def build_single_net(x, is_training, FLAGS):
with tf.variable_scope('pre'):
pre = layers.conv(x, num_outputs=32, kernel_size = [3, 3], scope='conv', b_norm=True, is_training=is_training,
weight_decay=FLAGS.weight_decay)
# pre = layers.max_pool2d(pre, [2, 2], padding='SAME', scope='pool')
h = pre
# for i in range(1, n + 1):
h = block(h, 32, FLAGS.weight_decay, '32_block{}'.format(1), is_training)
h = tf.nn.max_pool(h, [1, 2, 2, 1], [1, 1, 1, 1], 'SAME')
# h = block(h, 64, FLAGS.weight_decay, '64_block1', is_training, True)
h = layers.conv(h, num_outputs = 64, kernel_size=[3, 3], strides=[1, 1, 1, 1],
scope='conv1', b_norm=True, is_training=is_training, weight_decay=FLAGS.weight_decay)
# for i in range(2, n + 1):
h = block(h, 64, FLAGS.weight_decay, '64_block{}'.format(1), is_training)
h = tf.nn.max_pool(h, [1, 2, 2, 1], [1, 1, 1, 1], 'SAME')
h = layers.conv(h, num_outputs = 128, kernel_size=[3, 3], strides=[1, 1, 1, 1],
scope='conv2', b_norm=True, is_training=is_training, weight_decay=FLAGS.weight_decay)
# for i in range(2, n + 1):
h = block(h, 128, FLAGS.weight_decay, '128_block{}'.format(1), is_training)
shape = h.get_shape().as_list()
h = tf.contrib.layers.avg_pool2d(h, [shape[1], shape[2]], scope='global_pool')
return h
def build_single_net(x, is_training, FLAGS, nn=10):
n = FLAGS.blocks
# shape = x.get_shape().as_list()
with tf.variable_scope('pre'):
pre = layers.conv(x, num_outputs=24, kernel_size=[3, 3], scope='conv', b_norm=True, is_training=is_training,
weight_decay=FLAGS.weight_decay)
# pre = layers.max_pool2d(pre, [2, 2], padding='SAME', scope='pool')
h = pre
for i in range(1, n + 1):
h = block(h, 24, FLAGS.weight_decay, '24_block{}'.format(i), is_training)
h = block(h, 48, FLAGS.weight_decay, '46_block1', is_training, True)
for i in range(2, n + 1):
h = block(h, 48, FLAGS.weight_decay, '48_block{}'.format(i), is_training)
h = block(h, 96, FLAGS.weight_decay, '96_block1', is_training, True)
for i in range(2, n + 1):
h = block(h, 96, FLAGS.weight_decay, '96_block{}'.format(i), is_training)
shape = h.get_shape().as_list()
h = tf.contrib.layers.avg_pool2d(h, [shape[1], shape[2]], scope='global_pool')
res = layers.conv(h, num_outputs=nn, kernel_size=[1, 1], scope='fc1', padding='VALID',
b_norm=True, is_training=is_training, weight_decay=FLAGS.weight_decay, activation_fn=None)
return res
def build_net(x, is_training, FLAGS):
n = FLAGS.blocks
# shape = x.get_shape().as_list()
with tf.variable_scope('pre'):
pre = layers.conv(x,
num_outputs=16,
kernel_size=[3, 3],
scope='conv1',
b_norm=True,
is_training=is_training,
weight_decay=FLAGS.weight_decay)
pre = layers.conv(pre,
num_outputs=32,
kernel_size=[3, 3],
scope='conv2',
b_norm=True,
is_training=is_training,
weight_decay=FLAGS.weight_decay)
# pre = layers.max_pool2d(pre, [2, 2], padding='SAME', scope='pool')
h = pre
for i in range(1, n + 1):
h = block(h, 32, FLAGS.weight_decay, '32_block{}'.format(i),
is_training)
h = block(h, 64, FLAGS.weight_decay, '64_block1', is_training, True)
for i in range(2, n + 1):
h = block(h, 64, FLAGS.weight_decay, '64_block{}'.format(i),
