def rpn_head(featuremap, channel, num_anchors, tower_setup):
with tf.variable_scope('rpn'):
hidden = Conv(
'conv0', [featuremap],
channel,
tower_setup,
old_order=True,
bias=True,
W_initializer=tf.random_normal_initializer(stddev=0.01)).outputs[0]
label_logits = Conv(
'class', [hidden],
num_anchors,
tower_setup, (1, 1),
old_order=True,
bias=True,
activation="linear",
W_initializer=tf.random_normal_initializer(stddev=0.01)).outputs[0]
box_logits = Conv(
'box', [hidden],
4 * num_anchors,
tower_setup, (1, 1),
old_order=True,
bias=True,
activation="linear",
W_initializer=tf.random_normal_initializer(stddev=0.01)).outputs[0]
shp = tf.shape(box_logits)
box_logits = tf.reshape(
box_logits, tf.stack([shp[0], shp[1], shp[2], num_anchors, 4]))
return label_logits, box_logits
def resnet_shortcut(l, n_out, stride, tower_setup):
n_in = l.get_shape().as_list()[3]
if n_in != n_out: # change dimension when channel is not the same
if stride == 2:
l = l[:, :-1, :-1, :]
return Conv(name='convshortcut',
inputs=[l],
n_features=n_out,
tower_setup=tower_setup,
filter_size=(1, 1),
strides=(stride, stride),
activation='linear',
padding='VALID',
batch_norm=True,
old_order=True).outputs[0]
else:
return Conv(name='convshortcut',
inputs=[l],
n_features=n_out,
tower_setup=tower_setup,
filter_size=(1, 1),
strides=(stride, stride),
activation='linear',
batch_norm=True,
old_order=True).outputs[0]
else:
return l
def resnet_bottleneck(l, ch_out, stride, tower_setup):
l, shortcut = l, l
l = Conv(name='conv1', inputs=[l], n_features=ch_out, tower_setup=tower_setup, filter_size=(1, 1),
activation='relu', batch_norm=True, old_order=True).outputs[0]
if stride == 2:
l = tf.pad(l, [[0, 0], [0, 1], [0, 1], [0, 0]])
l = Conv(name='conv2', inputs=[l], n_features=ch_out, tower_setup=tower_setup, filter_size=(3, 3), strides=(2, 2),
activation='relu', padding='VALID', batch_norm=True, old_order=True).outputs[0]
else:
l = Conv(name='conv2', inputs=[l], n_features=ch_out, tower_setup=tower_setup, filter_size=(3, 3),
strides=(stride, stride), activation='relu', batch_norm=True, old_order=True).outputs[0]
l = Conv(name='conv3', inputs=[l], n_features=ch_out * 4, tower_setup=tower_setup, filter_size=(1, 1),
activation='linear', batch_norm=True, old_order=True).outputs[0]
return l + resnet_shortcut(shortcut, ch_out * 4, stride, tower_setup)
def maskrcnn_head(feature, num_class, tower_setup, class_agnostic_conv=False):
with tf.variable_scope('maskrcnn'):
# c2's MSRAFill is fan_out
l = ConvTranspose('deconv', [feature],
256,
tower_setup, (2, 2),
strides=(2, 2),
bias=True,
W_initializer=tf.variance_scaling_initializer(
scale=2.0,
mode='fan_out',
distribution='truncated_normal')).outputs[0]
if class_agnostic_conv:
num_output_channels = 1
else:
num_output_channels = num_class - 1
l = Conv('conv', [l],
num_output_channels,
tower_setup, (1, 1),
old_order=True,
bias=True,
activation="linear",
W_initializer=tf.variance_scaling_initializer(
scale=2.0,
mode='fan_out',
distribution='truncated_normal')).outputs[0]
return l
def pretrained_vgg16_2(image, num_blocks, tower_setup, freeze_c2=True):
assert len(num_blocks) == 3
l = tf.pad(image, [[0, 0], [2, 3], [2, 3], [0, 0]])
l = Conv(name='conv0', inputs=[l], n_features=64, tower_setup=tower_setup, filter_size=(7, 7),
activation='relu', padding='SAME', batch_norm=True, old_order=True).outputs[0]
l = Conv(name='conv1', inputs=[l], n_features=64, tower_setup=tower_setup, filter_size=(7, 7),
