def ssd300_blocks(net, end_points):
# block 6: 3x3 conv
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
net = slim.batch_norm(net)
end_points['block6'] = net
# block 7: 1x1 conv
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
net = slim.batch_norm(net)
end_points['block7'] = net
# block 8/9/10/11: 1x1 and 3x3 convolutions with stride 2 (except lasts)
end_point = 'block8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = slim.batch_norm(net)
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
net = slim.batch_norm(net)
end_points[end_point] = net
end_point = 'block9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.batch_norm(net)
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
net = slim.batch_norm(net)
end_points[end_point] = net
end_point = 'block10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.batch_norm(net)
net = slim.conv2d(net, 256, [3, 3], scope='conv3x3', padding='VALID')
net = slim.batch_norm(net)
end_points[end_point] = net
end_point = 'block11'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.batch_norm(net)
net = slim.conv2d(net, 256, [3, 3], scope='conv3x3', padding='VALID')
net = slim.batch_norm(net)
end_points[end_point] = net
return net, end_points
def mobilenet_block(net, num_out_channels, stride=[1, 1], scope=None):
"""Basic MobileNet block combining:
- depthwise conv + BN + relu
- 1x1 conv + BN + relu
"""
with tf.variable_scope(scope, 'block', [net]) as sc:
num_out_channels = int(num_out_channels * width_multiplier)
if stride[0] == 1 and stride[1] == 1:
# Depthwise convolution with stride=1
net = custom_layers.depthwise_convolution2d(net,
kernel_size,
depth_multiplier=1,
stride=stride,
scope='conv_dw')
else:
# Mimic CAFFE padding if stride > 1 => usually better accuracy.
net = custom_layers.pad2d(net, pad=padding)
net = custom_layers.depthwise_convolution2d(net,
kernel_size,
padding='VALID',
depth_multiplier=1,
stride=stride,
scope='conv_dw')
# Pointwise convolution.
net = slim.conv2d(net, num_out_channels, [1, 1], scope='conv_pw')
return net
def largenetwork(net, end_point, end_points):
print('Large network')
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block10'
print(net)
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net, 256, [3, 3], scope='conv3x3', padding='VALID')
end_points[end_point] = net
end_point = 'block11'
print(net)
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net, 256, [3, 3], scope='conv3x3', padding='VALID')
end_points[end_point] = net
print(net)
return end_points
def max_avg_pool2d(net, stride=[2, 2], scope=None):
with tf.variable_scope(scope, 'max_avg_pool', [net]) as sc:
ksize = [3, 3]
padding = [1, 1]
# Additional Caffe padding.
net = custom_layers.pad2d(net, pad=padding)
# Max + Avg pooling.
mnet = slim.max_pool2d(net, ksize, stride, padding='VALID')
anet = slim.avg_pool2d(net, ksize, stride, padding='VALID')
return mnet + anet
def mobilenet_block(net, num_out_channels, stride=[1, 1],
leaders=False, scope=None):
"""Basic MobileNet block combining:
- depthwise conv + BN + relu
- 1x1 conv + BN + relu
"""
with tf.variable_scope(scope, 'block', [net]) as sc:
num_out_channels = int(num_out_channels * width_multiplier)
kernel_size = [3, 3]
if stride[0] == 1 and stride[1] == 1:
# Classic depthwise convolution with stride=1
net = custom_layers.depthwise_convolution2d(
net, kernel_size,
depth_multiplier=1, stride=stride,
scope='conv_dw')
else:
if leaders:
# Special Depthwise Leader convolution when stride > 1
# net = custom_layers.pad2d(net, pad=(1, 1))
net = custom_layers.depthwise_leaders_convolution2d(
net,
kernel_size,
padding='SAME',
stride=stride,
rates=[1, 2, 3],
pooling_sizes=[5, 3, 1],
pooling_type='AVG',
activation_fn=tf.nn.relu,
scope='conv_lead_dw')
else:
# Mimic CAFFE padding if stride > 1.
net = custom_layers.pad2d(net, pad=(1, 1))
net = custom_layers.depthwise_convolution2d(
net, kernel_size, padding='VALID',
depth_multiplier=1, stride=stride,
scope='conv_dw')
# Pointwise convolution.
net = slim.conv2d(net, num_out_channels, [1, 1],
scope='conv_pw')
return net
def ssd_net(inputs,
num_classes=SSDNet.default_params.num_classes,
feat_layers=SSDNet.default_params.feat_layers,
anchor_sizes=SSDNet.default_params.anchor_sizes,
anchor_ratios=SSDNet.default_params.anchor_ratios,
normalizations=SSDNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
prediction_fn=slim.softmax,
reuse=None,
scope='ssd_300_vgg'):
"""SSD net definition.
"""
#if data_format == 'NCHW':
# inputs = tf.transpose(inputs, perm=(0, 3, 1, 2))
# End_points collect relevant activations for external use.
end_points = {}
with tf.variable_scope(scope, 'ssd_300_vgg', [inputs], reuse=reuse):
#inception
'''
net1 = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
end_points['block1'] = net1
net1 = slim.max_pool2d(net1, [2, 2], scope='pool1')
# Block 2.
#net2 = slim.repeat(net1, 2, slim.conv2d, 128, [3, 3], scope='conv2')
net2_1 = slim.conv2d(net1, 128, [3, 3], scope='conv2_0')
net2_2 = slim.conv2d(net2_1, 128, [3, 3], scope='conv2_1')
net2 = net2_1 + net2_2
end_points['block2'] = net2
net2 = slim.max_pool2d(net2, [2, 2], scope='pool2')
# Block 3.
#net3 = slim.repeat(net2, 3, slim.conv2d, 256, [3, 3], scope='conv3')
net3_1 = slim.conv2d(net2, 256, [3, 3], scope='conv3_1')
#net3_1 = slim.batch_norm(net3_1, scope='bn3_1')
net3_2 = slim.conv2d(net3_1, 256, [3, 3], scope='conv3_2')
#net3_2 = slim.batch_norm(net3_2, scope='bn3_2')
net3_3 = slim.conv2d(net3_2, 256, [3, 3], scope='conv3_3')
net3 = net3_1 + net3_3
end_points['block3'] = net3
net3 = slim.max_pool2d(net3, [2, 2], scope='pool3')
# Block 4.
net4 = slim.repeat(net3, 2, slim.conv2d, 512, [3, 3], scope='conv4')
net4 = slim.batch_norm(net4, scope='conv4_bn')
with tf.variable_scope('Mixed_4'):
with tf.variable_scope('Branch_0'):
tower_conv = slim.conv2d(net4, 256, 1, scope='Conv2d_1x1')
# 38,38,128
with tf.variable_scope('Branch_1'):
tower_conv1_0 = slim.conv2d(net4, 256, 1, scope='Conv2d_0a_1x1')
tower_conv1_1 = slim.conv2d(tower_conv1_0, 256, 3, scope='Conv2d_0b_3x3')
# 38,38,128
with tf.variable_scope('Branch_2'):
tower_conv2_0 = slim.conv2d(net4, 128, 1, scope='Conv2d_0a_1x1')
tower_conv2_1 = slim.conv2d(tower_conv2_0, 128, 3, scope='Conv2d_0b_3x3')
tower_conv2_2 = slim.conv2d(tower_conv2_1, 256, 3, scope='Conv2d_0c_3x3')
# 38,38,128
with tf.variable_scope('Branch_3'):
tower_pool = slim.avg_pool2d(net4, 3, stride=1, padding='SAME', scope='AvgPool_0a_3x3')
tower_pool_1 = slim.conv2d(tower_pool, 256, 1, scope='Conv2d_0b_1x1')
net4 = tf.concat([tower_conv, tower_conv1_1,
tower_conv2_2, tower_pool_1], 1)
end_points['block4'] = net4
net4 = slim.max_pool2d(net4, [2, 2], scope='pool4')
# Block 5.
net = slim.repeat(net4, 3, slim.conv2d, 512, [3, 3], scope='conv5')
end_points['block5'] = net
net = slim.max_pool2d(net, [3, 3], stride=1, scope='pool5')
# Additional SSD blocks.
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
end_points['block6'] = net
net = tf.layers.dropout(net, rate=dropout_keep_prob, training=is_training)
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
end_points['block7'] = net
net = tf.layers.dropout(net, rate=dropout_keep_prob, training=is_training)
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'block8'
net = tf.layers.batch_normalization(net, training=True)
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net, 512, [3, 3], stride=2, scope='conv3x3', padding='VALID')
end_points[end_point] = net
end_point = 'block9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net, 256, [3, 3], stride=2, scope='conv3x3', padding='VALID')
end_points[end_point] = net
end_point = 'block10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net, 256, [3, 3], scope='conv3x3', padding='VALID')
end_points[end_point] = net
end_point = 'block11'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net, 256, [3, 3], scope='conv3x3', padding='VALID')
end_points[end_point] = net
'''
#resnet
net1_ = slim.conv2d(inputs, 64, [3,3], scope="conv1")
# Block1.
net1 = slim.repeat(net1_, 3, slim.conv2d, 64, [3, 3], scope='conv1')
#net1_ = tf.concat([net1_]*2,1)
#print(net1_.shape)
net1 = net1 + net1_
end_points['block1'] = net1
net1 = slim.max_pool2d(net1, [2, 2], scope='pool1')
# 150,150,64
# Block 2.
net2 = slim.repeat(net1, 3, slim.conv2d, 128, [3, 3], scope='conv2')
net1 = tf.concat([net1]*2,1)
net2 = net2+net1
end_points['block2'] = net2
net2 = slim.max_pool2d(net2, [2, 2], scope='pool2')
# 75,75,128
# Block 3.
net3 = slim.repeat(net2, 3, slim.conv2d, 256, [3, 3], scope='conv3')
net2 = tf.concat([net2]*2,1)
net3_p = net3 + net2
#end_points['block3'] = net3
net3 = slim.max_pool2d(net3_p, [2, 2], scope='pool3')
# 38,38,256
# Block 3.
net4 = slim.repeat(net3, 3, slim.conv2d, 512, [3, 3], scope='conv4')
net3 = tf.concat([net3] * 2, 1)
net4_p = net4 + net3
end_points['block4'] = net4_p
net4 = slim.max_pool2d(net4_p, [2, 2], scope='pool4')
# 19,19,512
# Block 5.
net5 = slim.repeat(net4, 3, slim.conv2d, 512, [3, 3], scope='conv5')
net5 = net4 + net5
end_points['block5'] = net5
net5 = slim.max_pool2d(net5, [3, 3], stride=1, scope='pool5')
# 17
# Additional SSD blocks.
# Block 6: let's dilate the hell out of it!
net6 = slim.conv2d(net5, 1024, [3, 3], rate=6, scope='conv6')
end_points['block6'] = net6
net6 = tf.layers.dropout(net6, rate=dropout_keep_prob, training=is_training)
