def _seresblock_proj(name,
input,
weight1,
weight2,
weight_proj,
filter_num,
ratio=4):
with tf.variable_scope(name):
fc1_w = tf.get_variable(name=name + '_fc1_w',
shape=[filter_num,
int(filter_num / ratio)],
initializer=he())
fc2_w = tf.get_variable(name=name + '_fc2_w',
shape=[int(filter_num / ratio), filter_num],
initializer=he())
# residual block
conv = tf.nn.conv2d(input,
weight1,
strides=(1, 1, 1, 1),
padding='SAME')
bn = tf.layers.batch_normalization(conv)
af = tf.nn.relu(bn)
conv = tf.nn.conv2d(af, weight2, strides=(1, 1, 1, 1), padding='SAME')
bn = tf.layers.batch_normalization(conv)
proj = tf.nn.conv2d(input,
weight_proj,
strides=(1, 1, 1, 1),
padding='SAME')
bn_proj = tf.layers.batch_normalization(proj)
# squeeze & exitation block
gap = tf.reduce_mean(bn_proj, axis=[1, 2]) # (B x 1 x C)
fc1 = tf.reshape(gap, shape=[-1, filter_num]) # (B x C)
fc1 = tf.matmul(fc1, fc1_w) # (B x C/R)
relu = tf.nn.relu(fc1) # (B x C/R)
fc2 = tf.matmul(relu, fc2_w) # (B x C)
sig = tf.nn.sigmoid(fc2) # (B x C)
descriptor = tf.reshape(sig, shape=[-1, 1, 1, filter_num])
out = tf.nn.relu(bn + bn_proj * descriptor)
return out
def _seresblock_double(name, input, weight1, weight2, filter_num, ratio=4):
with tf.variable_scope(name):
fc1_w = tf.get_variable(name=name + '_fc1_w',
shape=[filter_num,
int(filter_num / ratio)],
initializer=he())
fc2_w = tf.get_variable(name=name + '_fc2_w',
shape=[int(filter_num / ratio), filter_num],
initializer=he())
# residual block
bn1 = tf.layers.batch_normalization(input)
relu1 = tf.nn.leaky_relu(bn1)
conv1 = tf.nn.conv2d(relu1,
weight1,
strides=(1, 1, 1, 1),
padding='SAME')
bn2 = tf.layers.batch_normalization(conv1)
relu2 = tf.nn.leaky_relu(bn2)
conv2 = tf.nn.conv2d(relu2,
weight2,
strides=(1, 1, 1, 1),
padding='SAME')
# squeeze & exitation block
gap = tf.reduce_mean(conv2, axis=[1, 2]) # (B x 1 x C)
fc1 = tf.reshape(gap, shape=[-1, filter_num]) # (B x C)
fc1 = tf.matmul(fc1, fc1_w) # (B x C/R)
relu = tf.nn.relu(fc1) # (B x C/R)
fc2 = tf.matmul(relu, fc2_w) # (B x C)
sig = tf.nn.sigmoid(fc2) # (B x C)
descriptor = tf.reshape(sig, shape=[-1, 1, 1, filter_num])
out = input + conv2 * descriptor
return out
def __init__(self, name, images, channel, classes):
with tf.variable_scope(name):
# 1 - Filters
self.conv_w = tf.get_variable(name='cls_conv_w',
shape=[3, 3, channel, 64],
initializer=he())
self.res_w1_1 = tf.get_variable(name='cls_res_w1_1',
shape=[3, 3, 64, 64],
initializer=he())
self.res_w1_2 = tf.get_variable(name='cls_res_w1_2',
shape=[3, 3, 64, 64],
initializer=he())
self.res_w1_3 = tf.get_variable(name='cls_res_w1_3',
shape=[3, 3, 64, 64],
initializer=he())
self.res_w1_4 = tf.get_variable(name='cls_res_w1_4',
shape=[3, 3, 64, 64],
initializer=he())
self.proj_w2 = tf.get_variable(name='cls_proj_w2',
