def _shadowdata_generator_model(netinput, is_training=True):
with slim.arg_scope(
[slim.conv2d, slim.conv2d_transpose, slim.convolution1d],
# weights_initializer=initializers.variance_scaling(scale=2.0),
weights_initializer=initializers.zeros(),
# weights_regularizer=slim.l1_l2_regularizer(),
# normalizer_fn=slim.batch_norm,
# normalizer_params={'is_training': is_training, 'decay': 0.95},
# normalizer_fn=slim.instance_norm,
# normalizer_params={'center': True, 'scale': True, 'epsilon': 0.001},
activation_fn=(lambda inp: slim.nn.leaky_relu(inp, alpha=0.1)),
trainable=is_training,
data_format="NHWC"):
num_filters = 1
band_size = netinput.get_shape()[3].value
kernel_size = band_size
net0 = tf.expand_dims(tf.squeeze(netinput, axis=[1, 2]), axis=2)
net1 = slim.convolution1d(net0,
num_filters,
kernel_size,
padding='SAME')
net1 = net1 + net0
net2 = slim.convolution1d(net1,
num_filters,
kernel_size // 2,
padding='SAME')
net2 = net2 + net1 + net0
net3 = slim.convolution1d(net2,
num_filters,
kernel_size // 4,
padding='SAME')
net3 = net3 + net2 + net1
net4 = slim.convolution1d(net3,
num_filters,
kernel_size // 8,
padding='SAME')
net4 = net4 + net3 + net2
net5 = slim.convolution1d(net4,
num_filters,
kernel_size // 4,
padding='SAME')
net5 = net5 + net4 + net3
net6 = slim.convolution1d(net5,
num_filters,
kernel_size // 2,
padding='SAME')
net6 = net6 + net5 + net4
net7 = slim.convolution1d(net6,
num_filters,
kernel_size,
padding='SAME',
normalizer_fn=None,
normalizer_params=None,
weights_regularizer=None,
activation_fn=None)
flatten = slim.flatten(net7)
# net9 = slim.fully_connected(flatten, band_size, activation_fn=None)
return tf.expand_dims(tf.expand_dims(flatten, axis=1), axis=1)
# experiment with small datasets
trainX = trainX[:1000]
trainY = trainY[:1000]
for neuron_count in neurons:
tf.reset_default_graph()
with tf.Session() as sess:
# Create the model
lr = tf.placeholder(tf.float32, [])
x = tf.placeholder(tf.float32, [None, NUM_FEATURES])
y_ = tf.placeholder(tf.float32, [None, 1])
# Build the graph for the deep net
hidden_layer = layers.dense(x, neuron_count, activation=tf.nn.relu, kernel_initializer=init.orthogonal(np.sqrt(2)),
bias_initializer=init.zeros(), kernel_regularizer=l2_regularizer(1e-3))
output_layer = layers.dense(hidden_layer, 1, kernel_initializer=init.orthogonal(), bias_initializer=init.zeros(),
kernel_regularizer=l2_regularizer(1e-3))
loss = tf.reduce_mean(tf.square(y_ - output_layer)
) + get_regularization_loss()
# Create the gradient descent optimizer with the given learning rate.
optimizer = tf.train.GradientDescentOptimizer(lr)
train_op = optimizer.minimize(loss)
error = tf.reduce_mean(tf.square(y_ - output_layer))
fold_errors = []
for o in range(NUM_FOLDS):
sess.run(tf.global_variables_initializer())
trainX, trainY = X_data[m:], Y_data[m:]
# experiment with small datasets
trainX = trainX[:1000]
trainY = trainY[:1000]
# Create the model
x = tf.placeholder(tf.float32, [None, NUM_FEATURES])
y_ = tf.placeholder(tf.float32, [None, 1])
# Build the graph for the deep net
hidden_layer = layers.dense(x,
100,
activation=tf.nn.relu,
kernel_initializer=init.orthogonal(np.sqrt(2)),
bias_initializer=init.zeros(),
kernel_regularizer=l2_regularizer(1e-3))
output_layer = layers.dense(hidden_layer,
1,
kernel_initializer=init.orthogonal(),
bias_initializer=init.zeros(),
kernel_regularizer=l2_regularizer(1e-3))
loss = tf.reduce_mean(tf.square(y_ - output_layer)) + get_regularization_loss()
# Create the gradient descent optimizer with the given learning rate.
optimizer = tf.train.GradientDescentOptimizer(learning_rate)
train_op = optimizer.minimize(loss)
error = tf.reduce_mean(tf.square(y_ - output_layer))
with tf.Session() as sess:
def __init__(self, name, session, observation_space, action_space):
self.name = name
assert isinstance(action_space, spaces.Box)
assert len(action_space.shape) == 1
assert (np.abs(action_space.low) == action_space.high
).all() # we assume symmetric actions.
