [4, 8, 12, 24]
TIME_STEP_MONTHS = [2, 3, 6, 9]
TIME_STEP = [TIME_STEP_DAYS, TIME_STEP_WEEKS, TIME_STEP_MONTHS]
step_name = ['time_series_step_days', 'time_series_step_weeks', 'time_series_step_months']
for index in range(NUM_TIME):
name = step_name[index]
step = TIME_STEP[index]
timeseries_params[name] = step[BitArray(params[20 + index * 2: 20 + index * 2 + 2]).uint]
DELAY = [7, 14, 30, 60, 90, 120, 150, 180]
delay_name_days = ['delay_google_days', 'delay_tweeter_days', 'delay_macro_days', 'delay_tweeter_re_days']
delay_name_weeks = ['delay_google_weeks', 'delay_tweeter_weeks', 'delay_macro_weeks', 'delay_tweeter_re_weeks']
delay_name_months = ['delay_google_months', 'delay_tweeter_months', 'delay_macro_months', 'delay_tweeter_re_months']
delay_name = [delay_name_days, delay_name_weeks, delay_name_months]
for type in range(NUM_DELAY_TYPE):
name_list = delay_name[type]
for index in range(NUM_DELAY):
name = name_list[index]
timeseries_params[name] = DELAY[BitArray(params[26 + index * 3: 26 + index * 3 + 3]).uint]
return network_params, timeseries_params
def main():
# XOR data set
data = np.array([[0, 0], [0, 1], [1, 0], [1, 1]], dtype=np.float32)
target = np.array([[0], [1], [1], [0]], dtype=np.float32)
# Show the 2D data
colors = np.array([
[1.0, 0.0, 0.0], # Red
[0.0, 0.0, 1.0]
]) # Blue
c = colors[np.squeeze(target == 0).astype(np.int)]
fig = plt.figure(figsize=(4, 4))
ax = fig.add_subplot(111)
ax.scatter(data[:, 0], data[:, 1], c=c, marker='x')
ax.set_title('XOR dataset (2D)')
ax.set_xlabel('First input')
ax.set_ylabel('Second input')
fig.tight_layout()
plt.show()
# Create neural network
model = Sequential()
model.add(
Dense(units=2,
activation='sigmoid',
input_shape=(2, ),
kernel_initializer=RandomUniform(minval=-0.01, maxval=0.01)))
model.add(Dense(units=1, activation='linear'))
print(model.summary())
# Define training parameters
model.compile(optimizer=SGD(lr=0.5, momentum=0.9), loss='mse')
# Perform training
model.fit(data,
target,
batch_size=len(data),
epochs=1000,
shuffle=True,
verbose=1)
# Save trained model to disk
model.save('xor.h5')
# Test model (loading from disk)
model = load_model('xor.h5')
targetPred = model.predict(data)
# Print the number of classification errors from the training data
nbErrors = np.sum(np.round(targetPred) != target)
accuracy = (len(data) - nbErrors) / len(data)
print('Classification accuracy: %0.3f' % (accuracy))
def __init__(self,
segments=2,
alpha_initializer=RandomUniform(0., 1.),
beta_initializer=RandomUniform(0., 1.),
alpha_regularizer=l2(1e-3),
beta_regularizer=l2(1e-3),
shared_axes=None,
**kwargs):
super(APLU, self).__init__(**kwargs)
self.segments = segments
self.alpha_initializer = alpha_initializer
self.beta_initializer = beta_initializer
self.alpha_regularizer = alpha_regularizer
self.beta_regularizer = beta_regularizer
if shared_axes is None:
self.shared_axes = None
elif not isinstance(shared_axes, (list, tuple)):
self.shared_axes = [shared_axes]
else:
self.shared_axes = list(shared_axes)