shape=[1, 1, 64, 128],
initializer=he())
self.res_w2_1 = tf.get_variable(name='cls_res_w2_1',
shape=[3, 3, 64, 128],
initializer=he())
self.res_w2_2 = tf.get_variable(name='cls_res_w2_2',
shape=[3, 3, 128, 128],
initializer=he())
self.res_w2_3 = tf.get_variable(name='cls_res_w2_3',
shape=[3, 3, 128, 128],
initializer=he())
self.res_w2_4 = tf.get_variable(name='cls_res_w2_4',
shape=[3, 3, 128, 128],
initializer=he())
self.proj_w3 = tf.get_variable(name='cls_proj_w3',
shape=[1, 1, 128, 256],
initializer=he())
self.res_w3_1 = tf.get_variable(name='cls_res_w3_1',
shape=[3, 3, 128, 256],
initializer=he())
self.res_w3_2 = tf.get_variable(name='cls_res_w3_2',
shape=[3, 3, 256, 256],
initializer=he())
self.res_w3_3 = tf.get_variable(name='cls_res_w3_3',
shape=[3, 3, 256, 256],
initializer=he())
self.res_w3_4 = tf.get_variable(name='cls_res_w3_4',
shape=[3, 3, 256, 256],
initializer=he())
self.conv_last_w1 = tf.get_variable(name='conv_last_w1',
shape=[3, 3, 256, 512],
initializer=he())
self.fc_w1 = tf.get_variable(name='fc_w1',
shape=[2048, classes
]) # bounding box coords
# self.concat_w = tf.get_variable(name='concat_w', shape=[3072, classes])
# 2 - Graphs
self.conv = tf.nn.relu(
tf.nn.conv2d(images,
self.conv_w,
strides=(1, 1, 1, 1),
padding='SAME'))
self.pool = tf.nn.max_pool(self.conv,
ksize=(1, 2, 2, 1),
strides=(1, 2, 2, 1),
padding='SAME') # (?,64,64,64)
self.res1_1 = _resblock_double('cls_res1_1', self.pool,
self.res_w1_1, self.res_w1_2)
self.res1_2 = _resblock_double('cls_res1_2', self.res1_1,
self.res_w1_3, self.res_w1_4)
self.pool1 = tf.nn.max_pool(self.res1_2,
ksize=(1, 2, 2, 1),
strides=(1, 2, 2, 1),
padding='SAME') # (?,32,32,64)
self.res2_1 = _resblock_proj('cls_res2_1', self.pool1,
self.res_w2_1, self.res_w2_2,
self.proj_w2)
self.res2_2 = _resblock_double('cls_res2_2', self.res2_1,
self.res_w2_3, self.res_w2_4)
self.pool2 = tf.nn.max_pool(self.res2_2,
ksize=(1, 2, 2, 1),
strides=(1, 2, 2, 1),
padding='SAME') # (?,16,16,128)
self.res3_1 = _resblock_proj('cls_res3_1', self.pool2,
self.res_w3_1, self.res_w3_2,
self.proj_w3)
self.res3_2 = _resblock_double('cls_res3_2', self.res3_1,
self.res_w3_3, self.res_w3_4)
self.pool3 = tf.nn.max_pool(self.res3_2,
ksize=(1, 2, 2, 1),
strides=(1, 2, 2, 1),
padding='SAME') # (?,8,8,256)
self.conv_last = tf.nn.relu(
tf.nn.conv2d(self.pool3,
self.conv_last_w1,
strides=(1, 1, 1, 1),
padding='SAME'))
self.pool_last = tf.nn.avg_pool(self.conv_last,
ksize=(1, 4, 4, 1),
strides=(1, 4, 4, 1),
padding='SAME')
self.flat = tf.reshape(self.pool_last, shape=[-1, 2048])
self.logits = tf.matmul(self.flat, self.fc_w1) # (?, classes * 4)
def __init__(self, name, img, channel, classes):
with tf.variable_scope(name):
# feature weights
self.proj_w1 = tf.get_variable(name='proj_w1',
