用原生Tensorflow编写的Inception-ResNet-V1网络（想了解模型实现细节的看）
# prepare input data
IMAGE_HEI = 128
IMAGE_WID = 128
IMAGE_CHA = 3
CLASS_NUM = 2
X = tf.placeholder(tf.float32, [None, IMAGE_HEI, IMAGE_WID, IMAGE_CHA])
Y_= tf.placeholder(tf.float32, [None, CLASS_NUM])
keep_prob = tf.placeholder(tf.float32)
lr = tf.placeholder(tf.float32) # learning rate
下边的代码片段实现了Inception-ResNet-V1网络的：Stem，Inception-resnet-A，Reduction-A，Inception-resnet-B，Reduction-B，Inception-resnet-C五个模块。
#---------- Stem Block ------------------------------------------#
# Input layer out: 128 * 128 * 3
# |
# 3*3 conv 32 out: 128 * 128 * 32
# |
# 3*3 conv 32, stride=2 (we use max_pool instead) out: 64 * 64 * 32
# |
# 3*3 conv 64 out: 64 * 64 * 64
# |
# 3*3 conv 64, stride=2 (we use max_pool instead) out: 32 * 32 * 64
# |
# 3*3 conv 128 out: 32 * 32 * 128
# |
# 3*3 conv 128,stride=2 (we use max_pool instead) out: 16 * 16 * 128
# |
# relu (output 16*16*128) out: 16 * 16 * 128
stem_conv1_k = 32
stem_conv1_w = tf.Variable(tf.truncated_normal([3, 3, IMAGE_CHA, stem_conv1_k], stddev = 0.1))
stem_conv1_b = tf.Variable(tf.ones([stem_conv1_k]) / 10)
stem_conv1_o = tf.nn.conv2d( X, stem_conv1_w, strides = [1, 1, 1, 1], padding = 'SAME')
stem_conv1_o = stem_conv1_o + stem_conv1_b;
# out: 64*64*32
stem_conv2_k = 32
stem_conv2_w = tf.Variable(tf.truncated_normal([3, 3, stem_conv1_k, stem_conv2_k], stddev = 0.1))
stem_conv2_b = tf.Variable(tf.ones([stem_conv2_k]) / 10)
stem_conv2_o = tf.nn.conv2d(stem_conv1_o, stem_conv2_w, strides = [1, 1, 1, 1], padding = 'SAME')
stem_conv2_o = stem_conv2_o + stem_conv2_b;
stem_conv2_o = tf.nn.max_pool(stem_conv2_o, ksize= [1,2,2,1], strides= [1,2,2,1], padding='SAME')
# out: 64*64*64
stem_conv3_k = 64
stem_conv3_w = tf.Variable(tf.truncated_normal([3, 3, stem_conv2_k, stem_conv3_k], stddev = 0.1))
stem_conv3_b = tf.Variable(tf.ones([stem_conv3_k]) / 10)
stem_conv3_o = tf.nn.conv2d(stem_conv2_o, stem_conv3_w, strides = [1, 1, 1, 1], padding = 'SAME')
stem_conv3_o = stem_conv3_o + stem_conv3_b;
# 32*32*64
stem_conv4_k = 64
stem_conv4_w = tf.Variable(tf.truncated_normal([3, 3, stem_conv3_k, stem_conv4_k], stddev = 0.1))
stem_conv4_b = tf.Variable(tf.ones([stem_conv4_k]) / 10)
stem_conv4_o = tf.nn.conv2d(stem_conv3_o, stem_conv4_w, strides = [1, 1, 1, 1], padding = 'SAME')
stem_conv4_o = stem_conv4_o + stem_conv4_b;
stem_conv4_o = tf.nn.max_pool(stem_conv4_o, ksize= [1,2,2,1], strides= [1,2,2,1], padding='SAME')
# In 32*32*64, Out 32*32*128
stem_conv5_k = 128
stem_conv5_w = tf.Variable(tf.truncated_normal([3, 3, stem_conv4_k, stem_conv5_k], stddev = 0.1))
stem_conv5_b = tf.Variable(tf.ones([stem_conv5_k]) / 10)
stem_conv5_o = tf.nn.conv2d(stem_conv4_o, stem_conv5_w, strides = [1, 1, 1, 1], padding = 'SAME')
stem_conv5_o = stem_conv5_o + stem_conv5_b;
# In 32*32*128, Out 16*16*128
stem_conv6_k = 128
stem_conv6_w = tf.Variable(tf.truncated_normal([3, 3, stem_conv5_k, stem_conv6_k], stddev = 0.1))
stem_conv6_b = tf.Variable(tf.ones([stem_conv6_k]) / 10)
stem_conv6_o = tf.nn.conv2d(stem_conv5_o, stem_conv6_w, strides = [1, 1, 1, 1], padding = 'SAME')
stem_conv6_o = stem_conv6_o + stem_conv6_b;
stem_conv6_o = tf.nn.max_pool(stem_conv6_o, ksize= [1,2,2,1], strides= [1,2,2,1], padding='SAME')
stem_conv6_o = tf.nn.relu(stem_conv6_o)
