def set_model(self, input_figs, z, batch_size, is_training = True, reuse = False):
assert(self.en_channels[0] == input_figs.get_shape().as_list()[3])
# reshape z
with tf.variable_scope(self.name_scopes[0], reuse = reuse):
h = linear('_r', z, get_dim(input_figs))
h = batch_norm('reshape', h, decay_rate= 0.99,
is_training = is_training)
h = tf.nn.relu(h)
height = input_figs.get_shape().as_list()[1]
width = input_figs.get_shape().as_list()[2]
h = tf.reshape(h, [-1, height, width, self.en_channels[0]])
h = tf.concat([h, input_figs], 3)
# convolution
encoded_list = []
# encode
with tf.variable_scope(self.name_scopes[1], reuse = reuse):
for i, out_dim in enumerate(self.en_channels[1:]):
h = conv(i, h, out_dim, 4, 4, 2)
if i == 0:
encoded_list.append(h)
h = lrelu(h)
else:
h = batch_norm(i, h, 0.99, is_training)
encoded_list.append(h)
h = lrelu(h)
# deconvolution
encoded_list.pop()
h = tf.nn.relu(h)
with tf.variable_scope(self.name_scopes[2], reuse = reuse):
for i, out_chan in enumerate(self.dec_channels[:-1]):
# get out shape
h_shape = h.get_shape().as_list()
out_width = 2 * h_shape[2]
out_height = 2 * h_shape[1]
out_shape = [batch_size, out_height, out_width, out_chan]
# deconvolution
deconved = deconv(i, h, out_shape, 4, 4, 2)
# batch normalization
h = batch_norm(i, deconved, 0.99, is_training)
if i <= 2:
h = tf.nn.dropout(h, 0.5)
h = tf.concat([h, encoded_list.pop()], 3)
# activation
h = tf.nn.relu(h)
height = 2 * h.get_shape().as_list()[1]
width = 2 * h.get_shape().as_list()[1]
out_shape = [batch_size, height, width, self.dec_channels[-1]]
h = deconv(i + 1, h, out_shape, 4, 4, 2)
return tf.nn.tanh(h)
def set_model(self, input_img, is_training = True, reuse = False):
assert(self.layer_channels[0] == input_img.get_shape().as_list()[-1])
h = input_img
# convolution
with tf.variable_scope(self.name_scopes[0], reuse = reuse):
for i, out_chan in enumerate(self.layer_channels[1:]):
h = conv(i, h, out_chan, 5, 5, 2)
h = lrelu(h)
# fully connect
h = flatten(h)
with tf.variable_scope(self.name_scopes[1], reuse = reuse):
h = linear('disc_fc', h, 1)
return h
def set_model(self, inputs, batch_size, is_training = True, reuse = False):
assert(self.layer_channels[0] == inputs.get_shape().as_list()[-1])
h = inputs
# convolution
with tf.variable_scope(self.name_scopes[0], reuse = reuse):
for i, out_chan in enumerate(self.layer_channels[1:]):
# convolution
conved = conv(i, h, out_chan, 5, 5, 2)
# batch normalization
bn_conved = batch_norm(i, conved, 0.99, is_training)
# activation
h = lrelu(bn_conved)
# fully connect
with tf.variable_scope(self.name_scopes[1], reuse = reuse):
encoded = linear('fc', flatten(h), self.out_dim)
return encoded
def build_graph(self):
def pad(board):
board = tf.pad(board, [[0, 0], [0, 0], [8, 8], [8, 8]],
constant_values=0.0)
edges = tf.constant(0, dtype=tf.float32, shape=[1, 1, 9, 9])
edges = tf.pad(edges, [[0, 0], [0, 0], [8, 8], [8, 8]],
constant_values=1.0)
board = tf.concat([board, edges], 1)
return board
# activation function
activation = tf.nn.selu
self.observation = tf.placeholder(tf.float32,
self.obs_space,
name='inputs')
board = pad(self.observation)
# Channel first to channel last as conv2d only support channel last
board = tf.transpose(board, [0, 2, 3, 1])
conv1 = U.conv(board,
name='conv1',
filters=16,
kernel_size=7,
strides=1,
use_bias=True,
activation=activation,
padding="valid",
summary=self.summary)
conv2 = U.conv(conv1,
name='conv2',
filters=16,
kernel_size=5,
strides=1,
use_bias=True,
activation=activation,
padding="valid",
summary=self.summary)
conv3 = U.conv(conv2,
name='conv3',
filters=44,
kernel_size=5,
strides=1,
use_bias=True,
activation=activation,
padding="valid",
summary=self.summary)
conv4 = U.conv(conv3,
name='conv4',
filters=16,
kernel_size=3,
strides=1,
use_bias=True,
activation=activation,
padding="valid",
summary=self.summary)
out = U.conv(conv4,
name='out',
filters=1,
kernel_size=1,
strides=1,
use_bias=True,
activation=None,
padding="valid",
summary=self.summary)
output = tf.layers.flatten(out)
#self.actions = tf.identity(output, 'outputs')
self.actions = output