# 17,17
# Block 7: 1x1 conv. Because the f**k.
net7 = slim.conv2d(net6, 1024, [1, 1], scope='conv7')
#end_points['block7'] = net7
net7_ = tf.layers.dropout(net7, rate=dropout_keep_prob, training=is_training)
# 17,17
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'block8'
net8 = tf.layers.batch_normalization(net7_, training=True)
with tf.variable_scope(end_point):
net8 = slim.conv2d(net8, 256, [1, 1], scope='conv1x1')
#17
net8 = custom_layers.pad2d(net8, pad=(1, 1))
#21
net8 = slim.conv2d(net8, 512, [3, 3], stride=2, scope='conv3x3', padding='VALID')
#10,10
#end_points[end_point] = net8
end_point = 'block9'
with tf.variable_scope(end_point):
net9 = slim.conv2d(net8, 128, [1, 1], scope='conv1x1')
net9 = custom_layers.pad2d(net9, pad=(1, 1))
net9 = slim.conv2d(net9, 256, [3, 3], stride=2, scope='conv3x3', padding='VALID')
#5，5
#end_points[end_point] = net9
end_point = 'block10'
with tf.variable_scope(end_point):
net10 = slim.conv2d(net9, 128, [1, 1], scope='conv1x1')
net10 = slim.conv2d(net10, 256, [3, 3], scope='conv3x3', padding='VALID')
#3，3
#end_points[end_point] = net10
end_point = 'block11'
with tf.variable_scope(end_point):
net11 = slim.conv2d(net10, 128, [1, 1], scope='conv1x1')
net11 = slim.conv2d(net11, 256, [3, 3], scope='conv3x3', padding='VALID')
#1，1
end_points[end_point] = net11
end_point = 'block10'
with tf.variable_scope(end_point):
net10_a = tf.transpose(net11, perm=(0, 2, 3, 1))#nchw
net10_a = tf.image.resize_nearest_neighbor(net10_a, (3,3))
net10_a = tf.transpose(net10_a, perm=(0, 3, 1, 2))#nchw
net10_a = slim.conv2d(net10_a, 256, [3,3], scope='pre10_3x3')
net10_b = slim.conv2d(net10, 256, [1, 1], scope='pre10_1x1')
net10_o = net10_a + net10_b
end_points[end_point] = net10_o #3
end_point = 'block9'
with tf.variable_scope(end_point):
net9_a = tf.transpose(net10_o, perm=(0, 2, 3, 1))#nchw
net9_a = tf.image.resize_nearest_neighbor(net9_a, (5,5))
net9_a = tf.transpose(net9_a, perm=(0, 3, 1, 2))#nchw
net9_a = slim.conv2d(net9_a, 256, [3,3], scope='pre9_3x3')
net9_b = slim.conv2d(net9, 256, [1, 1], scope='pre9_1x1')
net9_o = net9_a + net9_b
end_points[end_point] = net9_o#5
end_point = 'block8'
with tf.variable_scope(end_point):
net8_a = tf.transpose(net9_o, perm=(0, 2, 3, 1))#nchw
net8_a = tf.image.resize_nearest_neighbor(net8_a, (10,10))
net8_a = tf.transpose(net8_a, perm=(0, 3, 1, 2))#nchw
net8_a = slim.conv2d(net8_a, 512, [3,3], padding='SAME', scope='pre8_3x3')
net8_b = slim.conv2d(net8, 512, [1, 1], scope='pre8_1x1')#10
net8_o = net8_a + net8_b
end_points[end_point] = net8_o#10
end_point = 'block7'
with tf.variable_scope(end_point):
net7_a = tf.transpose(net8_o, perm=(0, 2, 3, 1))#nchw
net7_a = tf.image.resize_nearest_neighbor(net7_a, (19, 19))#
net7_a = tf.transpose(net7_a, perm=(0, 3, 1, 2))#nchw
net7_a = slim.conv2d(net7_a, 1024, [3, 3], padding='SAME', scope='pre7_3x3')
net7_b = slim.conv2d(net7, 1024, [1, 1], scope='pre7_1x1')
net7_o = net7_a + net7_b
end_points[end_point] = net7_o
end_point = 'block4'
with tf.variable_scope(end_point):
net4_a = tf.transpose(net7_o, perm=(0, 2, 3, 1))#nchw
net4_a = tf.image.resize_nearest_neighbor(net4_a, (38, 38))#
net4_a = tf.transpose(net4_a, perm=(0, 3, 1, 2))#nchw
net4_a = slim.conv2d(net4_a, 512, [3, 3], padding='SAME', scope='pre4_3x3')
net4_b = slim.conv2d(net4_p, 512, [1, 1], scope='pre4_1x1')
#print("asdfasdfasdf", net4_a.shape, net4_b.shape)
net4_o = net4_a + net4_b#38
end_points[end_point] = net4_o
end_point = 'block3'
with tf.variable_scope(end_point):
net3_a = tf.transpose(net4_o, perm=(0, 2, 3, 1))#nchw
net3_a = tf.image.resize_nearest_neighbor(net3_a, (75, 75))#
net3_a = tf.transpose(net3_a, perm=(0, 3, 1, 2))#nchw
net3_a = slim.conv2d(net3_a, 512, [3, 3], padding='SAME', scope='pre3_3x3')
net3_b = slim.conv2d(net3_p, 512, [1, 1], scope='pre3_1x1')
print("asdfasdfasdf", net3_a.shape, net3_b.shape)
net3_o = net3_a + net3_b#75
end_points[end_point] = net3_o
#original
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
end_points['block1'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool1')
# Block 2.
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
end_points['block2'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool2')
# Block 3.
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
end_points['block3'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool3')
# Block 4.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
end_points['block4'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool4')
# Block 5.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
end_points['block5'] = net
net = slim.max_pool2d(net, [3, 3], stride=1, scope='pool5')
# Additional SSD blocks.
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
end_points['block6'] = net
net = tf.layers.dropout(net, rate=dropout_keep_prob, training=is_training)
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
end_points['block7'] = net
net = tf.layers.dropout(net, rate=dropout_keep_prob, training=is_training)
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'block8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net, 512, [3, 3], stride=2, scope='conv3x3', padding='VALID')
end_points[end_point] = net
end_point = 'block9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net, 256, [3, 3], stride=2, scope='conv3x3', padding='VALID')
end_points[end_point] = net
end_point = 'block10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net, 256, [3, 3], scope='conv3x3', padding='VALID')
end_points[end_point] = net
end_point = 'block11'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net, 256, [3, 3], scope='conv3x3', padding='VALID')
end_points[end_point] = net
# Prediction and localisations layers.
predictions = []
logits = []
localisations = []
for i, layer in enumerate(feat_layers):
with tf.variable_scope(layer + '_box'):
p, l = ssd_multibox_layer(end_points[layer],
num_classes,
anchor_sizes[i],
anchor_ratios[i],
normalizations[i])
predictions.append(prediction_fn(p))
logits.append(p)
localisations.append(l)
return predictions, localisations, logits, end_points
def ssd_net(inputs,
num_classes=SSDNet.default_params.num_classes,
feat_layers=SSDNet.default_params.feat_layers,
anchor_sizes=SSDNet.default_params.anchor_sizes,
anchor_ratios=SSDNet.default_params.anchor_ratios,
normalizations=SSDNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
prediction_fn=slim.softmax,
reuse=None,
scope='ssd_300_vgg'):
"""SSD net definition.
"""
# if data_format == 'NCHW':
# inputs = tf.transpose(inputs, perm=(0, 3.txt, 1, 2))
# End_points collect relevant activations for external use.
end_points = {}
with tf.variable_scope(scope, 'ssd_300_vgg', [inputs], reuse=reuse):
# Original VGG-16 blocks.
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
end_points['block1'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool1')
# Block 2.
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
end_points['block2'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool2')
# Block 3.txt.
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
end_points['block3'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool3')
# Block 4.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
end_points['block4'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool4')
# Block 5.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
end_points['block5'] = net
net = slim.max_pool2d(net, [3, 3], stride=1, scope='pool5')
# Additional SSD blocks.
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
end_points['block6'] = net
net = tf.layers.dropout(net,
rate=dropout_keep_prob,
training=is_training)
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
end_points['block7'] = net
net = tf.layers.dropout(net,
rate=dropout_keep_prob,
training=is_training)
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'block8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block11'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
# Prediction and localisations layers.
predictions = []
logits = []
localisations = []
for i, layer in enumerate(
feat_layers
): #feat_layers=['block4', 'block7', 'block8', 'block9', 'block10', 'block11'],
with tf.variable_scope(layer + '_box'):
p, l = ssd_multibox_layer(end_points[layer], num_classes,
anchor_sizes[i], anchor_ratios[i],
normalizations[i])
predictions.append(prediction_fn(p))
logits.append(p)
localisations.append(l)
'''
p和l是没有进行softmax的，而p预测的是每个框的类别所以要加入一个prediction_fn(softmax)
'''
return predictions, localisations, logits, end_points
def text_net(inputs,
feat_layers=TextboxNet.default_params.feat_layers,
anchor_sizes=TextboxNet.default_params.anchor_sizes,
anchor_ratios = TextboxNet.default_params.anchor_ratios,
normalizations=TextboxNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
reuse=None,
scope='text_box_384'):
end_points = {}
with tf.variable_scope(scope, 'text_box_300', [inputs], reuse=reuse): # 300*300 384*383
# Original VGG-16 blocks.
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1') # 300 384
end_points['conv1'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool1') # 150
# Block 2.
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2') # 150 192
end_points['conv2'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool2') # 75
# Block 3.
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3') # 75 81
end_points['conv3'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool3') # 38
# Block 4.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4') # 38 40
end_point = 'conv4'
end_points[end_point] = net
net = slim.max_pool2d(net, [2, 2], scope='pool4') # 19
# Block 5.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5') # 19
end_points['conv5'] = net
net = slim.max_pool2d(net, [3, 3], stride=1, scope='pool5') # 19
# Additional SSD blocks.
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6') # 19
end_points['conv6'] = net
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(net, 1024, [1, 1], scope='conv7') # 19
end_point = 'conv7'
end_points[end_point] = net
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2
end_point = 'conv8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net, 512, [3, 3], stride=2, scope='conv3x3', padding='VALID')
end_points[end_point] = net # 10
end_point = 'conv9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net, 256, [3, 3], stride=2, scope='conv3x3', padding='VALID')
end_points[end_point] = net # 5
end_point = 'conv10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1', padding= 'VALID')
net = slim.conv2d(net, 256, [3, 3], scope='conv3x3', padding='VALID')
end_points[end_point] = net # 3
end_point = 'conv11'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net, 256, [3, 3], scope='conv3x3', padding='VALID')
end_points[end_point] = net #
end_point = feat_layers[0]
with tf.variable_scope(end_point):
net_dilation1 = slim.conv2d(end_points[end_point], 128, [3, 3], stride=1, scope='dilation1')
# net_dilation2 = custom_layers.pad2d(net, pad=(0, 4))
net_dilation2 = slim.conv2d(end_points[end_point], 128, [1, 9], padding='SAME', stride=1, scope='dilation2')
net_dilation3 = slim.conv2d(end_points[end_point], 128, [9, 1], stride=1, padding='SAME', scope='dilation3')
# net_dilation3 = custom_layers.pad2d(net_dilation3, pad=(4, 0))
net_inception = tf.concat(values=[net_dilation1, net_dilation2, net_dilation3], axis=3)
end_points[end_point] = net_inception
end_point= feat_layers[1]
with tf.variable_scope(end_point):
net_dilation1 = slim.conv2d(end_points[end_point], 1024, [1, 1], stride=1, scope='dilation1')
net_dilation2 = slim.conv2d(end_points[end_point], 1024, [1, 7], stride=1, scope='dilation2')
# net_dilation2 = custom_layers.pad2d(net_dilation2, pad=(0, 3))
net_dilation3 = slim.conv2d(end_points[end_point], 1024, [7, 1], stride=1, scope='dilation3')
# net_dilation3 = custom_layers.pad2d(net_dilation3, pad=(3, 0))
net_inception = tf.concat([net_dilation1, net_dilation2, net_dilation3], axis=3)
end_points[end_point] = net_inception
end_point = 'conv8'
with tf.variable_scope(end_point):
net_dilation1 = slim.conv2d(end_points[end_point], 128, [1, 1], stride=1,scope='dilation1')
net_dilation2 = slim.conv2d(end_points[end_point], 128, [1, 7], stride=1, scope='dilation2')
# net_dilation2 = custom_layers.pad2d(net_dilation2, pad=(0, 3))
net_dilation3 = slim.conv2d(end_points[end_point], 128, [7, 1], stride=1, scope='dilation3')
# net_dilation3 = custom_layers.pad2d(net_dilation3, pad=(3, 0))
net_inception = tf.concat([net_dilation1, net_dilation2, net_dilation3], axis=3)
end_points[end_point] = net_inception
end_point = feat_layers[3]
with tf.variable_scope(end_point):
net_dilation1 = slim.conv2d(end_points[end_point], 128, [1, 1], stride=1,scope='dilation1')
net_dilation2 = slim.conv2d(end_points[end_point], 128, [1, 7], stride=1, scope='dilation2')
# net_dilation2 = custom_layers.pad2d(net_dilation2, pad=(0, 3))
net_dilation3 = slim.conv2d(end_points[end_point], 128, [7, 1], stride=1, scope='dilation3')
# net_dilation3 = custom_layers.pad2d(net_dilation3, pad=(3, 0))
net_inception = tf.concat([net_dilation1, net_dilation2, net_dilation3], axis=3)
end_points[end_point] = net_inception # 5
end_point = 'conv10'
with tf.variable_scope(end_point):
net_dilation1 = slim.conv2d(end_points[end_point], 128, [1, 1], stride=1, scope='dilation1')
net_dilation2 = slim.conv2d(end_points[end_point], 128, [1, 7], stride=1, scope='dilation2')
# net_dilation2 = custom_layers.pad2d(net_dilation2, pad=(0, 3))
net_dilation3 = slim.conv2d(end_points[end_point], 128, [7, 1], stride=1, scope='dilation3')
# net_dilation3 = custom_layers.pad2d(net_dilation3, pad=(3, 0))
net_inception = tf.concat([net_dilation1, net_dilation2, net_dilation3], axis=3)
end_points[end_point] = net_inception # 3
end_point = 'conv11'
with tf.variable_scope(end_point):
net_dilation1 = slim.conv2d(end_points[end_point], 128, [1, 1], stride=1,scope='dilation1')
net_dilation2 = slim.conv2d(end_points[end_point], 128, [1, 5], stride=1, scope='dilation2')
# net_dilation2 = custom_layers.pad2d(net_dilation2, pad=(0, 2))
net_dilation3 = slim.conv2d(end_points[end_point], 128, [5, 1], stride=1, scope='dilation3')
# net_dilation3 = custom_layers.pad2d(net_dilation3, pad=(2, 0))
net_inception = tf.concat([net_dilation1, net_dilation2, net_dilation3], axis=3)
end_points[end_point] = net_inception # 1
# Prediction and localisations layers.
predictions = []
logits = []
localisations = []
for i, layer in enumerate(feat_layers):
with tf.variable_scope(layer + '_box'):
p, loc = text_multibox_layer(layer,
end_points[layer],
anchor_sizes[i],
anchor_ratios[i],
normalizations[i])
prediction_fn = slim.softmax
predictions.append(prediction_fn(p))
logits.append(p)
localisations.append(loc)
return predictions, localisations, logits, end_points
def ssd_net(inputs,
num_classes=SSDNet.default_params.num_classes,
feat_layers=SSDNet.default_params.feat_layers,
anchor_sizes=SSDNet.default_params.anchor_sizes,
anchor_ratios=SSDNet.default_params.anchor_ratios,
normalizations=SSDNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
prediction_fn=slim.softmax,
reuse=None,
scope='ssd_512_vgg'):
"""SSD net definition.
