def build_FF_controller(self,
input_,
output_list_prev,
hidden_list_prev,
scope="content"):
"""Build Feedforward controller."""
with tf.variable_scope("controller_" + scope):
output_list = []
hidden_list = []
for layer_idx in range(self.ctl_layer_size):
if layer_idx == 0:
z = Linear(input_,
self.ctl_hidden_size,
stddev=self.init_range,
name="ff_%s" % layer_idx)
else:
z = Linear(output_list[-1],
self.ctl_hidden_size,
stddev=self.init_range,
name="ff_%s" % layer_idx)
out = tf.nn.tanh(z)
hidden_list.append(out)
output_list.append(out)
return output_list, hidden_list
def init_state(self, dummy_value=0.0):
dummy = tf.Variable(tf.constant([[dummy_value]], dtype=tf.float32))
M = [tf.zeros(shape) for shape in self.M_shapes]
R = tf.zeros([1, self.G_shape[0] * self.G_shape[1]])
val = tf.zeros([1, sum(self.read_sizes)])
# controller state
output_init_list = []
hidden_init_list = []
for idx in range(self.ctl_layer_size):
output_init_idx = Linear(dummy, self.ctl_hidden_size, squeeze=True, stddev=self.init_range, name='output_init_%s' % idx)
output_init_list.append(tf.tanh(output_init_idx))
hidden_init_idx = Linear(dummy, self.ctl_hidden_size, squeeze=True, stddev=self.init_range, name='hidden_init_%s' % idx)
hidden_init_list.append(tf.tanh(hidden_init_idx))
state = {
'M': M,
'R': R,
'val': val,
'ptr': tf.zeros([1, self.ptr_size]), #weight('ptr_init', [1, self.ptr_size], init='constant'),
'dptr': tf.zeros([1, self.ptr_size]),
'seq_length': 0,
'output': output_init_list,
'hidden': hidden_init_list,
}
self.__call__(state,0)
return state
def build_RNN_controller(self, input_, output_list_prev, hidden_list_prev, scope="content"):
"""Build RNN controller."""
with tf.variable_scope("controller_" + scope):
output_list = []
hidden_list = []
for layer_idx in range(self.ctl_layer_size):
o_prev = output_list_prev[layer_idx]
h_prev = hidden_list_prev[layer_idx]
if layer_idx == 0:
hid = tf.nn.tanh(linear([input_, o_prev],
output_size = self.ctl_hidden_size,
bias = True,
stddev=self.init_range,
scope = "f1_%s" % (layer_idx)))
else:
hid = tf.nn.tanh(linear([output_list[-1], o_prev],
output_size = self.ctl_hidden_size,
bias = True,
stddev=self.init_range,
scope="f1_%s" % (layer_idx)))
out = tf.nn.tanh(Linear(hid, self.ctl_hidden_size, stddev=self.init_range, name="f2_%s" % (layer_idx)))
hidden_list.append(out)
output_list.append(out)
return output_list, hidden_list
def MobileNet(x, is_training=True):
with tf.variable_scope('mobilenet'):
x = Conv3x3(x, 3, 32, 2, is_training, 'conv1') # 224 -> 112
x = Separable(x, 32, 64, 1, is_training, 'sepa1')
x = Separable(x, 64, 128, 2, is_training, 'sepa2') # 112 -> 56
x = Separable(x, 128, 128, 1, is_training, 'sepa3_1')
x = Separable(x, 128, 256, 2, is_training, 'sepa3_2') # 56 -> 28
x = Separable(x, 256, 256, 1, is_training, 'sepa4_1')
x = Separable(x, 256, 512, 2, is_training, 'sepa4_2') # 28 -> 14
for i in range(5):
x = Separable(x, 512, 512, 1, is_training,
'sepa5_{}'.format(i + 1))
x = Separable(x, 512, 1024, 2, is_training, 'sepa5_6') # 14 -> 7
x = Separable(x, 1024, 1024, 1, is_training, 'sepa6')
x = AvgPool7x7(x)
x = Linear(x, 1024, 1000)
return x