activation='relu', padding='SAME', batch_norm=True, old_order=True).outputs[0]
l = tf.pad(l, [[0, 0], [0, 1], [0, 1], [0, 0]])
l = max_pool(l, shape=[2, 2], strides=[2, 2], padding='VALID')
l = Conv(name='conv2', inputs=[l], n_features=128, tower_setup=tower_setup, filter_size=(7, 7),
activation='relu', padding='SAME', batch_norm=True, old_order=True).outputs[0]
l = Conv(name='conv3', inputs=[l], n_features=128, tower_setup=tower_setup, filter_size=(7, 7),
activation='relu', padding='SAME', batch_norm=True, old_order=True).outputs[0]
l = tf.pad(l, [[0, 0], [0, 1], [0, 1], [0, 0]])
l = max_pool(l, shape=[2, 2], strides=[2, 2], padding='VALID')
l = Conv(name='conv4', inputs=[l], n_features=256, tower_setup=tower_setup, filter_size=(7, 7),
activation='relu', padding='SAME', batch_norm=True, old_order=True).outputs[0]
l = Conv(name='conv5', inputs=[l], n_features=256, tower_setup=tower_setup, filter_size=(7, 7),
activation='relu', padding='SAME', batch_norm=True, old_order=True).outputs[0]
l = Conv(name='conv6', inputs=[l], n_features=256, tower_setup=tower_setup, filter_size=(7, 7),
activation='relu', padding='SAME', batch_norm=True, old_order=True).outputs[0]
# l = tf.pad(l, [[0, 0], [0, 1], [0, 1], [0, 0]])
# l = max_pool(l, shape=[2, 2], strides=[2, 2], padding='VALID')
# l = Conv(name='conv7', inputs=[l], n_features=512, tower_setup=tower_setup, filter_size=(7, 7),
# activation='relu', padding='SAME', batch_norm=True, old_order=True).outputs[0]
# l = Conv(name='conv8', inputs=[l], n_features=512, tower_setup=tower_setup, filter_size=(7, 7),
# activation='relu', padding='SAME', batch_norm=True, old_order=True).outputs[0]
# l = Conv(name='conv9', inputs=[l], n_features=512, tower_setup=tower_setup, filter_size=(7, 7),
# activation='relu', padding='SAME', batch_norm=True, old_order=True).outputs[0]
# l = tf.pad(l, [[0, 0], [0, 1], [0, 1], [0, 0]])
# l = max_pool(l, shape=[2, 2], strides=[2, 2], padding='VALID')
# l = Conv(name='conv10', inputs=[l], n_features=512, tower_setup=tower_setup, filter_size=(7, 7),
# activation='relu', padding='SAME', batch_norm=True, old_order=True).outputs[0]
# l = Conv(name='conv11', inputs=[l], n_features=512, tower_setup=tower_setup, filter_size=(7, 7),
# activation='relu', padding='SAME', batch_norm=True, old_order=True).outputs[0]
# l = Conv(name='conv12', inputs=[l], n_features=512, tower_setup=tower_setup, filter_size=(7, 7),
# activation='relu', padding='SAME', batch_norm=True, old_order=True).outputs[0]
l = tf.pad(l, [[0, 0], [0, 1], [0, 1], [0, 0]])
l = max_pool(l, shape=[2, 2], strides=[2, 2], padding='VALID')
c2 = resnet_group(l, 'group0', 64, num_blocks[0], 1, tower_setup)
if freeze_c2:
c2 = tf.stop_gradient(c2)
c3 = resnet_group(c2, 'group1', 128, num_blocks[1], 2, tower_setup)
#if tower_setup.is_training:
# tf.add_to_collection('checkpoints', c3)
c4 = resnet_group(c3, 'group2', 256, num_blocks[2], 2, tower_setup)
return [c2, c3, c4]
def maskrcnn_head(feature, num_class, tower_setup):
with tf.variable_scope('maskrcnn'):
# TODO both operations should have kernel_initializer=tf.variance_scaling_initializer(scale=2.0, mode='fan_out', distribution='normal')
# c2's MSRAFill is fan_out
l = ConvTranspose('deconv', [feature],
256,
tower_setup, (2, 2),
strides=(2, 2),
bias=True).outputs[0]
l = Conv('conv', [l],
num_class - 1,
tower_setup, (1, 1),
old_order=True,
bias=True,
activation="linear").outputs[0]
return l
def pretrained_xception_done(image, num_blocks, tower_setup, freeze_c2=True):
