def qnet(observation_space, action_space, net_name, net_size):
num_actions = action_space.n
net_size = int(net_size)
net_name = net_name.lower()
state, feature, net = _atari_state_feature_net(observation_space, net_name)
# dueling or regular dqn/drqn
if 'dueling' in net_name:
value1 = net(net_size, activation='relu')(feature)
adv1 = net(net_size, activation='relu')(feature)
value2 = Dense(1)(value1)
adv2 = Dense(num_actions)(adv1)
mean_adv2 = Lambda(lambda x: K.mean(x, axis=1))(adv2)
ones = K.ones([1, num_actions])
lambda_exp = lambda x: K.dot(K.expand_dims(x, axis=1), -ones)
exp_mean_adv2 = Lambda(lambda_exp)(mean_adv2)
sum_adv = layers.add([exp_mean_adv2, adv2])
exp_value2 = Lambda(lambda x: K.dot(x, ones))(value2)
q_value = layers.add([exp_value2, sum_adv])
else:
hid = net(net_size, activation='relu')(feature)
q_value = Dense(num_actions)(hid)
# build model
return models.Model(inputs=state, outputs=q_value)
def spatial_attention(cost_volume):
feature = 4 * 9
k = 9
label = 9
dres0 = convbn_3d(cost_volume, feature / 2, 3, 1)
dres0 = Activation('relu')(dres0)
dres0 = convbn_3d(dres0, 1, 3, 1)
cost0 = Activation('relu')(dres0)
cost0 = Lambda(lambda x: K.permute_dimensions(K.squeeze(x, -1),
(0, 2, 3, 1)))(cost0)
cost1 = convbn(cost0, label // 2, (1, k), 1, 1)
cost1 = Activation('relu')(cost1)
cost1 = convbn(cost1, 1, (k, 1), 1, 1)
cost1 = Activation('relu')(cost1)
cost2 = convbn(cost0, label // 2, (k, 1), 1, 1)
cost2 = Activation('relu')(cost2)
cost2 = convbn(cost2, 1, (1, k), 1, 1)
cost2 = Activation('relu')(cost2)
cost = add([cost1, cost2])
cost = Activation('sigmoid')(cost)
cost = Lambda(lambda y: K.repeat_elements(K.expand_dims(y, 1), 9, 1))(cost)
cost = Lambda(lambda y: K.repeat_elements(y, feature, 4))(cost)
return multiply([cost, cost_volume])
def BasicBlock(input, planes, stride, downsample, dilation):
conv1 = convbn(input, planes, 3, stride, dilation)
conv1 = Activation('relu')(conv1)
conv2 = convbn(conv1, planes, 3, 1, dilation)
if downsample is not None:
input = downsample
conv2 = add([conv2, input])
return conv2
def basic(cost_volume):
feature = 2 * 75
dres0 = convbn_3d(cost_volume, feature, 3, 1)
dres0 = Activation('relu')(dres0)
dres0 = convbn_3d(dres0, feature, 3, 1)
cost0 = Activation('relu')(dres0)
dres1 = convbn_3d(cost0, feature, 3, 1)
dres1 = Activation('relu')(dres1)
dres1 = convbn_3d(dres1, feature, 3, 1)
cost0 = add([dres1, cost0])
dres4 = convbn_3d(cost0, feature, 3, 1)
dres4 = Activation('relu')(dres4)
dres4 = convbn_3d(dres4, feature, 3, 1)
cost0 = add([dres4, cost0])
classify = convbn_3d(cost0, feature, 3, 1)
classify = Activation('relu')(classify)
cost = Conv3D(1, 3, 1, 'same', data_format='channels_last',
use_bias=False)(classify)
return cost
def Conv_Block(inpt,
nb_filter,
kernel_size,
strides=(1, 1),
with_conv_shortcut=False):
x = _BN_ReLU_Conv2d(inpt,
nb_filter=nb_filter,
kernel_size=kernel_size,
strides=strides,
padding='same')
x = _BN_ReLU_Conv2d(x,
nb_filter=nb_filter,
kernel_size=kernel_size,
padding='same')
if with_conv_shortcut:
shortcut = _BN_ReLU_Conv2d(inpt,
nb_filter=nb_filter,
strides=strides,
kernel_size=kernel_size)
x = add([x, shortcut])
return x
else:
x = add([x, inpt])
return x
def _create_model(self,
middle_flow_repeats: int = 8,
dropout: float = 0.,
weight_decay: float = 0.,
**kwargs) -> None:
"""
Craete XCeption model.