"""
# End_points collect relevant activations for external use.
end_points = {}
with tf.variable_scope(scope, 'ssd_512_vgg', [inputs], reuse=reuse):
# Original VGG-16 blocks.
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
end_points['block1'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool1')
# Block 2.
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
end_points['block2'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool2')
# Block 3.
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
end_points['block3'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool3')
# Block 4.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
end_points['block4'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool4')
# Block 5.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
end_points['block5'] = net
net = slim.max_pool2d(net, [3, 3], 1, scope='pool5')
# Additional SSD blocks.
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
end_points['block6'] = net
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
end_points['block7'] = net
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'block8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net, 512, [3, 3], stride=2, scope='conv3x3', padding='VALID')
end_points[end_point] = net
end_point = 'block9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net, 256, [3, 3], stride=2, scope='conv3x3', padding='VALID')
end_points[end_point] = net
end_point = 'block10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net, 256, [3, 3], stride=2, scope='conv3x3', padding='VALID')
end_points[end_point] = net
end_point = 'block11'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net, 256, [3, 3], stride=2, scope='conv3x3', padding='VALID')
end_points[end_point] = net
end_point = 'block12'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net, 256, [4, 4], scope='conv4x4', padding='VALID')
# Fix padding to match Caffe version (pad=1).
# pad_shape = [(i-j) for i, j in zip(layer_shape(net), [0, 1, 1, 0])]
# net = tf.slice(net, [0, 0, 0, 0], pad_shape, name='caffe_pad')
end_points[end_point] = net
# Prediction and localisations layers.
predictions = []
logits = []
localisations = []
for i, layer in enumerate(feat_layers):
with tf.variable_scope(layer + '_box'):
p, l = ssd_vgg_300.ssd_multibox_layer(end_points[layer],
num_classes,
anchor_sizes[i],
anchor_ratios[i],
normalizations[i])
predictions.append(prediction_fn(p))
logits.append(p)
localisations.append(l)
return predictions, localisations, logits, end_points
def text_net(inputs,
feat_layers=TextboxNet.default_params.feat_layers,
normalizations=TextboxNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
reuse=None,
scope='text_box_300'):
end_points = {}
with tf.variable_scope(scope, 'text_box_300', [inputs], reuse=reuse):
# Original VGG-16 blocks.
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
end_points['conv1'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool1')
# Block 2.
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
end_points['conv2'] = net # 150,150 128
net = slim.max_pool2d(net, [2, 2], scope='pool2')
# Block 3. # 75 75 256
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
end_points['conv3'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool3', padding='SAME')
# Block 4. # 38 38 512
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
end_points['conv4'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool4')
# Block 5. # 19 19 512
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
end_points['conv5'] = net
net = slim.max_pool2d(net, [3, 3],
stride=1,
scope='pool5',
padding='SAME')
# Additional SSD blocks.
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
end_points['conv6'] = net
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
end_points['conv7'] = net
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'conv8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'conv9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'conv10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'global'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
# Prediction and localisations layers.
predictions = []
logits = []
localisations = []
for i, layer in enumerate(feat_layers):
with tf.variable_scope(layer + '_box'):
p, l = text_multibox_layer(layer, end_points[layer],
normalizations[i])
#predictions.append(prediction_fn(p))
logits.append(p)
localisations.append(l)
return localisations, logits, end_points
def text_net(inputs,
feat_layers=TextboxNet.default_params.feat_layers,
anchor_sizes=TextboxNet.default_params.anchor_sizes,
anchor_ratios=TextboxNet.default_params.anchor_ratios,
normalizations=TextboxNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
reuse=None,
scope='text_box_384',
update_feat_shapes=False):
end_points = {}
with tf.compat.v1.variable_scope(scope,
'text_box_300', [inputs],
reuse=reuse): # 300*300 384*383
# Original VGG-16 blocks.
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3],
scope='conv1') # 300 384
end_points['conv1'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool1') # 150
# Block 2.
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3],
scope='conv2') # 150 192
end_points['conv2'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool2') # 75
# Block 3.
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3],
scope='conv3') # 75 81
end_points['conv3'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool3') # 38
# Block 4.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3],
scope='conv4') # 38 40
end_point = 'conv4'
end_points[end_point] = net
net = slim.max_pool2d(net, [2, 2], scope='pool4') # 19
# Block 5.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3],
scope='conv5') # 19
end_points['conv5'] = net
net = slim.max_pool2d(net, [3, 3], stride=1, scope='pool5') # 19
# Additional SSD blocks.
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6') # 19
end_points['conv6'] = net
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(net, 1024, [1, 1], scope='conv7') # 19
end_point = 'conv7'
end_points[end_point] = net
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2
end_point = 'conv8'
with tf.compat.v1.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net # 10
end_point = 'conv9'
with tf.compat.v1.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net # 5
end_point = 'conv10'
with tf.compat.v1.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net # 3
end_point = 'conv11'
with tf.compat.v1.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net #
# Prediction and localisations layers.
predictions = []
logits = []
localisations = []
shape_list = []
for i, layer in enumerate(feat_layers):
with tf.compat.v1.variable_scope(layer + '_box'):
p, loc, shape = text_multibox_layer(layer, end_points[layer],
anchor_sizes[i],
anchor_ratios[i],
normalizations[i])
prediction_fn = slim.softmax
predictions.append(prediction_fn(p))
logits.append(p)
localisations.append(loc)
shape_list.append(shape)
if update_feat_shapes is True:
return predictions, localisations, logits, end_points, shape_list
else:
return predictions, localisations, logits, end_points
def ssd_net(inputs,
num_classes=SSDNet.default_params.num_classes,
feat_layers=SSDNet.default_params.feat_layers,
anchor_sizes=SSDNet.default_params.anchor_sizes,
anchor_ratios=SSDNet.default_params.anchor_ratios,
normalizations=SSDNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
prediction_fn=slim.softmax,
reuse=None,
scope='ssd_300_vgg'):
"""SSD net definition.
SSD网络的定义
ssd_net（输入，类的数量，特征层的名称列表，anchor的实际大小，anchor的长宽比，正则化，是否训练，扔出网络的概率
，预测，重用，命名绑定）
"""
# if data_format == 'NCHW':
# inputs = tf.transpose(inputs, perm=(0, 3, 1, 2))
# End_points collect relevant activations for external use.
end_points = {}
with tf.variable_scope(
scope, 'ssd_300_vgg', [inputs],
reuse=reuse): #命名空间管理：创建命名空间，reuse=none\false时只能在该命名空间创建变量
# Original VGG-16 blocks.
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3],
scope='conv1') #conv1卷积
end_points['block1'] = net #300X300
net = slim.max_pool2d(net, [2, 2], scope='pool1')
# Block 2.
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
end_points['block2'] = net #150x150
net = slim.max_pool2d(net, [2, 2], scope='pool2')
# Block 3.
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
end_points['block3'] = net #75x75
net = slim.max_pool2d(net, [2, 2], scope='pool3')
# Block 4.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
end_points['block4'] = net #38x38
net = slim.max_pool2d(net, [2, 2], scope='pool4')
# Block 5.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
end_points['block5'] = net #19x19
net = slim.max_pool2d(net, [3, 3], stride=1, scope='pool5')
# Additional SSD blocks.
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6') #13x13的感受野
end_points['block6'] = net
net = tf.layers.dropout(net,
rate=dropout_keep_prob,
training=is_training)
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
end_points['block7'] = net
net = tf.layers.dropout(net,
rate=dropout_keep_prob,
training=is_training)
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'block8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net #10x10
end_point = 'block9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net #5x5
end_point = 'block10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net #3X3
end_point = 'block11'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net #1X1
# Prediction and localisations layers. 预测和位置层
predictions = []
logits = []
localisations = []
for i, layer in enumerate(feat_layers):
with tf.variable_scope(layer + '_box'):
p, l = ssd_multibox_layer(end_points[layer], num_classes,
anchor_sizes[i], anchor_ratios[i],
normalizations[i])
predictions.append(prediction_fn(p))
logits.append(p) #对每个特征图类别的网络
localisations.append(l) #对每个特征图测量位置的网络
return predictions, localisations, logits, end_points
def ssd_net(inputs,
num_classes=SSDNet.default_params.num_classes,
feat_layers=SSDNet.default_params.feat_layers,
anchor_sizes=SSDNet.default_params.anchor_sizes,
anchor_ratios=SSDNet.default_params.anchor_ratios,
normalizations=SSDNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
prediction_fn=slim.softmax,
reuse=None,
scope='ssd_300_mobilenetv1'):
"""SSD net definition.
"""
# if data_format == 'NCHW':
# inputs = tf.transpose(inputs, perm=(0, 3, 1, 2))
# End_points collect relevant activations for external use.
end_points = {}
min_depth = 32
depth_multiplier = 1.0
with tf.variable_scope(scope, 'ssd_300_mobilenetv1', [inputs], reuse=reuse):
input_shape = inputs.get_shape().as_list()
if len(input_shape) != 4:
raise ValueError('Invalid input tensor rank, expected 4, was: %d' %
len(input_shape))
with slim.arg_scope(mobilenet_v1.mobilenet_v1_arg_scope(is_training=is_training)):
with slim.arg_scope([slim.batch_norm, slim.dropout], is_training=is_training):
net, end_points = mobilenet_v1.mobilenet_v1_base(inputs, scope='MobilenetV1',
min_depth=min_depth,
depth_multiplier=depth_multiplier,
conv_defs=None)
'''
# Additional SSD blocks.
# Block 6: let's dilate the hell out of it!
end_point = 'block13'
with tf.variable_scope(end_point)
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='atrous_conv')
end_points['block13_atrous'] = net
net = tf.layers.dropout(net, rate=dropout_keep_prob, training=is_training)
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(net, 1024, [1, 1], scope='conv1x1')
end_points['block13'] = net
net = tf.layers.dropout(net, rate=dropout_keep_prob, training=is_training)
'''
# Block 14/15/16/17: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'block14'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], biases_initializer=None, trainable=is_training, scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net, 512, [3, 3], biases_initializer=None, trainable=is_training, stride=2, scope='conv3x3', padding='VALID')
end_points[end_point] = net
end_point = 'block15'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], biases_initializer=None, trainable=is_training, scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net, 256, [3, 3], biases_initializer=None, trainable=is_training, stride=2, scope='conv3x3', padding='VALID')
end_points[end_point] = net
end_point = 'block16'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], biases_initializer=None, trainable=is_training, scope='conv1x1')
net = slim.conv2d(net, 256, [3, 3], biases_initializer=None, trainable=is_training, scope='conv3x3', padding='VALID')
end_points[end_point] = net
'''
end_point = 'block17'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 64, [1, 1], biases_initializer=None, trainable=is_training, scope='conv1x1')
net = slim.conv2d(net, 128, [3, 3], biases_initializer=None, trainable=is_training, scope='conv3x3', padding='VALID')
end_points[end_point] = net
'''
# Prediction and localisations layers.
predictions = []
logits = []
localisations = []
for i, layer in enumerate(feat_layers):
with tf.variable_scope(layer + '_box'):
p, l = ssd_multibox_layer(end_points[layer],
num_classes,
anchor_sizes[i],
anchor_ratios[i],
normalizations[i],
is_training=is_training)
predictions.append(prediction_fn(p))
logits.append(p)
localisations.append(l)
# end_points['logits'] = logits
# end_points['predictions'] = predictions
# end_points['localisations'] = localisations
return predictions, localisations, logits, end_points
def mobilenets(inputs,
num_classes=1000,
width_multiplier=1.0,
is_training=True,
dropout_keep_prob=0.5,
pad_logits=True,
scope='MobileNets'):
"""MobileNets implementation.