shape=[1, 1, channel, 64],
initializer=he())
self.res_w1 = tf.get_variable(name='res_w1',
shape=[3, 3, channel, 64],
initializer=he())
self.proj_w2 = tf.get_variable(name='proj_w2',
shape=[1, 1, 64, 128],
initializer=he())
self.res_w2 = tf.get_variable(name='res_w2',
shape=[3, 3, 64, 128],
initializer=he())
self.proj_w3 = tf.get_variable(name='proj_w3',
shape=[1, 1, 128, 256],
initializer=he())
self.res_w3 = tf.get_variable(name='res_w3',
shape=[3, 3, 128, 256],
initializer=he())
self.proj_w4 = tf.get_variable(name='proj_w4',
shape=[1, 1, 256, 512],
initializer=he())
self.res_w4 = tf.get_variable(name='res_w4',
shape=[3, 3, 256, 512],
initializer=he())
self.proj_w5 = tf.get_variable(name='proj_w5',
shape=[1, 1, 512, 1024],
initializer=he())
self.res_w5 = tf.get_variable(name='res_w5',
shape=[3, 3, 512, 1024],
initializer=he())
self.proj_w6 = tf.get_variable(name='proj_w6',
shape=[1, 1, 1024, 2048],
initializer=he())
self.res_w6 = tf.get_variable(name='res_w6',
shape=[3, 3, 1024, 2048],
initializer=he())
self.proj_w7 = tf.get_variable(name='proj_w7',
shape=[1, 1, 2048, 4096],
initializer=he())
self.res_w7 = tf.get_variable(name='res_w7',
shape=[3, 3, 2048, 4096],
initializer=he())
self.fc_w = tf.get_variable(name='fc_w',
shape=[4096, classes],
initializer=he())
# common graphs
self.res1_1 = ops.resblock_single('res1_1', img, self.res_w1,
self.proj_w1)
self.pool1 = tf.nn.max_pool(self.res1_1,
ksize=(1, 2, 2, 1),
strides=(1, 2, 2, 1),
padding='SAME') # 40,40,64
self.res2_1 = ops.resblock_single('res2_1', self.pool1,
self.res_w2, self.proj_w2)
self.pool2 = tf.nn.max_pool(self.res2_1,
ksize=(1, 2, 2, 1),
strides=(1, 2, 2, 1),
padding='SAME') # 20,20,128
self.res3_1 = ops.resblock_single('res3_1', self.pool2,
self.res_w3, self.proj_w3)
self.pool3 = tf.nn.max_pool(self.res3_1,
ksize=(1, 2, 2, 1),
strides=(1, 2, 2, 1),
padding='SAME') # 10,10,256
self.res4_1 = ops.resblock_single('res4_1', self.pool3,
self.res_w4, self.proj_w4)
self.pool4 = tf.nn.max_pool(self.res4_1,
ksize=(1, 2, 2, 1),
strides=(1, 2, 2, 1),
padding='SAME') # 5,5,512
self.res5_1 = ops.resblock_single('res5_1', self.pool4,
self.res_w5, self.proj_w5)
self.pool5 = tf.nn.max_pool(self.res5_1,
ksize=(1, 2, 2, 1),
strides=(1, 2, 2, 1),
padding='SAME') # 3,3,1024
self.res6_1 = ops.resblock_single('res6_1', self.pool5,
self.res_w6, self.proj_w6)
self.pool6 = tf.nn.max_pool(self.res6_1,
ksize=(1, 2, 2, 1),
strides=(1, 2, 2, 1),
padding='SAME') # 2,2,2048
self.res7_1 = ops.resblock_single('res7_1', self.pool6,
self.res_w7, self.proj_w7)
self.pool7 = tf.nn.avg_pool(self.res7_1,
ksize=(1, 2, 2, 1),
strides=(1, 2, 2, 1),
padding='SAME') # 1,1,4096
self.logits = tf.nn.softmax(
tf.matmul(tf.squeeze(self.pool7), self.fc_w))