self.session = session
with tf.variable_scope(self.name):
self.observations_input = tf.placeholder(
observation_space.dtype,
shape=tuple([None] + list(observation_space.shape)),
name="observations_input")
self.actions_input = tf.placeholder(
action_space.dtype,
shape=tuple([None] + list(action_space.shape)),
name="actions_input")
self.actions_size = action_space.shape[0]
self.combined_input = tf.concat(
[self.observations_input, self.actions_input],
axis=1,
name="combined_input")
with tf.variable_scope("actor"):
self.actor01 = layers.dense(
inputs=self.observations_input,
units=64,
activation=tf.tanh,
name="layer_one",
bias_initializer=initializers.zeros(),
kernel_initializer=initializers.orthogonal(
gain=np.sqrt(2)))
self.actor02 = layers.dense(
inputs=self.actor01,
units=64,
activation=tf.tanh,
name="layer_two",
bias_initializer=initializers.zeros(),
kernel_initializer=initializers.orthogonal(
gain=np.sqrt(2)))
self.actor_head = layers.dense(
inputs=self.actor02,
units=self.actions_input.shape[1].value,
activation=tf.tanh,
name="head",
bias_initializer=initializers.zeros(),
kernel_initializer=initializers.orthogonal(gain=0.01))
self.actor_head_rescaled = self.actor_head * action_space.high
self.model_computed_combined = tf.concat(
[self.observations_input, self.actor_head_rescaled],
axis=1,
name="model_computed_combined")
with tf.variable_scope("critic"):
self.critic01 = layers.dense(
inputs=self.combined_input,
units=64,
activation=tf.tanh,
name="layer_one",
bias_initializer=initializers.zeros(),
kernel_initializer=initializers.orthogonal(
gain=np.sqrt(2)))
self.critic02 = layers.dense(
inputs=self.critic01,
units=64,
activation=tf.tanh,
name="layer_two",
bias_initializer=initializers.zeros(),
kernel_initializer=initializers.orthogonal(
gain=np.sqrt(2)))
self.action_value_head = layers.dense(
inputs=self.critic02,
units=1,
activation=None,
name="head",
bias_initializer=initializers.zeros(),
kernel_initializer=initializers.random_uniform(
-3.0e-3, 3.0e-3))
self.model_critic01 = layers.dense(
inputs=self.model_computed_combined,
units=64,
activation=tf.tanh,
name="layer_one",
bias_initializer=initializers.zeros(),
kernel_initializer=initializers.orthogonal(
gain=np.sqrt(2)),
reuse=
True # Use the same weights for 'critic' and for 'model critic'
)
self.model_critic02 = layers.dense(
inputs=self.model_critic01,
units=64,
activation=tf.tanh,
name="layer_two",
bias_initializer=initializers.zeros(),
kernel_initializer=initializers.orthogonal(
gain=np.sqrt(2)),
reuse=
True # Use the same weights for 'critic' and for 'model critic'
)
self.state_value_head = layers.dense(
inputs=self.model_critic02,
units=1,
activation=None,
name="head",
bias_initializer=initializers.zeros(),
kernel_initializer=initializers.random_uniform(
-3.0e-3, 3.0e-3),
reuse=
True # Use the same weights for 'critic' and for 'model critic'
)
# experiment with small datasets
trainX = trainX[:1000]
trainX = (trainX - np.mean(trainX, axis=0)) / np.std(trainX, axis=0)
trainY = trainY[:1000]
n = trainX.shape[0]
# Create the model
x = tf.placeholder(tf.float32, [None, NUM_FEATURES])
y_ = tf.placeholder(tf.float32, [None, NUM_CLASSES])
# Build the graph for the deep nets
hidden_layer = layers.dense(x, 10, activation=tf.nn.selu, kernel_initializer=init.orthogonal(),
bias_initializer=init.zeros(),
kernel_regularizer=l2_regularizer(1e-6))
output_layer = layers.dense(x, NUM_CLASSES, kernel_initializer=init.orthogonal(), bias_initializer=init.zeros(),
kernel_regularizer=l2_regularizer(1e-6))
cross_entropy = tf.nn.softmax_cross_entropy_with_logits_v2(labels=y_, logits=output_layer)
loss = tf.reduce_mean(cross_entropy) + get_regularization_loss()
# Create the gradient descent optimizer with the given learning rate.
optimizer = tf.train.GradientDescentOptimizer(learning_rate)
train_op = optimizer.minimize(loss)
correct_prediction = tf.cast(tf.equal(tf.argmax(output_layer, 1), tf.argmax(y_, 1)), tf.float32)
accuracy = tf.reduce_mean(correct_prediction)
with tf.Session() as sess:
# experiment with small datasets
trainX = trainX[:1000]
trainY = trainY[:1000]
n = trainX.shape[0]
# Create the model
x = tf.placeholder(tf.float32, [None, NUM_FEATURES])
y_ = tf.placeholder(tf.float32, [None, NUM_CLASSES])
# Build the graph for the deep net
hidden_layer = layers.dense(x, 10, activation=tf.nn.relu, kernel_initializer=init.orthogonal(
np.sqrt(2)), bias_initializer=init.zeros(), kernel_regularizer=l2_regularizer(1e-6))
output_layer = layers.dense(hidden_layer, 6, kernel_initializer=init.orthogonal(
), bias_initializer=init.zeros(), kernel_regularizer=l2_regularizer(1e-6))
cross_entropy = tf.nn.softmax_cross_entropy_with_logits_v2(
labels=y_, logits=output_layer)
loss = tf.reduce_mean(cross_entropy) + get_regularization_loss()
# Create the gradient descent optimizer with the given learning rate.
optimizer = tf.train.GradientDescentOptimizer(learning_rate)
train_op = optimizer.minimize(loss)
correct_prediction = tf.cast(
tf.equal(tf.argmax(output_layer, 1), tf.argmax(y_, 1)), tf.float32)
accuracy = tf.reduce_mean(correct_prediction)
def build(self, input_shape: Union[list, tuple]):
super().build(input_shape)
self.__create_weights__(init.zeros()(self.__num_outputs__),
'log_std_params')