Args:
inputs: a tensor of size [batch_size, height, width, channels].
num_classes: number of predicted classes.
is_training: whether or not the model is being trained.
dropout_keep_prob: the probability that activations are kept in the dropout
layers during training.
scope: Optional scope for the variables.
Returns:
the last op containing the log predictions and end_points dict.
"""
# MobileNets kernel size and padding (for layers with stride > 1).
kernel_size = [3, 3]
padding = [(kernel_size[0] - 1) // 2, (kernel_size[1] - 1) // 2]
def mobilenet_block(net, num_out_channels, stride=[1, 1], scope=None):
"""Basic MobileNet block combining:
- depthwise conv + BN + relu
- 1x1 conv + BN + relu
"""
with tf.variable_scope(scope, 'block', [net]) as sc:
num_out_channels = int(num_out_channels * width_multiplier)
if stride[0] == 1 and stride[1] == 1:
# Depthwise convolution with stride=1
net = custom_layers.depthwise_convolution2d(net,
kernel_size,
depth_multiplier=1,
stride=stride,
scope='conv_dw')
else:
# Mimic CAFFE padding if stride > 1 => usually better accuracy.
net = custom_layers.pad2d(net, pad=padding)
net = custom_layers.depthwise_convolution2d(net,
kernel_size,
padding='VALID',
depth_multiplier=1,
stride=stride,
scope='conv_dw')
# Pointwise convolution.
net = slim.conv2d(net, num_out_channels, [1, 1], scope='conv_pw')
return net
with tf.variable_scope(scope, 'MobileNets', [inputs]) as sc:
end_points = {}
# First full convolution...
net = custom_layers.pad2d(inputs, pad=padding)
net = slim.conv2d(net,
32,
kernel_size,
stride=[2, 2],
padding='VALID',
scope='conv1')
# net = slim.conv2d(inputs, 32, kernel_size, stride=[2, 2],
# padding='SAME', scope='conv1')
# Then, MobileNet blocks!
net = mobilenet_block(net, 64, scope='block2')
net = mobilenet_block(net, 128, stride=[2, 2], scope='block3')
net = mobilenet_block(net, 128, scope='block4')
net = mobilenet_block(net, 256, stride=[2, 2], scope='block5')
net = mobilenet_block(net, 256, scope='block6')
net = mobilenet_block(net, 512, stride=[2, 2], scope='block7')
# Intermediate blocks...
for i in range(5):
net = mobilenet_block(net, 512, scope='block%i' % (i + 8))
# Final blocks.
net = mobilenet_block(net, 1024, stride=[2, 2], scope='block13')
net = mobilenet_block(net, 1024, scope='block14')
# Spatial pooling + fully connected layer.
net = custom_layers.spatial_mean(net,
keep_dims=True,
scope='spatial_mean14')
net = slim.conv2d(net,
1000, [1, 1],
activation_fn=None,
normalizer_fn=None,
normalizer_params=None,
biases_initializer=tf.zeros_initializer(),
scope='conv_fc15')
net = custom_layers.spatial_squeeze(net)
# Logits padding: get everyone to the same number of classes.
if pad_logits:
net = custom_layers.pad_logits(net, pad=(num_classes - 1000, 0))
return net, end_points
def ssd_net(inputs,
num_classes=SSDNet.default_params.num_classes,
feat_layers=SSDNet.default_params.feat_layers,
anchor_sizes=SSDNet.default_params.anchor_sizes,
anchor_ratios=SSDNet.default_params.anchor_ratios,
normalizations=SSDNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
prediction_fn=slim.softmax,
reuse=None,
scope='ssd_300_vgg'):
"""SSD net definition.
"""
#if data_format == 'NCHW':
#inputs = tf.transpose(inputs, perm=(0, 3, 1, 2))
print("====", inputs.shape)
# End_points collect relevant activations for external use.
end_points = {}
with tf.variable_scope(scope, 'ssd_300_vgg', [inputs], reuse=reuse):
#original
'''
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
end_points['block1'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool1')
# Block 2.
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
end_points['block2'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool2')
# Block 3.
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
end_points['block3'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool3')
# Block 4.
net4 = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
# 38x38
#end_points['block4'] = net
net = slim.max_pool2d(net4, [2, 2], scope='pool4')
# Block 5.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
end_points['block5'] = net
net = slim.max_pool2d(net, [3, 3], stride=1, scope='pool5')
# Additional SSD blocks.
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
end_points['block6'] = net
net = tf.layers.dropout(net, rate=dropout_keep_prob, training=is_training)
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
#end_points['block7'] = net
net7 = tf.layers.dropout(net, rate=dropout_keep_prob, training=is_training)
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'block8'
with tf.variable_scope(end_point):
net = slim.conv2d(net7, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net8 = slim.conv2d(net, 512, [3, 3], stride=2, scope='conv3x3', padding='VALID')
#end_points[end_point] = net
end_point = 'block9'
with tf.variable_scope(end_point):
net = slim.conv2d(net8, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net9 = slim.conv2d(net, 256, [3, 3], stride=2, scope='conv3x3', padding='VALID')
#end_points[end_point] = net
end_point = 'block10'
with tf.variable_scope(end_point):
net = slim.conv2d(net9, 128, [1, 1], scope='conv1x1')
net10 = slim.conv2d(net, 256, [3, 3], scope='conv3x3', padding='VALID')
#end_points[end_point] = net
end_point = 'block11'
with tf.variable_scope(end_point):
net11 = slim.conv2d(net10, 128, [1, 1], scope='conv1x1')
net11 = slim.conv2d(net11, 256, [3, 3], scope='conv3x3', padding='VALID')
end_points[end_point] = net11
#print("================net11 shape", net11.shape)
# NCHW
#################################################################################
end_point = 'block10be'
with tf.variable_scope(end_point):
net10_a = tf.transpose(net11, perm=(0, 2, 3, 1))#nchw
net10_a = tf.image.resize_nearest_neighbor(net10_a, (3,3))
net10_a = tf.transpose(net10_a, perm=(0, 3, 1, 2))#nchw
net10_a = slim.conv2d(net10_a, 256, [3,3], scope='pre10_3x3')
net10_b = slim.conv2d(net10, 256, [1, 1], scope='pre10_1x1')
net10_o = net10_a + net10_b
end_points[end_point] = net10_o #3
end_point = 'block9be'
with tf.variable_scope(end_point):
net9_a = tf.transpose(net10_o, perm=(0, 2, 3, 1))#nchw
net9_a = tf.image.resize_nearest_neighbor(net9_a, (5,5))
net9_a = tf.transpose(net9_a, perm=(0, 3, 1, 2))#nchw
net9_a = slim.conv2d(net9_a, 256, [3,3], scope='pre9_3x3')
net9_b = slim.conv2d(net9, 256, [1, 1], scope='pre9_1x1')
net9_o = net9_a + net9_b
end_points[end_point] = net9_o#5
end_point = 'block8be'
with tf.variable_scope(end_point):
net8_a = tf.transpose(net9_o, perm=(0, 2, 3, 1))#nchw
net8_a = tf.image.resize_nearest_neighbor(net8_a, (10,10))
net8_a = tf.transpose(net8_a, perm=(0, 3, 1, 2))#nchw
net8_a = slim.conv2d(net8_a, 512, [3,3], padding='SAME', scope='pre8_3x3')
net8_b = slim.conv2d(net8, 512, [1, 1], scope='pre8_1x1')#10
net8_o = net8_a + net8_b
end_points[end_point] = net8_o#10
end_point = 'block7be'
with tf.variable_scope(end_point):
net7_a = tf.transpose(net8_o, perm=(0, 2, 3, 1))#nchw
net7_a = tf.image.resize_nearest_neighbor(net7_a, (19, 19))#
net7_a = tf.transpose(net7_a, perm=(0, 3, 1, 2))#nchw
net7_a = slim.conv2d(net7_a, 1024, [3, 3], padding='SAME', scope='pre7_3x3')
net7_b = slim.conv2d(net7, 1024, [1, 1], scope='pre7_1x1')
net7_o = net7_a + net7_b
end_points[end_point] = net7_o
end_point = 'block4be'
with tf.variable_scope(end_point):
net4_a = tf.transpose(net7_o, perm=(0, 2, 3, 1))#nchw
net4_a = tf.image.resize_nearest_neighbor(net4_a, (38, 38))#
net4_a = tf.transpose(net4_a, perm=(0, 3, 1, 2))#nchw
net4_a = slim.conv2d(net4_a, 512, [3, 3], padding='SAME', scope='pre4_3x3')
net4_b = slim.conv2d(net4, 512, [1, 1], scope='pre4_1x1')
#print("asdfasdfasdf", net4_a.shape, net4_b.shape)
net4_o = net4_a + net4_b#38
end_points[end_point] = net4_o
'''
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
end_points['block1'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool1')
# Block 2.
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
end_points['block2'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool2')
# Block 3.
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
end_points['block3'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool3')
# Block 4.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
end_points['block4'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool4')
# Block 5.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
end_points['block5'] = net
net = slim.max_pool2d(net, [3, 3], stride=1, scope='pool5')
# Additional SSD blocks.
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
end_points['block6'] = net
net = tf.layers.dropout(net,
rate=dropout_keep_prob,
training=is_training)
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
end_points['block7'] = net
net = tf.layers.dropout(net,
rate=dropout_keep_prob,
training=is_training)
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'block8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block11'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
# Prediction and localisations layers.
predictions = []
logits = []
localisations = []
for i, layer in enumerate(feat_layers):
with tf.variable_scope(layer + '_box'):
p, l = ssd_multibox_layer(end_points[layer], num_classes,
anchor_sizes[i], anchor_ratios[i],
normalizations[i])
predictions.append(prediction_fn(p))
logits.append(p)
localisations.append(l)
return predictions, localisations, logits, end_points
def text_net(inputs,
feat_layers=TextboxNet.default_params.feat_layers,
anchor_sizes=TextboxNet.default_params.anchor_sizes,
anchor_ratios=TextboxNet.default_params.anchor_ratios,
normalizations=TextboxNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
reuse=None,
scope='text_box_384',
update_feat_shapes=False):
"""
Define the backbone (13 original vgg layers + 10 extra conv layers to extract multi-scale feature maps form as SSD)
of the textboxes and the neck net -- 6 textbox layers.