assert len(num_blocks) == 3
l = tf.pad(image, [[0, 0], [2, 3], [2, 3], [0, 0]])
l = Conv(name='conv0',
inputs=[l],
n_features=32,
tower_setup=tower_setup,
filter_size=(3, 3),
activation='relu',
padding='VALID',
batch_norm=True,
old_order=True).outputs[0]
l = Conv(name='conv1',
inputs=[l],
n_features=64,
tower_setup=tower_setup,
filter_size=(3, 3),
activation='relu',
padding='VALID',
batch_norm=True,
old_order=True).outputs[0]
l = Conv(name='conv2',
inputs=[l],
n_features=128,
tower_setup=tower_setup,
filter_size=(3, 3),
activation='relu',
padding='VALID',
batch_norm=True,
old_order=True).outputs[0]
l = Conv(name='conv3',
inputs=[l],
n_features=128,
tower_setup=tower_setup,
filter_size=(3, 3),
activation='relu',
padding='VALID',
batch_norm=True,
old_order=True).outputs[0]
l = max_pool(l, shape=[3, 3], strides=[2, 2], padding='SAME')
l = Conv(name='conv4',
inputs=[l],
n_features=256,
tower_setup=tower_setup,
filter_size=(3, 3),
activation='relu',
padding='VALID',
batch_norm=True,
old_order=True).outputs[0]
l = Conv(name='conv5',
inputs=[l],
n_features=256,
tower_setup=tower_setup,
filter_size=(3, 3),
activation='relu',
padding='VALID',
batch_norm=True,
old_order=True).outputs[0]
l = max_pool(l, shape=[3, 3], strides=[2, 2], padding='SAME')
l = Conv(name='conv6',
inputs=[l],
n_features=728,
tower_setup=tower_setup,
filter_size=(3, 3),
activation='relu',
padding='VALID',
batch_norm=True,
old_order=True).outputs[0]
l = Conv(name='conv7',
inputs=[l],
n_features=728,
tower_setup=tower_setup,
filter_size=(3, 3),
activation='relu',
padding='VALID',
batch_norm=True,
old_order=True).outputs[0]
l = max_pool(l, shape=[3, 3], strides=[2, 2], padding='SAME')
l = Conv(name='conv8',
inputs=[l],
n_features=728,
tower_setup=tower_setup,
filter_size=(3, 3),
activation='relu',
padding='VALID',
batch_norm=True,
old_order=True).outputs[0]
l = Conv(name='conv9',
inputs=[l],
n_features=728,
tower_setup=tower_setup,
filter_size=(3, 3),
activation='relu',
padding='VALID',
batch_norm=True,
old_order=True).outputs[0]
l = Conv(name='conv10',
inputs=[l],
n_features=728,
tower_setup=tower_setup,
filter_size=(3, 3),
activation='relu',
padding='VALID',
batch_norm=True,
old_order=True).outputs[0]
l = Conv(name='conv11',
inputs=[l],
n_features=728,
tower_setup=tower_setup,
filter_size=(3, 3),
activation='relu',
padding='VALID',
batch_norm=True,
old_order=True).outputs[0]
l = Conv(name='conv12',
inputs=[l],
n_features=1024,
tower_setup=tower_setup,
filter_size=(3, 3),
activation='relu',
padding='VALID',
batch_norm=True,
old_order=True).outputs[0]
l = max_pool(l, shape=[3, 3], strides=[2, 2], padding='SAME')
l = Conv(name='conv13',
inputs=[l],
n_features=1536,
tower_setup=tower_setup,
filter_size=(3, 3),
activation='relu',
padding='VALID',
batch_norm=True,
old_order=True).outputs[0]
l = Conv(name='conv14',
inputs=[l],
n_features=2048,
tower_setup=tower_setup,
filter_size=(3, 3),
activation='relu',
padding='VALID',
batch_norm=True,
old_order=True).outputs[0]
l = max_pool(l, shape=[3, 3], strides=[2, 2], padding='SAME')
l = Conv(name='conv15',
inputs=[l],
n_features=2048,
tower_setup=tower_setup,
filter_size=(1, 1),
activation='relu',
padding='VALID',
batch_norm=True,
old_order=True).outputs[0]
c2 = resnet_group(l, 'group0', 64, num_blocks[0], 1, tower_setup)
if freeze_c2:
c2 = tf.stop_gradient(c2)
c3 = resnet_group(c2, 'group1', 128, num_blocks[1], 2, tower_setup)
c4 = resnet_group(c3, 'group2', 256, num_blocks[2], 2, tower_setup)
return [c2, c3, c4]