def add_backbone_model(self, cnn):
# --------------------------------------------------------------------------
# Resnet-34 backbone model -- modified for SSD
# --------------------------------------------------------------------------
# Input 300x300, output 150x150
cnn.conv(64, 7, 7, 2, 2, mode='SAME_RESNET', use_batch_norm=True)
cnn.mpool(3, 3, 2, 2, mode='SAME')
resnet34_layers = [3, 4, 6, 3]
version = 'v1'
# ResNet-34 block group 1
# Input 150x150, output 75x75
for i in range(resnet34_layers[0]):
# Last argument forces residual_block to use projection shortcut, even
# though the numbers of input and output channels are equal
resnet_model.residual_block(cnn, 64, 1, version)
# ResNet-34 block group 2
# Input 75x75, output 38x38
for i in range(resnet34_layers[1]):
stride = 2 if i == 0 else 1
resnet_model.residual_block(cnn, 128, stride, version, i == 0)
# ResNet-34 block group 3
# This block group is modified: first layer uses stride=1 so that the image
# size does not change in group of layers
# Input 38x38, output 38x38
for i in range(resnet34_layers[2]):
# The following line is intentionally commented out to differentiate from
# the original ResNet-34 model
# stride = 2 if i == 0 else 1
resnet_model.residual_block(cnn, 256, stride, version, i == 0)
def add_inference(self, cnn):
# TODO(haoyuzhang): check batch norm params for resnet34 in reference model?
cnn.use_batch_norm = True
cnn.batch_norm_config = {'decay': 0.9, 'epsilon': 1e-5, 'scale': True}
# --------------------------------------------------------------------------
# Resnet-34 backbone model -- modified for SSD
# --------------------------------------------------------------------------
# Input 300x300, output 150x150
cnn.conv(64, 7, 7, 2, 2, mode='SAME_RESNET', use_batch_norm=True)
cnn.mpool(3, 3, 2, 2, mode='SAME')
resnet34_layers = [3, 4, 6, 3]
version = 'v1'
# ResNet-34 block group 1
# Input 150x150, output 75x75
for i in range(resnet34_layers[0]):
# Last argument forces residual_block to use projection shortcut, even
# though the numbers of input and output channels are equal
resnet_model.residual_block(cnn, 64, 1, version, i == 0)
# ResNet-34 block group 2
# Input 75x75, output 38x38
for i in range(resnet34_layers[1]):
stride = 2 if i == 0 else 1
resnet_model.residual_block(cnn, 128, stride, version, i == 0)
# ResNet-34 block group 3
# This block group is modified: first layer uses stride=1 so that the image
# size does not change in group of layers
# Input 38x38, output 38x38
for i in range(resnet34_layers[2]):
# The following line is intentionally commented out to differentiate from
# the original ResNet-34 model
# stride = 2 if i == 0 else 1
resnet_model.residual_block(cnn, 256, stride, version, i == 0)
# ResNet-34 block group 4: removed final block group
# The following 3 lines are intentially commented out to differentiate from
# the original ResNet-34 model
# for i in range(resnet34_layers[3]):
# stride = 2 if i == 0 else 1
# resnet_model.residual_block(cnn, 512, stride, version, i == 0)
# Create saver with mapping from variable names in checkpoint of backbone
# model to variables in SSD model
if not self.backbone_saver:
backbone_var_list = self._collect_backbone_vars()
self.backbone_saver = tf.train.Saver(backbone_var_list)
# --------------------------------------------------------------------------
# SSD additional layers
# --------------------------------------------------------------------------
def add_ssd_layer(cnn, depth, k_size, stride, mode):
return cnn.conv(
depth,
k_size,
k_size,
stride,
stride,
mode=mode,
bias=None,
kernel_initializer=tf.contrib.layers.xavier_initializer())
# Activations for feature maps of different layers
self.activations = [cnn.top_layer]
# Conv7_1, Conv7_2
# Input 38x38, output 19x19
add_ssd_layer(cnn, 256, 1, 1, 'valid')
self.activations.append(add_ssd_layer(cnn, 512, 3, 2, 'same'))
# Conv8_1, Conv8_2
# Input 19x19, output 10x10
add_ssd_layer(cnn, 256, 1, 1, 'valid')
self.activations.append(add_ssd_layer(cnn, 512, 3, 2, 'same'))
# Conv9_1, Conv9_2
# Input 10x10, output 5x5
add_ssd_layer(cnn, 128, 1, 1, 'valid')
self.activations.append(add_ssd_layer(cnn, 256, 3, 2, 'same'))
# Conv10_1, Conv10_2
# Input 5x5, output 3x3
add_ssd_layer(cnn, 128, 1, 1, 'valid')
self.activations.append(add_ssd_layer(cnn, 256, 3, 1, 'valid'))
# Conv11_1, Conv11_2
# Input 3x3, output 1x1
add_ssd_layer(cnn, 128, 1, 1, 'valid')
self.activations.append(add_ssd_layer(cnn, 256, 3, 1, 'valid'))
self.loc = []
self.conf = []
for nd, ac, oc in zip(self.num_dboxes, self.activations,
self.out_chan):
self.loc.append(
tf.reshape(
cnn.conv(nd * 4,
3,
3,
1,
1,
input_layer=ac,
num_channels_in=oc,
activation=None,
bias=None,
kernel_initializer=tf.contrib.layers.
xavier_initializer()),
[ac.get_shape()[0], 4, -1]))
self.conf.append(
tf.reshape(
cnn.conv(nd * self.label_num,
3,
3,
1,
1,
input_layer=ac,
num_channels_in=oc,
activation=None,
bias=None,
kernel_initializer=tf.contrib.layers.
xavier_initializer()),
[ac.get_shape()[0], self.label_num, -1]))
# Shape of locs: [batch_size, 4, NUM_SSD_BOXES]
# Shape of confs: [batch_size, label_num, NUM_SSD_BOXES]
locs, confs = tf.concat(self.loc, 2), tf.concat(self.conf, 2)
# Pack location and confidence outputs into a single output layer
# Shape of logits: [batch_size, 4+label_num, NUM_SSD_BOXES]
logits = tf.concat([locs, confs], 1)
cnn.top_layer = logits
cnn.top_size = 4 + self.label_num
return cnn.top_layer