:param inputs: input image size
:param feat_layers: feature map which connect to the textbox layer
:param anchor_sizes: multi-scale anchor sizes
:param anchor_ratios: multi-scale anchor aspect ratios [2.0, 1. / 2, 3.0, 1. / 3, 4.0, 1. / 4, 5., 1. / 5]
:param normalizations:
:param is_training: train or not
:param dropout_keep_prob:
:param reuse:
:param scope:
:param update_feat_shapes:
:return: [predictions, localisations, logits, end_points, shape_list(if update_feat_shapes=True)]
"""
# End_points collect relevant activations for external use.
end_points = {}
with tf.variable_scope(scope, 'text_box_300', [inputs],
reuse=reuse): # 300*300 384*384
######################################
# 前五个 Blocks，首先照搬 VGG16 架构 #
# 注意这里使用 end_points 标注中间结果 #
######################################
# ——————————————————Original VGG-16 blocks (total 13 conv layers)———————————————————————
# Block 1.
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3],
scope='conv1') # 300 384
end_points['conv1'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool1') # 150
# Block 2.
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3],
scope='conv2') # 150 192
end_points['conv2'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool2') # 75
# Block 3.
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3],
scope='conv3') # 75 81
end_points['conv3'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool3') # 38
# Block 4.
end_point = 'conv4'
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3],
scope='conv4') # 38 40
end_points[end_point] = net
net = slim.max_pool2d(net, [2, 2], scope='pool4') # 19
# Block 5.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3],
scope='conv5') # 19
end_points['conv5'] = net
net = slim.max_pool2d(net, [3, 3], stride=1,
scope='pool5') # 19 Pooling size 2 -> 3
####################################
# 后六个 Blocks，使用额外卷积层 #
####################################
# ————————————Additional SSD blocks.——————————————————————
# Block 6: let's dilate the hell out of it! dilation -> 6
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6') # 19
end_points['conv6'] = net
# Block 7: 1x1 conv. Because the f**k.
end_point = 'conv7'
net = slim.conv2d(net, 1024, [1, 1], scope='conv7') # 19
end_points[end_point] = net
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'conv8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net # 10
end_point = 'conv9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net # 5
end_point = 'conv10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net # 3
end_point = 'conv11'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net #
######################################
# 每个中间层 end_points 返回中间结果 #
# 将各层预测结果存入列表，返回给优化函数 #
######################################
# Prediction and localisations layers.
predictions = []
logits = []
localisations = []
shape_list = []
# feat_layers=['conv4', 'conv7', 'conv8', 'conv9', 'conv10', 'conv11']
for i, layer in enumerate(feat_layers):
with tf.variable_scope(layer + '_box'):
cls, loc, shape = text_multibox_layer(layer, end_points[layer],
anchor_sizes[i],
anchor_ratios[i],
normalizations[i])
prediction_fn = slim.softmax
# Prediction of conference and bbox location of each textbox layer.
predictions.append(prediction_fn(cls))
logits.append(cls)
localisations.append(loc)
shape_list.append(shape)
if update_feat_shapes is True:
return predictions, localisations, logits, end_points, shape_list
else:
return predictions, localisations, logits, end_points
def ssd_net(inputs,
num_classes=SSDNet.default_params.num_classes,
feat_layers=SSDNet.default_params.feat_layers,
anchor_sizes=SSDNet.default_params.anchor_sizes,
anchor_ratios=SSDNet.default_params.anchor_ratios,
normalizations=SSDNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
prediction_fn=slim.softmax,
reuse=None,
scope='ssd_300_vgg'):
""" SSD net definition. """
end_points = {}
with tf.variable_scope(scope, 'ssd_300_vgg', [inputs], reuse=reuse):
# Original VGG-16 blocks.
# Block 1.
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
end_points['block1'] = net # shape = (batch_size, 300, 300, 64)
net = slim.max_pool2d(
net, [2, 2], scope='pool1') # shape = (batch_size, 150, 150, 64)
# Block 2.
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
end_points['block2'] = net # shape=(2, 150, 150, 128)
net = slim.max_pool2d(net, [2, 2],
scope='pool2') # shape=(2, 75, 75, 128)
# Block 3.
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
end_points['block3'] = net # shape=(2, 75, 75, 256)
net = slim.max_pool2d(
net, [2, 2],
scope='pool3') # shape=(2, 38, 38, 256), default padding='VALID'
# Block 4.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
end_points['block4'] = net # shape=(2, 38, 38, 256)
net = slim.max_pool2d(net, [2, 2],
scope='pool4') # shape=(2, 19, 19, 512)
# Block 5.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
end_points['block5'] = net # shape=(2, 19, 19, 512)
net = slim.max_pool2d(net, [3, 3], stride=1,
scope='pool5') # shape=(2, 19, 19, 512)
# Additional SSD blocks.
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
end_points['block6'] = net # shape=(2, 19, 19, 1024)
net = tf.layers.dropout(
net, rate=dropout_keep_prob,
training=is_training) # shape=(2, 16, 16, 1024)
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
end_points['block7'] = net # shape=(2, 19, 19, 1024)
net = tf.layers.dropout(
net, rate=dropout_keep_prob,
training=is_training) # shape=(2, 19, 19, 1024)
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'block8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1],
scope='conv1x1') # shape=(2, 19, 19, 256)
net = custom_layers.pad2d(net,
pad=(1, 1)) # shape=(2, 21, 21, 256)
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID') # shape=(2, 10, 10, 512)
end_points[end_point] = net
end_point = 'block9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1],
scope='conv1x1') # shape=(2, 10, 10, 128)
net = custom_layers.pad2d(net,
pad=(1, 1)) # shape=(2, 12, 12, 128)
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID') # shape=(2, 5, 5, 256)
end_points[end_point] = net
end_point = 'block10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1],
scope='conv1x1') # shape=(2, 5, 5, 128)
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID') # shape=(2, 3, 3, 256)
end_points[end_point] = net
end_point = 'block11'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1],
scope='conv1x1') # shape=(2, 3, 3, 128)
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID') # shape=(2, 1, 1, 256)
end_points[end_point] = net
# Prediction and localisations layers.
predictions = []
logits = []
localisations = []
for i, layer in enumerate(feat_layers):
with tf.variable_scope(layer + '_box'):
p, l = ssd_multibox_layer(end_points[layer], num_classes,
anchor_sizes[i], anchor_ratios[i],
normalizations[i])
predictions.append(prediction_fn(p))
logits.append(p)
localisations.append(l)
return predictions, localisations, logits, end_points
def ssd_net(inputs,
num_classes=SSDNet.default_params.num_classes,
feat_layers=SSDNet.default_params.feat_layers,
anchor_sizes=SSDNet.default_params.anchor_sizes,
anchor_ratios=SSDNet.default_params.anchor_ratios,
normalizations=SSDNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
prediction_fn=slim.softmax,
reuse=None,
scope='ssd_300_vgg'):
"""SSD net definition.
"""
# if data_format == 'NCHW':
# inputs = tf.transpose(inputs, perm=(0, 3, 1, 2))
# End_points collect relevant activations for external use.
end_points = {}
# 定义ssd网络，前半部分是vgg-16网络
with tf.variable_scope(scope, 'ssd_300_vgg', [inputs], reuse=reuse):
# Original VGG-16 blocks.
# 原始vgg-16
# 两个卷积层、一个最大池化层
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
end_points['block1'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool1')
# Block 2.
# 特征层1：block2,两个卷积层、1个最大池化层
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
end_points['block2'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool2')
# Block 3.
# 特征层2：block3,三个卷积层、1个最大池化层
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
end_points['block3'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool3')
# Block 4.
# 特征层3：block4,三个卷积层、1个最大池化层
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
end_points['block4'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool4')
# Block 5.
# 特征层4：block5,三个卷积层、1个最大池化层
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
end_points['block5'] = net
net = slim.max_pool2d(net, [3, 3], stride=1, scope='pool5')
# Additional SSD blocks.
# ssd相比起vgg多加的层
# Block 6: let's dilate the hell out of it!
# 特征层5：block6,一个卷积层、1个最大池化层，使用了dropout训练技巧
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
end_points['block6'] = net
net = tf.layers.dropout(net,
rate=dropout_keep_prob,
training=is_training)
# Block 7: 1x1 conv. Because the f**k.
# 特征层6：block7,一个卷积层(1x1卷积核，相当于在3个图像通道做叠加平均)、1个最大池化层，使用了dropout训练技巧
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
end_points['block7'] = net
net = tf.layers.dropout(net,
rate=dropout_keep_prob,
training=is_training)
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
# block8-block11都是特征层
# padding='VALID'经过filter kernel之后，尺寸可能会变小；padding='SAME'表示经过filter kernel之后，尺寸保持跟输入不变
end_point = 'block8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block11'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
# Prediction and localisations layers.
# 预测层与定位层
predictions = []
logits = []
localisations = []
for i, layer in enumerate(feat_layers):
with tf.variable_scope(layer + '_box'):
p, l = ssd_multibox_layer(end_points[layer], num_classes,
anchor_sizes[i], anchor_ratios[i],
normalizations[i])
predictions.append(prediction_fn(p))
logits.append(p)
localisations.append(l)
return predictions, localisations, logits, end_points
def ssd_net(inputs,
num_classes=SSDNet.default_params.num_classes,
feat_layers=SSDNet.default_params.feat_layers,
anchor_sizes=SSDNet.default_params.anchor_sizes,
anchor_ratios=SSDNet.default_params.anchor_ratios,
normalizations=SSDNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
prediction_fn=slim.softmax,
reuse=None,
scope='ssd_512_vgg'):
"""SSD net definition.
"""
# End_points collect relevant activations for external use.
end_points = {}
with tf.variable_scope(scope, 'ssd_512_vgg', [inputs], reuse=reuse):
# Original VGG-16 blocks.
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
end_points['block1'] = net
# print("====net.shape = {}".format(net))
net = slim.max_pool2d(net, [2, 2], scope='pool1')
# Block 2.
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
end_points['block2'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool2')
# Block 3.txt.
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
end_points['block3'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool3')
# Block 4.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
end_points['block4'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool4')
# Block 5.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
end_points['block5'] = net
net = slim.max_pool2d(net, [3, 3], 1, scope='pool5')
# Additional SSD blocks.
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
end_points['block6'] = net
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
end_points['block7'] = net
if IS_FPN:
stride_b8_to_b12 = 1
else:
stride_b8_to_b12 = 2
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'block8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=stride_b8_to_b12,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=stride_b8_to_b12,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block11'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=stride_b8_to_b12,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'concat'
with tf.variable_scope(end_point):
high_feature_list = []
for i in range(9, 12):
high_feature_list.append(end_points['block' + str(i)])
high_feature = tf.concat(high_feature_list, 1)
end_points[end_point] = high_feature
end_point = 'block4_concat'
with tf.variable_scope(end_point):
net_block4 = slim.conv2d(end_points['concat'],
128, [1, 1],
scope='conv1x1')
# print("net_block4 ========{}".format(net_block4))
net_block4 = tf.layers.conv2d_transpose(
net_block4,
512,
kernel_size=3,
strides=(4, 4),
data_format='channels_first')
# print("net_block4 ========{}".format(net_block4))
end_points['block4'] += net_block4
end_point = 'block7_concat'
with tf.variable_scope(end_point):
net_block7 = slim.conv2d(end_points['concat'],
128, [1, 1],
scope='conv1x1')
# print("net_block4 ========{}".format(net_block4))
net_block7 = tf.layers.conv2d_transpose(
net_block7,
1024,
kernel_size=1,
strides=(2, 2),
data_format='channels_first')
# print("net_block4 ========{}".format(net_block4))
end_points['block7'] += net_block7
end_point = 'block8_concat'
with tf.variable_scope(end_point):
net_block8 = slim.conv2d(end_points['concat'],
512, [1, 1],
scope='conv1x1')
end_points['block8'] += net_block8
# with tf.variable_scope(end_point):
# print("=debug===========high_feature = {}".format(high_feature))
# for i in range(1,12):
# print("=debug===== "+str(i)+" = {}".format(end_points['block'+str(i)]))
# Prediction and localisations layers.
predictions_pest = []
logits_pest = []
predictions = []
logits = []
localisations = []
for i, layer in enumerate(feat_layers):
with tf.variable_scope(layer + '_box'):
# print("=debug===== "+str(i)+" = {}".format(end_points[layer]))
# 这里获取每一个anchor的预测值
pest, p, l = ssd_vgg_300.ssd_multibox_layer(
end_points[layer], num_classes, anchor_sizes[i],
anchor_ratios[i], normalizations[i])
predictions_pest.append(prediction_fn(pest))
logits_pest.append(pest)
predictions.append(prediction_fn(p))
logits.append(p)
localisations.append(l)
return predictions_pest, logits_pest, predictions, localisations, logits, end_points
def ssd_net(inputs,
num_classes=SSDNet.default_params.num_classes,
feat_layers=SSDNet.default_params.feat_layers,
anchor_sizes=SSDNet.default_params.anchor_sizes,
anchor_ratios=SSDNet.default_params.anchor_ratios,
normalizations=SSDNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
prediction_fn=slim.softmax,
reuse=None,
scope='ssd_512_vgg',
DSSD_FLAG=False):
"""SSD net definition.
