def __call__(self, x, n_channels, n_codes):
non_linearity = tf.nn.relu
with tf.variable_scope(self.name, reuse=self.reuse):
conv1_a = conv2d(x, 'conv1_a', 16, 3, 2, 'SAME', True,
non_linearity, self.is_train)
conv1_b = conv2d(conv1_a, 'conv1_b', 16, 3, 1, 'SAME', True,
non_linearity, self.is_train)
conv2_a = conv2d(conv1_b, 'conv2_a', 32, 3, 2, 'SAME', True,
non_linearity, self.is_train)
conv2_b = conv2d(conv2_a, "conv2_b", 32, 3, 1, 'SAME', True,
non_linearity, self.is_train)
conv3 = conv2d(conv2_b, 'conv3', 1, 3, 2, 'SAME', True,
non_linearity, self.is_train)
fc1 = dense(tf.reshape(conv3, [-1, np.prod([8, 8, 1])]), 'fc1',
n_codes, False, None, self.is_train)
fc2 = dense(fc1, 'fc2', np.prod([8, 8, 1]), False, non_linearity,
self.is_train)
deconv4 = upconv2d(tf.reshape(fc2, [-1] + [8, 8, 1]), 2, 'deconv4',
32, 3, 1, 'SAME', True, non_linearity,
self.is_train)
conv4 = conv2d(deconv4, 'conv4', 32, 3, 1, 'SAME', True,
non_linearity, self.is_train)
deconv5 = upconv2d(conv4, 2, 'deconv5', 16, 3, 1, 'SAME', True,
non_linearity, self.is_train)
conv5 = conv2d(deconv5, 'conv5', 16, 3, 1, 'SAME', True,
non_linearity, self.is_train)
deconv6 = upconv2d(conv5, 2, 'deconv6', 16, 3, 1, 'SAME', True,
non_linearity, self.is_train)
conv_out = conv2d(deconv6, 'conv_out', n_channels, 3, 1, 'SAME',
False, None, self.is_train)
if self.reuse is None:
self.var_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope=self.name)
self.saver = tf.train.Saver(var_list=self.var_list,
max_to_keep=100)
self.reuse = True
return fc1, conv_out
def __call__(self, x):
non_linearity = tf.nn.elu
with tf.variable_scope(self.name, reuse=self.reuse):
conv1_a = conv2d(x, 'conv1_a', 16, 3, 1, 'SAME', True,
non_linearity, self.is_train)
conv1_b = conv2d(conv1_a, 'conv1_b', 16, 3, 1, 'SAME', True, None,
self.is_train)
act1 = non_linearity(conv1_b)
pool1 = tf.nn.avg_pool(act1, [1, 2, 2, 1], [1, 2, 2, 1], 'SAME')
conv2_a = conv2d(pool1, 'conv2_a', 32, 3, 1, 'SAME', True,
non_linearity, self.is_train)
conv2_b = conv2d(conv2_a, 'conv2_b', 32, 3, 1, 'SAME', True, None,
self.is_train)
act2 = non_linearity(conv2_b)
pool2 = tf.nn.avg_pool(act2, [1, 2, 2, 1], [1, 2, 2, 1], 'SAME')
conv3_a = conv2d(pool2, 'conv3_a', 64, 3, 1, 'SAME', True,
non_linearity, self.is_train)
conv3_b = conv2d(conv3_a, 'conv3_b', 64, 3, 1, 'SAME', True, None,
self.is_train)
act3 = non_linearity(conv3_b)
pool3 = tf.nn.avg_pool(act3, [1, 2, 2, 1], [1, 2, 2, 1], 'SAME')
conv4_a = conv2d(pool3, 'conv4_a', 128, 3, 1, 'SAME', True,
non_linearity, self.is_train)
conv4_b = conv2d(conv4_a, 'conv4_b', 128, 3, 1, 'SAME', True, None,
self.is_train)
act4 = non_linearity(conv4_b)
drop4 = tf.layers.dropout(act4,
rate=0.5,
training=self.is_train,
name='drop4')
deconv4 = upconv2d(drop4, 2, 'deconv4', 64, 3, 1, 'SAME', True,
None, self.is_train)
concat5 = non_linearity(tf.add(deconv4, conv3_b))
conv5_a = conv2d(concat5, 'conv5_a', 64, 3, 1, 'SAME', True,
non_linearity, self.is_train)
conv5_b = conv2d(conv5_a, 'conv5_b', 64, 3, 1, 'SAME', True,
non_linearity, self.is_train)
drop5 = tf.layers.dropout(conv5_b,
rate=0.5,
training=self.is_train,
name='drop5')
deconv5 = upconv2d(drop5, 2, 'deconv5', 32, 3, 1, 'SAME', True,
None, self.is_train)
concat6 = non_linearity(tf.add(deconv5, conv2_b))
conv6_a = conv2d(concat6, 'conv6_a', 32, 3, 1, 'SAME', True,
non_linearity, self.is_train)
conv6_b = conv2d(conv6_a, 'conv6_b', 32, 3, 1, 'SAME', True,
non_linearity, self.is_train)
drop6 = tf.layers.dropout(conv6_b,
rate=0.5,
training=self.is_train,
name='drop6')
deconv6 = upconv2d(drop6, 2, 'deconv6', 16, 3, 1, 'SAME', True,
None, self.is_train)
concat7 = non_linearity(tf.add(deconv6, conv1_b))
conv7_a = conv2d(concat7, 'conv7_a', 16, 3, 1, 'SAME', True,
non_linearity, self.is_train)
conv7_b = conv2d(conv7_a, 'conv7_b', 16, 3, 1, 'SAME', True,
non_linearity, self.is_train)
conv_out = conv2d(conv7_b, 'conv_out', 2, 3, 1, 'SAME', True, None,
self.is_train)
if self.reuse is None:
self.var_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope=self.name)
self.saver = tf.train.Saver(var_list=self.var_list,
max_to_keep=100)
self.reuse = True
return conv_out