is_training)
h = block(h, 128, FLAGS.weight_decay, '128_block1', is_training, True)
for i in range(2, n + 1):
h = block(h, 128, FLAGS.weight_decay, '128_block{}'.format(i),
is_training)
shape = h.get_shape().as_list()
h = tf.contrib.layers.avg_pool2d(h, [shape[1], shape[2]],
scope='global_pool')
h = layers.conv(h,
num_outputs=64,
kernel_size=[1, 1],
scope='fc1',
padding='VALID',
b_norm=True,
is_training=is_training,
weight_decay=FLAGS.weight_decay,
activation_fn=None)
h = layers.conv(h,
num_outputs=FLAGS.num_classes,
kernel_size=[1, 1],
scope='fc2',
padding='VALID',
b_norm=True,
is_training=is_training,
weight_decay=FLAGS.weight_decay,
activation_fn=None)
return tf.reshape(h, [-1, FLAGS.num_classes])
def build_net(x, is_training, FLAGS):
n = FLAGS.blocks
# shape = x.get_shape().as_list()
with tf.variable_scope('pre'):
pre = layers.conv(x,
num_outputs=64,
kernel_size=[7, 7],
scope='conv',
b_norm=True,
is_training=is_training,
weight_decay=FLAGS.weight_decay)
pre = tf.contrib.layers.max_pool2d(pre, [3, 3],
stride=2,
padding='SAME',
scope='pool') #32
h = pre
for i in range(1, 4):
h = block(h, 64, FLAGS.weight_decay, '64_block{}'.format(i),
is_training)
h = block(h, 128, FLAGS.weight_decay, '128_block_s2', is_training,
True) #16
for i in range(1, 4):
h = block(h, 128, FLAGS.weight_decay, '128_block{}'.format(i),
is_training)
h = block(h, 256, FLAGS.weight_decay, '256_block_s2', is_training, True)
for i in range(1, 6):
h = block(h, 256, FLAGS.weight_decay, '256_block{}'.format(i),
is_training)
h = block(h, 512, FLAGS.weight_decay, '512_block_s2', is_training, True)
for i in range(1, 3):
h = block(h, 512, FLAGS.weight_decay, '512_block{}'.format(i),
is_training)
shape = h.get_shape().as_list()
h = tf.contrib.layers.avg_pool2d(h, [shape[1], shape[2]],
scope='global_pool')
shape = h.get_shape().as_list()
h = layers.conv(h,
num_outputs=FLAGS.num_classes,
kernel_size=[shape[1], shape[2]],
scope='fc1',
padding='VALID',
b_norm=True,
is_training=is_training,
weight_decay=FLAGS.weight_decay,
activation_fn=None)
return tf.reshape(h, [-1, FLAGS.num_classes])
def build_net(x, is_training, FLAGS):
n = FLAGS.blocks
with tf.variable_scope('pre'):
pre = layers.conv(x, num_outputs=16, kernel_size = [3, 3], scope='conv', b_norm=True, is_training=is_training,
weight_decay=FLAGS.weight_decay)
h1 =
def build_net(x, is_training, FLAGS):
num_branches = FLAGS.num_branches
y = []
for i in range(1, num_branches + 1):
with tf.variable_scope('branch%d' % i):
tmp = build_single_net(x, is_training, FLAGS)
if FLAGS.p_relu:
tmp = layers.p_relu(tmp)
y.append(tmp)
with tf.variable_scope('branch%d' % (num_branches + 1)):
y4 = build_single_net_w(x, is_training, FLAGS)
con = tf.concat(y, axis=3) #192
con = tf.reshape(con, [-1, 1, num_branches * 224])
w = layers.conv(y4,
num_outputs=FLAGS.num_classes * num_branches * 224,
kernel_size=[1, 1],
scope='fc2',
padding='VALID',
b_norm=True,
is_training=is_training,
weight_decay=FLAGS.weight_decay,
activation_fn=None)
w = tf.reshape(w, [-1, num_branches * 224, FLAGS.num_classes])
res = tf.matmul(con, w)