"""
# End_points collect relevant activations for external use.
end_points = {}
if inputs.shape[2] == inputs.shape[3]:
mode = 'bnwh'
else:
mode = 'bwhn'
with tf.variable_scope(scope, 'ssd_512_vgg', [inputs], reuse=reuse):
# Original VGG-16 blocks.
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
end_points['block1'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool1')
# Block 2.
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
end_points['block2'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool2')
# Block 3.
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
end_points['block3'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool3')
# Block 4.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
end_points['block4'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool4')
# Block 5.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
end_points['block5'] = net
net = slim.max_pool2d(net, [3, 3], 1, scope='pool5')
# Additional SSD blocks.
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
end_points['block6'] = net
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
end_points['block7'] = net
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'block8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block11'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block12'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [4, 4],
scope='conv4x4',
padding='VALID')
# Fix padding to match Caffe version (pad=1).
# pad_shape = [(i-j) for i, j in zip(layer_shape(net), [0, 1, 1, 0])]
# net = tf.slice(net, [0, 0, 0, 0], pad_shape, name='caffe_pad')
end_points[end_point] = net
# Prediction and localisations layers.
# rever_feat_layers = list(reversed(feat_layers))
# for i, l in enumerate(rever_feat_layers):
# if i == 0: continue
# l_ = rever_feat_layers[i - 1]
#
# end_points[l] = tf.concat([upbilinear([end_points[l_], end_points[l]], name=l_), end_points[l]],axis=1)
# with tf.variable_scope("fpn11"):
#
#
# end_points['block11'] = tf.add(upbilinear([end_points['block12'], end_points['block11']],name ="up_11",mode=mode),
# end_points['block11'],name= "fpn_block11")
#
# with tf.variable_scope("fpn10"):
#
# end_points['block10'] = tf.add(upbilinear([end_points['block11'],end_points['block10']],name ="up_10",mode=mode),
# end_points['block10'],
# name="fpn_block10")
# with tf.variable_scope("fpn9"):
# b9 = slim.conv2d(end_points['block10'], 256, [1, 1], scope='9conv1x1')
# b9_ = slim.conv2d(end_points['block9'], 256, [1, 1], scope='9_conv1x1')
# end_points['block9'] = tf.add(upbilinear([end_points['block10'],end_points['block9']],name ="up_9",mode=mode),
# end_points['block9'],
# name="fpn_block9")
# with tf.variable_scope("fpn8"):
# b8 = slim.conv2d(end_points['block9'], 512, [1, 1], scope='8conv1x1')
#
# end_points['block8'] = tf.add(upbilinear([b8, end_points['block8']],name ="up_8",mode=mode), end_points['block8'],
# name="fpn_block8")
# with tf.variable_scope("fpn7"):
# b7 = slim.conv2d(end_points['block8'], 1024, [1, 1], scope='7conv1x1')
#
# end_points['block7'] = tf.add(upbilinear([b7, end_points['block7']],name ="up_7",mode=mode), end_points['block7'],
# name="fpn_block7")
# with tf.variable_scope("fpn4"):
# b4 = slim.conv2d(end_points['block7'], 512, [1, 1], scope='4conv1x1')
#
# end_points['block4'] = tf.add(upbilinear([b4, end_points['block4']],name ="up_4",mode=mode), end_points['block4'],
# name="fpn_block4")
if DSSD_FLAG:
with tf.variable_scope("dssd11"):
de_12 = slim.conv2d_transpose(end_points['block12'],
512, [3, 3],
stride=2,
scope="de_12")
con_12 = slim.conv2d(de_12, 512, [3, 3], scope='conv_12')
bn_12 = slim.batch_norm(con_12, is_training=is_training)
con_11 = slim.conv2d(end_points["block11"],
512, [3, 3],
scope="conv11")
bn_11 = slim.batch_norm(con_11, is_training=is_training)
relu_11 = tf.nn.relu(bn_11)
con_11 = slim.conv2d(relu_11, 512, [3, 3], scope="conv11_2")
bn_11 = slim.batch_norm(con_11, is_training=is_training)
end_points["block11"] = tf.nn.relu(tf.multiply(bn_12, bn_11))
with tf.variable_scope("dssd10"):
de_11 = slim.conv2d_transpose(end_points['block11'],
512, [3, 3],
stride=2,
scope="de_11")
con_11 = slim.conv2d(de_11, 512, [3, 3], scope='conv_11')
bn_11 = slim.batch_norm(con_11, is_training=is_training)
con_10 = slim.conv2d(end_points["block10"],
512, [3, 3],
scope="conv10")
bn_10 = slim.batch_norm(con_10, is_training=is_training)
relu_10 = tf.nn.relu(bn_10)
con_10 = slim.conv2d(relu_10, 512, [3, 3], scope="conv10_2")
bn_10 = slim.batch_norm(con_10, is_training=is_training)
end_points["block10"] = tf.nn.relu(tf.multiply(bn_11, bn_10))
with tf.variable_scope("dssd9"):
de_10 = slim.conv2d_transpose(end_points['block10'],
512, [3, 3],
stride=2,
scope="de_10")
con_10 = slim.conv2d(de_10, 512, [3, 3], scope='conv_10')
bn_10 = slim.batch_norm(con_10, is_training=is_training)
con_9 = slim.conv2d(end_points["block9"],
512, [3, 3],
scope="conv9")
bn_9 = slim.batch_norm(con_9, is_training=is_training)
relu_9 = tf.nn.relu(bn_9)
con_9 = slim.conv2d(relu_9, 512, [3, 3], scope="conv9_2")
bn_9 = slim.batch_norm(con_9, is_training=is_training)
end_points["block9"] = tf.nn.relu(tf.multiply(bn_10, bn_9))
with tf.variable_scope("dssd8"):
de_9 = slim.conv2d_transpose(end_points['block9'],
512, [3, 3],
stride=2,
scope="de_9")
con_9 = slim.conv2d(de_9, 512, [3, 3], scope='conv_9')
bn_9 = slim.batch_norm(con_9, is_training=is_training)
con_8 = slim.conv2d(end_points["block8"],
512, [3, 3],
scope="conv8")
bn_8 = slim.batch_norm(con_8, is_training=is_training)
relu_8 = tf.nn.relu(bn_8)
con_8 = slim.conv2d(relu_8, 512, [3, 3], scope="conv8_2")
bn_8 = slim.batch_norm(con_8, is_training=is_training)
end_points["block8"] = tf.nn.relu(tf.multiply(bn_9, bn_8))
with tf.variable_scope("dssd7"):
de_8 = slim.conv2d_transpose(end_points['block8'],
512, [3, 3],
stride=2,
scope="de_8")
con_8 = slim.conv2d(de_8, 512, [3, 3], scope='conv_8')
bn_8 = slim.batch_norm(con_8, is_training=is_training)
con_7 = slim.conv2d(end_points["block7"],
512, [3, 3],
scope="conv7")
bn_7 = slim.batch_norm(con_7, is_training=is_training)
relu_7 = tf.nn.relu(bn_7)
con_7 = slim.conv2d(relu_7, 512, [3, 3], scope="conv7_2")
bn_7 = slim.batch_norm(con_7, is_training=is_training)
end_points["block7"] = tf.nn.relu(tf.multiply(bn_8, bn_7))
with tf.variable_scope("dssd4"):
de_7 = slim.conv2d_transpose(end_points['block7'],
512, [3, 3],
stride=2,
scope="de_7")
con_7 = slim.conv2d(de_7, 512, [3, 3], scope='conv_7')
bn_7 = slim.batch_norm(con_7, is_training=is_training)
con_4 = slim.conv2d(end_points["block4"],
512, [3, 3],
scope="conv4")
bn_4 = slim.batch_norm(con_4, is_training=is_training)
relu_4 = tf.nn.relu(bn_4)
con_4 = slim.conv2d(relu_4, 512, [3, 3], scope="conv4_2")
bn_4 = slim.batch_norm(con_4, is_training=is_training)
end_points["block4"] = tf.nn.relu(tf.multiply(bn_7, bn_4))
#
predictions = []
logits = []
localisations = []
for i, layer in enumerate(feat_layers):
with tf.variable_scope(layer + '_box'):
p, l = ssd_vgg_300.ssd_multibox_layer(end_points[layer],
num_classes,
anchor_sizes[i],
anchor_ratios[i],
normalizations[i])
predictions.append(prediction_fn(p))
logits.append(p)
localisations.append(l)
return predictions, localisations, logits, end_points
def ssd_net(inputs,
num_classes=SSDNet.default_params.num_classes,
feat_layers=SSDNet.default_params.feat_layers,
anchor_sizes=SSDNet.default_params.anchor_sizes,
anchor_ratios=SSDNet.default_params.anchor_ratios,
normalizations=SSDNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
prediction_fn=slim.softmax,
reuse=None,
scope='ssd_300_vgg'):
"""SSD net definition.
"""
# if data_format == 'NCHW':
# inputs = tf.transpose(inputs, perm=(0, 3, 1, 2))
# End_points collect relevant activations for external use.
"""
net = layers_lib.repeat(
inputs, 2, layers.conv2d, 64, [3, 3], scope='conv1')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool1')
net = layers_lib.repeat(net, 2, layers.conv2d, 128, [3, 3], scope='conv2')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool2')
net = layers_lib.repeat(net, 3, layers.conv2d, 256, [3, 3], scope='conv3')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool3')
net = layers_lib.repeat(net, 3, layers.conv2d, 512, [3, 3], scope='conv4')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool4')
net = layers_lib.repeat(net, 3, layers.conv2d, 512, [3, 3], scope='conv5')
net = layers_lib.max_pool2d(net, [2, 2], scope='pool5')
"""
end_points = {}
with tf.variable_scope(scope, 'ssd_300_vgg', [inputs], reuse=reuse):
######################################
# 前五个 Blocks，首先照搬 VGG16 架构 #
# 注意这里使用 end_points 标注中间结果 #
######################################
# ——————————————————Original VGG-16 blocks.———————————————————————
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
end_points['block1'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool1')
# Block 2.
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
end_points['block2'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool2')
# Block 3.
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
end_points['block3'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool3')
# Block 4.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
end_points['block4'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool4')
# Block 5.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
end_points['block5'] = net
net = slim.max_pool2d(net, [3, 3], stride=1,
scope='pool5') # 池化层由2修改到三
####################################
# 后六个 Blocks，使用额外卷积层 #
####################################
# ————————————Additional SSD blocks.——————————————————————
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
end_points['block6'] = net
net = tf.layers.dropout(net,
rate=dropout_keep_prob,
training=is_training)
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
end_points['block7'] = net
net = tf.layers.dropout(net,
rate=dropout_keep_prob,
training=is_training)
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'block8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block11'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
######################################
# 每个中间层 end_points 返回中间结果 #
# 将各层预测结果存入列表，返回给优化函数 #
######################################
# Prediction and localisations layers.
predictions = []
logits = []
localisations = []
# feat_layers=['block4', 'block7', 'block8', 'block9', 'block10', 'block11']
for i, layer in enumerate(feat_layers):
with tf.variable_scope(layer + '_box'):
p, l = ssd_multibox_layer(end_points[layer], num_classes,
anchor_sizes[i], anchor_ratios[i],
normalizations[i])
"""
框的数目等于anchor_sizes[i]和anchor_ratios[i]的长度和
anchor_sizes=[(21., 45.),
(45., 99.),
(99., 153.),
(153., 207.),
(207., 261.),
(261., 315.)]
anchor_ratios=[[2, .5],
[2, .5, 3, 1./3],
[2, .5, 3, 1./3],
[2, .5, 3, 1./3],
[2, .5],
[2, .5]]
normalizations=[20, -1, -1, -1, -1, -1]
"""
predictions.append(prediction_fn(p))
logits.append(p)
localisations.append(l)
return predictions, localisations, logits, end_points
def ssd_net(inputs,
num_classes=SSDNet.default_params.num_classes,
feat_layers=SSDNet.default_params.feat_layers,
anchor_sizes=SSDNet.default_params.anchor_sizes,
anchor_ratios=SSDNet.default_params.anchor_ratios,
normalizations=SSDNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
prediction_fn=slim.softmax,
reuse=None,
scope='ssd_300_vgg',
large=True,
medium=True):
"""SSD net definition.
