def compute_transfweights(source_size,
target_size,
conditioned,
is_training=True,
batch_norm=True,
reuse=False):
last_conv = tf.get_collection('last_conv')[0]
batch_sz = last_conv.get_shape().as_list()[0]
indexing_tensor = compute_indexing(source_size, target_size)
indexing_tensor = tf.tile(indexing_tensor, [batch_sz, 1, 1, 1])
H, W = indexing_tensor.get_shape().as_list()[1:3]
if conditioned:
# compute image global features
with nets.base_arg_scope(is_training, batch_norm):
with tf.variable_scope('condition_net', reuse=reuse) as scope:
fc1 = slim.conv2d(last_conv, 64, [1, 1], scope='fc1')
tf.summary.histogram('fc1', fc1)
fc2 = slim.conv2d(fc1,
1, [1, 1],
scope='fc2',
activation_fn=None,
normalizer_fn=None)
tf.summary.histogram('fc2', fc2)
fc2 = tf.reshape(fc2, [batch_sz, 1, 1, -1])
fc2 = tf.tile(fc2, [1, H, W, 1])
fc2 = tf.reshape(fc2, [batch_sz, H, W, -1])
# concatenate with indexing tensor
net = tf.concat([indexing_tensor, fc2], 3)
else:
net = indexing_tensor
with nets.base_arg_scope(is_training, batch_norm, weight_decay=0.):
with tf.variable_scope('weight_net', reuse=reuse) as scope:
net = slim.conv2d(net, 128, [1, 1], scope='fc1')
tf.summary.histogram('fc1_wn', net)
net = slim.conv2d(net, 64, [1, 1], scope='fc2')
tf.summary.histogram('fc2_wn', net)
net = slim.conv2d(net,
1, [1, 1],
scope='fc3',
activation_fn=None,
normalizer_fn=None)
tf.summary.histogram('fc3_wn', net)
weights = tf.reshape(
net,
[-1, source_size[0] * source_size[1], target_size[0] * target_size[1]])
weights = tf.nn.softmax(weights, dim=1) # sum columns to 1
return weights
def pixelnet(inputs,
num_class,
is_training=True,
reuse=False,
batch_norm=True):
with nets.base_arg_scope(is_training, batch_norm):
net = pixelnet_convs(inputs, num_class, is_training, reuse)
with tf.variable_scope('aerial_mlp', reuse=reuse):
fc1 = slim.conv2d(net, 512, [1, 1], scope='fc1')
fc2 = slim.conv2d(fc1, 512, [1, 1], scope='fc2')
fc3 = slim.conv2d(fc2,
num_class, [1, 1],
scope='fc3',
activation_fn=None,
normalizer_fn=None)
tf.add_to_collection('aerial_mlp', fc1)
tf.add_to_collection('aerial_mlp', fc2)
tf.add_to_collection('aerial_mlp', fc3)
return fc3
def pixelnet_g(inputs,
num_class,
is_training=True,
reuse=False,
batch_norm=True):
with nets.base_arg_scope(is_training, batch_norm):
net = pixelnet_convs_g(inputs, num_class, is_training, reuse)
net = tf.image.resize_bilinear(net, size=[8, 40])
with tf.variable_scope('ground_mlp', reuse=reuse):
fc1 = slim.conv2d(net, 512, [1, 1], scope='fc1_g')
fc2 = slim.conv2d(fc1, 512, [1, 1], scope='fc2_g')
fc3 = slim.conv2d(fc2,
num_class, [1, 1],
scope='fc3_g',
activation_fn=None,
normalizer_fn=None)
tf.add_to_collection('ground_mlp', fc1)
tf.add_to_collection('ground_mlp', fc2)
tf.add_to_collection('ground_mlp', fc3)
return fc3