res = layers.batch_norm(res, is_training=is_training, scope='bn')
return tf.reshape(res, [-1, FLAGS.num_classes])
def build_net(x, is_training, FLAGS):
n = FLAGS.blocks
# shape = x.get_shape().as_list()
with tf.variable_scope('pre'):
pre = layers.conv(x, num_outputs=64, kernel_size = [3, 3], scope='conv', b_norm=True, is_training=is_training,
weight_decay=FLAGS.weight_decay, activation_fn=tf.nn.elu)
# pre = layers.max_pool2d(pre, [2, 2], padding='SAME', scope='pool')
h = pre
shape = h.get_shape().as_list()
pool0 = tf.contrib.layers.avg_pool2d(h, [shape[1], shape[2]], scope='global_pool0')
fc0 = layers.conv(pool0, num_outputs=FLAGS.num_classes, kernel_size=[1, 1], scope='fc0', padding='VALID',
b_norm=True, is_training=is_training, weight_decay=FLAGS.weight_decay, activation_fn=None)
for i in range(1, 4):
h = block(h, 64, FLAGS.weight_decay, '64_block{}'.format(i), is_training)
shape = h.get_shape().as_list()
pool1 = tf.contrib.layers.avg_pool2d(h, [shape[1], shape[2]], scope='global_pool1')
fc1 = layers.conv(pool1, num_outputs=FLAGS.num_classes, kernel_size=[1, 1], scope='fc1', padding='VALID',
b_norm=True, is_training=is_training, weight_decay=FLAGS.weight_decay, activation_fn=None)
h = block(h, 128, FLAGS.weight_decay, '128_block1', is_training, True)
for i in range(2, 4):
h = block(h, 128, FLAGS.weight_decay, '128_block{}'.format(i), is_training)
shape = h.get_shape().as_list()
pool2 = tf.contrib.layers.avg_pool2d(h, [shape[1], shape[2]], scope='global_pool2')
fc2 = layers.conv(pool2, num_outputs=FLAGS.num_classes, kernel_size=[1, 1], scope='fc2', padding='VALID',
b_norm=True, is_training=is_training, weight_decay=FLAGS.weight_decay, activation_fn=None)
h = block(h, 256, FLAGS.weight_decay, '256_block1', is_training, True)
for i in range(2, 6):
h = block(h, 256, FLAGS.weight_decay, '256_block{}'.format(i), is_training)
shape = h.get_shape().as_list()
pool3 = tf.contrib.layers.avg_pool2d(h, [shape[1], shape[2]], scope='global_pool3')
fc3 = layers.conv(pool3, num_outputs=FLAGS.num_classes, kernel_size=[1, 1], scope='fc3', padding='VALID',
b_norm=True, is_training=is_training, weight_decay=FLAGS.weight_decay, activation_fn=None)
h = fc0 + fc1 + fc2 + fc3
return tf.reshape(h, [-1, FLAGS.num_classes])
def block(inputs, num_outputs, weight_decay, scope, is_training, down_sample = False):
with tf.variable_scope(scope):
num_inputs = inputs.get_shape().as_list()[3]
res = layers.conv(inputs, num_outputs = num_outputs, kernel_size=[3, 3], activation_fn=tf.nn.elu,
strides=[1, 2, 2, 1] if down_sample else [1, 1, 1, 1],
scope='conv1', b_norm=True, is_training=is_training, weight_decay=weight_decay)
res = layers.conv(res, num_outputs=num_outputs, kernel_size=[3, 3], activation_fn=None,
scope='conv2', b_norm=True, is_training=is_training, weight_decay=weight_decay)
if num_inputs != num_outputs:
inputs = layers.conv(inputs, num_outputs=num_outputs, kernel_size=[1, 1], activation_fn=None,
scope='short_cut', strides=[1, 2, 2,1 ], b_norm=True, is_training=is_training,
weight_decay=weight_decay)
res = tf.nn.elu(res + inputs)
return res