"""
# if data_format == 'NCHW':
# inputs = tf.transpose(inputs, perm=(0, 3, 1, 2))
# End_points collect relevant activations for external use.
end_points = {}
with tf.variable_scope(scope, 'ssd_300_vgg', [inputs], reuse=reuse):
# Original VGG-16 blocks.
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
end_points['block1'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool1')
# Block 2.
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
end_points['block2'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool2')
# Block 3.
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
end_points['block3'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool3')
# Block 4.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
print(net)
end_points['block4'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool4')
# Block 5.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
end_points['block5'] = net
net = slim.max_pool2d(net, [3, 3], stride=1, scope='pool5')
# Additional SSD blocks.
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
end_points['block6'] = net
net = tf.layers.dropout(net,
rate=dropout_keep_prob,
training=is_training)
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
print(net)
end_points['block7'] = net
net = tf.layers.dropout(net,
rate=dropout_keep_prob,
training=is_training)
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'block8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
print(net)
end_point = 'block9'
end_points = tf.cond(
large, lambda: largenetwork(net, end_point, end_points),
lambda: tf.cond(medium, lambda: mediumnetwork(
net, end_point, end_points), lambda: foveanetwork(
net, end_point, end_points)))
# Prediction and localisations layers.
#print(feat_layers)
#find_predictions_large(end_points,num_classes,anchor_sizes,anchor_ratios,normalizations,prediction_fn)
#find_predictions_fovea(end_points,num_classes,anchor_sizes,anchor_ratios,normalizations,prediction_fn)
predictions, localisations, logits = tf.cond(
large, lambda: find_predictions_large(
end_points, num_classes, anchor_sizes, anchor_ratios,
normalizations, prediction_fn), lambda: tf.cond(
medium, lambda: find_predictions_medium(
end_points, num_classes, anchor_sizes, anchor_ratios,
normalizations,
prediction_fn), lambda: find_predictions_fovea(
end_points, num_classes, anchor_sizes,
anchor_ratios, normalizations, prediction_fn)))
return predictions, localisations, logits, end_points
def ssd_net(inputs,
num_classes=SSDNet.default_params.num_classes,
feat_layers=SSDNet.default_params.feat_layers,
anchor_sizes=SSDNet.default_params.anchor_sizes,
anchor_ratios=SSDNet.default_params.anchor_ratios,
normalizations=SSDNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
prediction_fn=slim.softmax,
reuse=None,
scope='ssd_300_vgg'):
"""SSD net definition.
"""
#if data_format == 'NCHW':
#inputs = tf.transpose(inputs, perm=(0, 3, 1, 2))
# End_points collect relevant activations for external use.
end_points = {}
with tf.variable_scope(scope, 'ssd_300_vgg', [inputs], reuse=reuse):
# Original VGG-16 blocks.
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
end_points['block1'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool1')
# Block 2.
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
end_points['block2'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool2')
# Block 3.
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
end_points['block3'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool3')
# Block 4.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
end_points['block4'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool4')
# transform net4 to net7
net4_net7 = net
net4_net7 = slim.conv2d(net4_net7, 1024, [1, 1], scope='net4_net7')
# Block 5.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
end_points['block5'] = net
net = slim.max_pool2d(net, [3, 3], stride=1, scope='pool5')
# Additional SSD blocks.
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
end_points['block6'] = net
net = tf.layers.dropout(net,
rate=dropout_keep_prob,
training=is_training)
# Block 7: 1x1 conv.
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
end_points['block7'] = net #19*19*1024
net7_net4 = tensor_resize.tensor4_resize(net, [38, 38])
net7_net4 = slim.conv2d(net7_net4, 512, [1, 1], scope='net7_net4')
# # merge block4 to block7
end_points['block7'] = end_points['block7'] + net4_net7
# # Transform net7 to net8
net7_net8 = end_points['block7']
net7_net8 = slim.max_pool2d(net7_net8, [3, 3],
stride=2,
scope='net7_net8_mp')
net7_net8 = slim.conv2d(net7_net8, 512, [1, 1], scope='net7_net8')
net = tf.layers.dropout(net,
rate=dropout_keep_prob,
training=is_training)
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'block8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
net8_net7 = tensor_resize.tensor4_resize(net, [19, 19])
net8_net7 = slim.conv2d(net8_net7, 1024, [1, 1], scope='net8_net7')
# merge block7 to block8
end_points[end_point] = end_points[end_point] + net7_net8
# # Transform net8 to net9
net8_net9 = end_points[end_point]
net8_net9 = slim.max_pool2d(net8_net9, [3, 3],
stride=2,
scope='net8_net9_mp')
net8_net9 = slim.conv2d(net8_net9, 256, [1, 1], scope='net8_net9')
end_point = 'block9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
net9_net8 = tensor_resize.tensor4_resize(net, [10, 10])
net9_net8 = slim.conv2d(net9_net8, 512, [1, 1], scope='net9_net8')
# merge block8 to block9
end_points[end_point] = end_points[end_point] + net8_net9
# Transform net9 to net10
net9_net10 = end_points[end_point]
net9_net10 = slim.max_pool2d(net9_net10, [3, 3],
stride=2,
scope='net9_net10_mp')
net9_net10 = slim.conv2d(net9_net10, 256, [1, 1], scope='net9_net10')
end_point = 'block10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net #3*3*256
net10_net9 = tensor_resize.tensor4_resize(net, [5, 5])
net10_net9 = slim.conv2d(net10_net9, 256, [1, 1], scope='net10_net9')
# # merge block9 to block10
end_points[end_point] = end_points[end_point] + net9_net10
# # Transform net10 to net11
net10_net11 = end_points[end_point]
net10_net11 = slim.max_pool2d(net10_net11, [3, 3],
stride=2,
scope='net10_net11_mp',
padding='VALID')
net10_net11 = slim.conv2d(net10_net11,
256, [1, 1],
scope='net10_net11')
end_point = 'block11'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net #1*1*256
#upsample net11 and Transform net11 to net10 !
net11_net10 = tensor_resize.tensor4_resize(net, [3, 3])
net11_net10 = slim.conv2d(net11_net10,
256, [1, 1],
scope='net11_net10')
# # merge block10 to block11
end_points[end_point] = end_points[end_point] + net10_net11
end_points['block4'] = end_points['block4'] + net7_net4
end_points['block7'] = end_points['block7'] + net8_net7
end_points['block8'] = end_points['block8'] + net9_net8
end_points['block9'] = end_points['block9'] + net10_net9
end_points['block10'] = end_points['block10'] + net11_net10
# Prediction and localisations layers.
predictions = []
logits = []
localisations = []
for i, layer in enumerate(feat_layers):
with tf.variable_scope(layer + '_box'):
p, l = ssd_multibox_layer(end_points[layer], num_classes,
anchor_sizes[i], anchor_ratios[i],
normalizations[i])
predictions.append(prediction_fn(p))
logits.append(p)
localisations.append(l)
return predictions, localisations, logits, end_points
def ssd_net(inputs,
num_classes,
feature_layers,
anchor_sizes,
anchor_ratios,
normalizations,
is_training=True,
dropout_keep_prob=0.5,
prediction_fn=slim.softmax,
reuse=None,
scope='ssd_vgg'):
#structure of SSD net
outputs = {}
with tf.variable_scope(scope, 'ssd_vgg', [inputs], reuse=reuse):
# Structure of vgg16
# Block1
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
outputs['block1'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool1')
# Block 2
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
outputs['block2'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool2')
# Block 3
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
outputs['block3'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool3')
# Block 4
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
outputs['block4'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool4')
# Block 5
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
outputs['block5'] = net
net = slim.max_pool2d(net, [3, 3], 1, scope='pool5')
# Additional SSD blocks
# Block 6
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
outputs['block6'] = net
net = tf.layers.dropout(net,
rate=dropout_keep_prob,
training=is_training)
# block 7
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
outputs['block7'] = net
net = tf.layers.dropout(net,
rate=dropout_keep_prob,
training=is_training)
# Block8
with tf.variable_scope('block8'):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
outputs['block8'] = net
# Block 9
with tf.variable_scope('block9'):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
outputs['block9'] = net
# Block 10
with tf.variable_scope('block10'):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
outputs['block10'] = net
# Block 11
with tf.variable_scope('block11'):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
outputs['block11'] = net
# Block 12
with tf.variable_scope('block12'):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [4, 4],
stride=2,
scope='conv4x4',
padding='VALID')
outputs['block12'] = net
# Prediction and locolization
predictions = [] # Class prediction
logits = [] # Probability of class
locations = [] # Location prediction
for i, layer in enumerate(feature_layers): # Block 4,7,8,9,10,11,12
with tf.variable_scope(layer + '_box'):
p, l = ssd_multibox_layers(outputs[layer], num_classes,
anchor_sizes[i], anchor_ratios[i],
normalizations[i])
predictions.append(
prediction_fn(p)) # Here use softmax to predict classs
logits.append(p)
locations.append(l)
return predictions, locations, logits, outputs
def text_net(inputs,
feat_layers=default_params.feat_layers,
normalizations=default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
scope='vgg_16'): # checked
feature_layers = {}
with tf.variable_scope(scope, 'vgg_16', [inputs], reuse=None):
# Original VGG-16 blocks.
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
net = slim.max_pool2d(net, [2, 2], scope='pool1')
# Block 2.
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
net = slim.max_pool2d(net, [2, 2], scope='pool2')
# Block 3.
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
net = slim.max_pool2d(net, [2, 2], scope='pool3')
# Block 4.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
feature_layers['conv4'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool4')
# Block 5.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
net = slim.max_pool2d(net, [3, 3], stride=1, scope='pool5')
# Additional SSD blocks.
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
feature_layers['conv7'] = net
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'conv8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
feature_layers['conv8'] = net
end_point = 'conv9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
feature_layers['conv9'] = net
end_point = 'conv10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
feature_layers['conv10'] = net
end_point = 'conv11'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
feature_layers['conv11'] = net
localisations, logits = text_detect_net(feature_layers, feat_layers)
return localisations, logits, feature_layers
def ssd_net(
inputs, #定义ssd网络结构
num_classes=SSDNet.default_params.num_classes, #分类数
feat_layers=SSDNet.default_params.feat_layers, #特征层
anchor_sizes=SSDNet.default_params.anchor_sizes,
anchor_ratios=SSDNet.default_params.anchor_ratios,
normalizations=SSDNet.default_params.normalizations, #正则化
is_training=True,
dropout_keep_prob=0.5,
prediction_fn=slim.softmax,
reuse=None,
scope='ssd_300_vgg'):
"""SSD net definition.
"""
# if data_format == 'NCHW':
# inputs = tf.transpose(inputs, perm=(0, 3, 1, 2))
# End_points collect relevant activations for external use.
end_points = {} #用于收集每一层输出结果
with tf.variable_scope(scope, 'ssd_300_vgg', [inputs], reuse=reuse):
# Original VGG-16 blocks.
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3],
scope='conv1') #VGG16网络的第一个conv，重复2次卷积，核为3x3,64个特征
end_points['block1'] = net #conv1_2结果存入end_points，name='block1'
net = slim.max_pool2d(net, [2, 2], scope='pool1')
# Block 2.
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3],
scope='conv2') #重复2次卷积，核为3x3,128个特征
end_points['block2'] = net #conv2_2结果存入end_points，name='block2'
net = slim.max_pool2d(net, [2, 2], scope='pool2')
# Block 3.
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3],
scope='conv3') #重复3次卷积，核为3x3,256个特征
end_points['block3'] = net #conv3_3结果存入end_points，name='block3'
net = slim.max_pool2d(net, [2, 2], scope='pool3')
# Block 4.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3],
scope='conv4') #重复3次卷积，核为3x3,512个特征
end_points['block4'] = net #conv4_3结果存入end_points，name='block4'
net = slim.max_pool2d(net, [2, 2], scope='pool4')
# Block 5.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3],
scope='conv5') #重复3次卷积，核为3x3,512个特征
end_points['block5'] = net #conv5_3结果存入end_points，name='block5'
net = slim.max_pool2d(net, [3, 3], stride=1, scope='pool5')
# Additional SSD blocks. #去掉了VGG的全连接层
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6,
scope='conv6') #将VGG基础网络最后的池化层结果做扩展卷积（带孔卷积）；
end_points['block6'] = net #conv6结果存入end_points，name='block6'
net = tf.layers.dropout(net,
rate=dropout_keep_prob,
training=is_training) #dropout层
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(
net, 1024, [1, 1],
scope='conv7') #将dropout后的网络做1x1卷积，输出1024特征，name='block7'
end_points['block7'] = net
net = tf.layers.dropout(net,
rate=dropout_keep_prob,
training=is_training) #将卷积后的网络继续做dropout
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'block8'
with tf.variable_scope(end_point):
net = slim.conv2d(
net, 256, [1, 1], scope='conv1x1'
) #对上述dropout的网络做1x1卷积，然后做3x3卷积，,输出512特征图，name=‘block8’
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block9'
with tf.variable_scope(end_point):
net = slim.conv2d(
net, 128, [1, 1],
scope='conv1x1') #对上述网络做1x1卷积，然后做3x3卷积，输出256特征图，name=‘block9’
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block10'
with tf.variable_scope(end_point):
net = slim.conv2d(
net, 128, [1, 1],
scope='conv1x1') #对上述网络做1x1卷积，然后做3x3卷积，输出256特征图，name=‘block10’
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block11'
with tf.variable_scope(end_point):
net = slim.conv2d(
net, 128, [1, 1],
scope='conv1x1') #对上述网络做1x1卷积，然后做3x3卷积，输出256特征图，name=‘block11’
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
# Prediction and localisations layers. #预测和定位
predictions = []
logits = []
localisations = []
for i, layer in enumerate(feat_layers): #遍历特征层
with tf.variable_scope(layer + '_box'): #起个命名范围
p, l = ssd_multibox_layer(
end_points[
layer], #做多尺度大小box预测的特征层，返回每个cell中每个先验框预测的类别p和预测的位置l
num_classes, #种类数
anchor_sizes[i], #先验框尺度（同一特征图上的先验框尺度和长宽比一致）
anchor_ratios[i], #先验框长宽比
normalizations[i]) #每个特征正则化信息，目前是只对第一个特征图做归一化操作；
#把每一层的预测收集
predictions.append(prediction_fn(p)) #prediction_fn为softmax，预测类别
logits.append(p) #把每个cell每个先验框预测的类别的概率值存在logits中
localisations.append(l) #预测位置信息
return predictions, localisations, logits, end_points #返回类别预测结果，位置预测结果，所属某个类别的概率值，以及特征层
def ssd_net(inputs,
num_classes,
feat_layers,
normalizations,
is_training,
dropout_keep_prob,
prediction_fn,
reuse,
scope):