:param middle_flow_repeats: number of middle flow block repeats
:param dropout: dropout rate of the extracted features
:param weight_decay: weight decay regularization
"""
images = tf.placeholder(tf.float32,
shape=[None] + self._dataset.shape,
name='images')
labels = tf.placeholder(tf.int64, shape=[None], name='labels')
regularizer = K.regularizers.l2(weight_decay)
with tf.variable_scope('model'):
net = Conv2D(32, (3, 3),
strides=(2, 2),
use_bias=False,
name='block1_conv1',
kernel_regularizer=regularizer)(images)
net = BatchNormalization(name='block1_conv1_bn')(
net, training=self.is_training)
net = Activation('relu', name='block1_conv1_act')(net)
net = Conv2D(64, (3, 3),
use_bias=False,
name='block1_conv2',
kernel_regularizer=regularizer)(net)
net = BatchNormalization(name='block1_conv2_bn')(
net, training=self.is_training)
net = Activation('relu', name='block1_conv2_act')(net)
residual = Conv2D(128, (1, 1),
strides=(2, 2),
padding='same',
use_bias=False,
kernel_regularizer=regularizer)(net)
residual = BatchNormalization()(residual,
training=self.is_training)
net = SeparableConv2D(128, (3, 3),
padding='same',
use_bias=False,
name='block2_sepconv1',
depthwise_regularizer=regularizer,
pointwise_regularizer=regularizer)(net)
net = BatchNormalization(name='block2_sepconv1_bn')(
net, training=self.is_training)
net = Activation('relu', name='block2_sepconv2_act')(net)
net = SeparableConv2D(128, (3, 3),
padding='same',
use_bias=False,
name='block2_sepconv2',
depthwise_regularizer=regularizer,
pointwise_regularizer=regularizer)(net)
net = BatchNormalization(name='block2_sepconv2_bn')(
net, training=self.is_training)
net = MaxPooling2D((3, 3),
strides=(2, 2),
padding='same',
name='block2_pool')(net)
net = layers.add([net, residual])
residual = Conv2D(256, (1, 1),
strides=(2, 2),
padding='same',
use_bias=False,
kernel_regularizer=regularizer)(net)
residual = BatchNormalization()(residual,
training=self.is_training)
net = Activation('relu', name='block3_sepconv1_act')(net)
net = SeparableConv2D(256, (3, 3),
padding='same',
use_bias=False,
name='block3_sepconv1',
depthwise_regularizer=regularizer,
pointwise_regularizer=regularizer)(net)
net = BatchNormalization(name='block3_sepconv1_bn')(
net, training=self.is_training)
net = Activation('relu', name='block3_sepconv2_act')(net)
net = SeparableConv2D(256, (3, 3),
padding='same',
use_bias=False,
name='block3_sepconv2',
depthwise_regularizer=regularizer,
pointwise_regularizer=regularizer)(net)
net = BatchNormalization(name='block3_sepconv2_bn')(
net, training=self.is_training)
net = MaxPooling2D((3, 3),
strides=(2, 2),
padding='same',
name='block3_pool')(net)
net = layers.add([net, residual])
residual = Conv2D(728, (1, 1),
strides=(2, 2),
padding='same',
use_bias=False,
kernel_regularizer=regularizer)(net)
residual = BatchNormalization()(residual,
training=self.is_training)
net = Activation('relu', name='block4_sepconv1_act')(net)
net = SeparableConv2D(728, (3, 3),
padding='same',
use_bias=False,
name='block4_sepconv1',
depthwise_regularizer=regularizer,
pointwise_regularizer=regularizer)(net)
net = BatchNormalization(name='block4_sepconv1_bn')(
net, training=self.is_training)
net = Activation('relu', name='block4_sepconv2_act')(net)
net = SeparableConv2D(728, (3, 3),
padding='same',
use_bias=False,
name='block4_sepconv2',
depthwise_regularizer=regularizer,