# End_points collect relevant activations for external use.
end_points = {}
with tf.variable_scope(scope, 'ssd_640_vgg', [inputs], reuse=reuse):
# Original VGG-16 blocks.
print("nnnn-block1 begin")
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='r2_crcr1')
end_points['block1'] = net
print("uuuu-block1 end")
print("nnnn-block2 begin")
net = slim.max_pool2d(net, [2, 2], scope='bbpool1')
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='r2_crcr2')
end_points['block2'] = net
print("uuuu-block2 end")
print("nnnn-block3 begin")
net = slim.max_pool2d(net, [2, 2], scope='ddpool2')
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='r3_crcr3')
end_points['block3'] = net
print("uuuu-block3 end")
print("nnnn-block4 begin")
net = slim.max_pool2d(net, [2, 2], scope='ffpool3')
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='r3_crcr4')
end_points['block4'] = net
print("uuuu-block4 end")
print("nnnn-block5 begin")
net = slim.max_pool2d(net, [2, 2], scope='hhpool4')
# rate as `[dilation]`/`pad` in prototxt?, if is `[dilation]` then set rate=1
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], rate=1, scope='r3_crcr5')
end_points['block5'] = net
print("uuuu-block5 end")
print("nnnn-block6 begin")
# pool5: kernel_size: 3->2, stride: 1->2, +pad:1,, where to put `pad:1`?
net = slim.max_pool2d(net, [2, 2], stride=2, scope='jjpool5')
net = slim.conv2d(net, 1024, [3, 3], rate=1, scope='kkfc6')
end_points['block6'] = net
print("uuuu-block6 end")
print("nnnn-block7 begin")
net = tf.layers.dropout(net, rate=dropout_keep_prob, training=is_training)
net = slim.conv2d(net, 1024, [1, 1], scope='llfc7')
end_points['block7'] = net
print("uuuu-block7 end")
print("nnnn-block8 begin")
# conv61->conv62
net = tf.layers.dropout(net, rate=dropout_keep_prob, training=is_training)
end_point = 'block8'
with tf.variable_scope(end_point):
# paper: 1x1x128
net = slim.conv2d(net, 256, [1, 1], scope='mmconv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
# paper: 3x3x512-s2
net = slim.conv2d(net, 512, [3, 3], stride=2, scope='nnconv3x3', padding='VALID')
end_points[end_point] = net
print("uuuu-block8 end")
print("nnnn-block9 begin")
end_point = 'block9'
# conv71->conv72
with tf.variable_scope(end_point):
# paper: 1x1x128
net = slim.conv2d(net, 128, [1, 1], scope='ooconv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
# paper: 3x3x256-s2
net = slim.conv2d(net, 256, [3, 3], stride=2, scope='ppconv3x3', padding='VALID')
end_points[end_point] = net
print("uuuu-block9 end")
# Prediction and localisations layers.
predictions = []
logits = []
localisations = []
addn = 1
for i, layer in enumerate(feat_layers):
with tf.variable_scope(layer + '_box'):
print("nnnn-begin process----" + layer + '_box')
p, l = MultiboxLayer(addn,
end_points[layer],
num_classes,
normalizations[i])
addn = 0
print("uuuu-end process----" + layer + '_box')
predictions.append(prediction_fn(p))
logits.append(p)
localisations.append(l)
print("[final end]")
return predictions, localisations, logits, end_points
def __additional_ssd_block(self, end_points, net):
# Additional SSD blocks.
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
net = slim.batch_norm(net)
net = self.__dropout(net)
end_points['block6'] = net
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
net = slim.batch_norm(net)
net = self.__dropout(net)
end_points['block7'] = net
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'block8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = slim.batch_norm(net)
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
net = slim.batch_norm(net)
end_points[end_point] = net
end_point = 'block9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.batch_norm(net)
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
net = slim.batch_norm(net)
end_points[end_point] = net
end_point = 'block10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.batch_norm(net)
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
net = slim.batch_norm(net)
end_points[end_point] = net
end_point = 'block11'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.batch_norm(net)
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
net = slim.batch_norm(net)
end_points[end_point] = net
end_point = 'block12'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.batch_norm(net)
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [4, 4],
scope='conv4x4',
padding='VALID')
net = slim.batch_norm(net)
end_points[end_point] = net
# Prediction and localisations layers.
predictions = []
logits = []
localisations = []
for i, layer in enumerate(self.feat_layers):
with tf.variable_scope(layer + '_box'):
p, l = self.ssd_multibox_layer(end_points[layer],
self.num_classes,
self.anchor_sizes[i],
self.anchor_ratios[i],
self.normalizations[i])
predictions.append(slim.softmax(p))
logits.append(p)
localisations.append(l)
return predictions, localisations, logits, end_points
def ssd_net(inputs,
num_classes=SSDNet.default_params.num_classes,
feat_layers=SSDNet.default_params.feat_layers,
anchor_sizes=SSDNet.default_params.anchor_sizes,
anchor_ratios=SSDNet.default_params.anchor_ratios,
normalizations=SSDNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
prediction_fn=slim.softmax,
reuse=None,
scope='ssd_300_vgg'):
"""
SSD net definition.
"""
# End_points collect relevant activations for external use.
end_points = {}
with tf.variable_scope(scope, 'ssd_300_vgg', [inputs], reuse=reuse):
# 基础 VGG-16 blocks.
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
end_points['block1'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool1') # 150*150*64
# Block 2.
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
end_points['block2'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool2') # 75*75*128
# Block 3.
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
end_points['block3'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool3') # 38*38*256
# Block 4.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
end_points['block4'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool4') # 19*19*512
# Block 5.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
end_points['block5'] = net
net = slim.max_pool2d(net, [3, 3], stride=1,
scope='pool5') # 19*19*512
# 添加的 SSD blocks.
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
end_points['block6'] = net
net = tf.layers.dropout(net,
rate=dropout_keep_prob,
training=is_training)
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
end_points['block7'] = net
net = tf.layers.dropout(net,
rate=dropout_keep_prob,
training=is_training)
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'block8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block11'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = slim.conv2d(net,
256, [3, 3],
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
# Prediction and localisations layers.
predictions = []
logits = []
localisations = []
# 每一层特征图的预测
for i, layer in enumerate(feat_layers):
with tf.variable_scope(layer + '_box'):
# 特征图的需要框的点数*每个点的框数
# 每一层特征图框的大小和框的变化已经定好了
pred, loc = ssd_multibox_layer(end_points[layer], num_classes,
anchor_sizes[i],
anchor_ratios[i],
normalizations[i])
predictions.append(prediction_fn(pred))
logits.append(pred)
localisations.append(loc)
return predictions, localisations, logits, end_points
pass
def ssd_net(inputs,
num_classes=SSDNet.default_params.num_classes,
feat_layers=SSDNet.default_params.feat_layers,
anchor_sizes=SSDNet.default_params.anchor_sizes,
anchor_ratios=SSDNet.default_params.anchor_ratios,
normalizations=SSDNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
prediction_fn=slim.softmax,
reuse=None,
scope='ssd_512_vgg'):
"""SSD net definition.
"""
# End_points collect relevant activations for external use.
end_points = {}
with tf.variable_scope(scope, 'ssd_512_vgg', [inputs], reuse=reuse):
# Original VGG-16 blocks.
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
end_points['block1'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool1')
# Block 2.
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
end_points['block2'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool2')
# Block 3.
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
end_points['block3'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool3')
# Block 4.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
end_points['block4'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool4')
# Block 5.
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
end_points['block5'] = net
net = slim.max_pool2d(net, [3, 3], 1, scope='pool5')
# Additional SSD blocks.
# Block 6: let's dilate the hell out of it!
net = slim.conv2d(net, 1024, [3, 3], rate=6, scope='conv6')
end_points['block6'] = net
# Block 7: 1x1 conv. Because the f**k.
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
end_points['block7'] = net
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts).
end_point = 'block8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block11'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block12'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [4, 4],
scope='conv4x4',
padding='VALID')
# Fix padding to match Caffe version (pad=1).
# pad_shape = [(i-j) for i, j in zip(layer_shape(net), [0, 1, 1, 0])]
# net = tf.slice(net, [0, 0, 0, 0], pad_shape, name='caffe_pad')
end_points[end_point] = net
# Prediction and localisations layers.
predictions = []
logits = []
localisations = []
for i, layer in enumerate(feat_layers):
with tf.variable_scope(layer + '_box'):
p, l = ssd_vgg_300.ssd_multibox_layer(end_points[layer],
num_classes,
anchor_sizes[i],
anchor_ratios[i],
normalizations[i])
predictions.append(prediction_fn(p))
logits.append(p)
localisations.append(l)
return predictions, localisations, logits, end_points
def ssd_net(inputs,
num_classes=SSDNet.default_params.num_classes,
feat_layer=SSDNet.default_params.feat_layers,
anchor_sizes=SSDNet.default_params.anchor_sizes,
anchor_ratios=SSDNet.default_params.anchor_raios,
normalizations=SSDNet.default_params.normalizations,
is_training=True,
dropout_keep_prob=0.5,
prediction_fn=slim.softmax,
reues=None,
scope='ssd_512_vgg'):
"""SSD net definition"""
end_points = {}
with tf.variable_scope(scope, 'ssd_512_vgg', [inputs], reuse=reues):
# Original VGG-16 blocks
net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')
end_points['block1'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool1')
# Block 2
net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')
end_points['block2'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool2')
# Block 3
net = slim.repeat(net, 3, slim.conv2d, 256, [3, 3], scope='conv3')
end_points['block3'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool3')
# Block 4
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')
end_points['block4'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool4')
# Block 5
net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')
end_points['block5'] = net
net = slim.max_pool2d(net, [2, 2], scope='pool5')
# Additional SSD blocks
# Block 6
net = slim.conv2d(net, 1024, [3, 3], rate=6,
scope='conv6') #TODO: find the meaning of this rate
end_points['block6'] = net
# Block 7
net = slim.conv2d(net, 1024, [1, 1], scope='conv7')
end_points['block7'] = net
# Block 8/9/10/11: 1x1 and 3x3 convolutions stride 2 (except lasts)
end_point = 'block8'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 256, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
512, [3, 3],
stride=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block9'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
srtide=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block10'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
srtide=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block11'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
srtide=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
end_point = 'block12'
with tf.variable_scope(end_point):
net = slim.conv2d(net, 128, [1, 1], scope='conv1x1')
net = custom_layers.pad2d(net, pad=(1, 1))
net = slim.conv2d(net,
256, [3, 3],
srtide=2,
scope='conv3x3',
padding='VALID')
end_points[end_point] = net
# Prediction and localisations layers
predictions = []
logits = []
localisations = []
for i, layer in enumerate(feat_layer):
with tf.variable_scope(layer + '_box'):
p, l = ssd_multibox_layer(end_points[layer], num_classes,
anchor_sizes[i], anchor_ratios[i],
normalizations[i])
predictions.append(prediction_fn(p))
logits.append(p)
localisations.append(l)
return predictions, localisations, logits, end_point