pointwise_regularizer=regularizer)(net)
net = BatchNormalization(name='block4_sepconv2_bn')(
net, training=self.is_training)
net = MaxPooling2D((3, 3),
strides=(2, 2),
padding='same',
name='block4_pool')(net)
net = layers.add([net, residual])
for i in range(middle_flow_repeats):
residual = net
prefix = 'block' + str(i + 5)
net = Activation('relu', name=prefix + '_sepconv1_act')(net)
net = SeparableConv2D(728, (3, 3),
padding='same',
use_bias=False,
name=prefix + '_sepconv1',
depthwise_regularizer=regularizer,
pointwise_regularizer=regularizer)(net)
net = BatchNormalization(name=prefix + '_sepconv1_bn')(
net, training=self.is_training)
net = Activation('relu', name=prefix + '_sepconv2_act')(net)
net = SeparableConv2D(728, (3, 3),
padding='same',
use_bias=False,
name=prefix + '_sepconv2',
depthwise_regularizer=regularizer,
pointwise_regularizer=regularizer)(net)
net = BatchNormalization(name=prefix + '_sepconv2_bn')(
net, training=self.is_training)
net = Activation('relu', name=prefix + '_sepconv3_act')(net)
net = SeparableConv2D(728, (3, 3),
padding='same',
use_bias=False,
name=prefix + '_sepconv3',
depthwise_regularizer=regularizer,
pointwise_regularizer=regularizer)(net)
net = BatchNormalization(name=prefix + '_sepconv3_bn')(
net, training=self.is_training)
net = layers.add([net, residual])
residual = Conv2D(1024, (1, 1),
strides=(2, 2),
padding='same',
use_bias=False,
kernel_regularizer=regularizer)(net)
residual = BatchNormalization()(residual,
training=self.is_training)
net = Activation('relu', name='block13_sepconv1_act')(net)
net = SeparableConv2D(728, (3, 3),
padding='same',
use_bias=False,
name='block13_sepconv1',
depthwise_regularizer=regularizer,
pointwise_regularizer=regularizer)(net)
net = BatchNormalization(name='block13_sepconv1_bn')(
net, training=self.is_training)
net = Activation('relu', name='block13_sepconv2_act')(net)
net = SeparableConv2D(1024, (3, 3),
padding='same',
use_bias=False,
name='block13_sepconv2',
depthwise_regularizer=regularizer,
pointwise_regularizer=regularizer)(net)
net = BatchNormalization(name='block13_sepconv2_bn')(
net, training=self.is_training)
net = MaxPooling2D((3, 3),
strides=(2, 2),
padding='same',
name='block13_pool')(net)
net = layers.add([net, residual])
net = SeparableConv2D(1536, (3, 3),
padding='same',
use_bias=False,
name='block14_sepconv1',
depthwise_regularizer=regularizer,
pointwise_regularizer=regularizer)(net)
net = BatchNormalization(name='block14_sepconv1_bn')(
net, training=self.is_training)
net = Activation('relu', name='block14_sepconv1_act')(net)
net = SeparableConv2D(2048, (3, 3),
padding='same',
use_bias=False,
name='block14_sepconv2',
depthwise_regularizer=regularizer,
pointwise_regularizer=regularizer)(net)
net = BatchNormalization(name='block14_sepconv2_bn')(
net, training=self.is_training)
net = Activation('relu', name='block14_sepconv2_act')(net)
logging.info('Output shape: %s', net.shape)
with tf.variable_scope('classifier'):
net = GlobalAveragePooling2D(name='avg_pool')(net)
if dropout > 0:
net = Dropout(dropout)(net, training=self.is_training)
logits = Dense(self._dataset.num_classes, activation=None)(net)
# outputs
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=labels,
logits=logits)
tf.identity(loss, name='loss')
tf.nn.softmax(logits, 1, name='predictions')
tf.reduce_mean(
tf.cast(tf.equal(tf.argmax(logits, 1), labels),
tf.float32,
name='accuracy'))