#---------- Inception-ResNet-A 1 --------------------------------#
# We only use 1 Inception-ResNet-A block. There are 5 Inception-ResNet-A
# blocks in origin paper.
#
# previous layer (from stem block, input 16*16*128)
# |
# _______|_____________________________________________________
# / | | \
# 1*1 conv 32 1*1 conv 32 1*1 conv 32 |
# | | | |
# | 3*3 conv 32 3*3 conv 32 |
# | | | |
# | | 3*3 conv 32 |
# \_________________________C______________________/ |
# | |
# | three signals concat |
# | |
# 1*1 conv 128 |
# \_________________+_________________/
# |
# | two signal add (must have the same shape)
# |
# relu(output 16*16*128)
# Through the graph above, we know the input and output data have the same shape.
# path1
ia1_conv11_k = 32
ia1_conv11_w = tf.Variable(tf.truncated_normal([1, 1, stem_conv6_k, ia1_conv11_k], stddev = 0.1))
ia1_conv11_b = tf.Variable(tf.ones([ia1_conv11_k]) / 10)
ia1_conv11_o = tf.nn.conv2d(stem_conv6_o, ia1_conv11_w, strides = [1, 1, 1, 1], padding = 'SAME')
ia1_conv11_o = ia1_conv11_o + ia1_conv11_b;
#path 2
ia1_conv21_k = 32
ia1_conv21_w = tf.Variable(tf.truncated_normal([1, 1, stem_conv6_k, ia1_conv21_k], stddev = 0.1))
ia1_conv21_b = tf.Variable(tf.ones([ia1_conv21_k]) / 10)
ia1_conv21_o = tf.nn.conv2d(stem_conv6_o, ia1_conv21_w, strides = [1, 1, 1, 1], padding = 'SAME')
ia1_conv21_o = ia1_conv21_o + ia1_conv21_b;
ia1_conv22_k = 32
ia1_conv22_w = tf.Variable(tf.truncated_normal([3, 3, ia1_conv21_k, ia1_conv22_k], stddev = 0.1))
ia1_conv22_b = tf.Variable(tf.ones([ia1_conv22_k]) / 10)
ia1_conv22_o = tf.nn.conv2d(ia1_conv21_o, ia1_conv22_w, strides = [1, 1, 1, 1], padding = 'SAME')
ia1_conv22_o = ia1_conv22_o + ia1_conv22_b;
# path 3
ia1_conv31_k = 32
ia1_conv31_w = tf.Variable(tf.truncated_normal([1, 1, stem_conv6_k, ia1_conv31_k], stddev = 0.1))
ia1_conv31_b = tf.Variable(tf.ones([ia1_conv31_k]) / 10)
ia1_conv31_o = tf.nn.conv2d(stem_conv6_o, ia1_conv31_w, strides = [1, 1, 1, 1], padding = 'SAME')
ia1_conv31_o = ia1_conv31_o + ia1_conv31_b;
ia1_conv32_k = 32
ia1_conv32_w = tf.Variable(tf.truncated_normal([3, 3, ia1_conv31_k, ia1_conv32_k], stddev = 0.1))
ia1_conv32_b = tf.Variable(tf.ones([ia1_conv32_k]) / 10)
ia1_conv32_o = tf.nn.conv2d(ia1_conv31_o, ia1_conv32_w, strides = [1, 1, 1, 1], padding = 'SAME')
ia1_conv32_o = ia1_conv32_o + ia1_conv32_b;
ia1_conv33_k = 32
ia1_conv33_w = tf.Variable(tf.truncated_normal([3, 3, ia1_conv32_k, ia1_conv33_k], stddev = 0.1))
ia1_conv33_b = tf.Variable(tf.ones([ia1_conv33_k]) / 10)
ia1_conv33_o = tf.nn.conv2d(ia1_conv32_o, ia1_conv33_w, strides = [1, 1, 1, 1], padding = 'SAME')
ia1_conv33_o = ia1_conv33_o + ia1_conv33_b;
# concat path1 - 3
ia1_concat_i = tf.concat(axis=3, values=[ia1_conv11_o, ia1_conv22_o, ia1_conv33_o])
ia1_conv41_i = ia1_conv11_k + ia1_conv22_k + ia1_conv33_k
# conv after concat
ia1_conv41_k = 128
ia1_conv41_w = tf.Variable(tf.truncated_normal([1, 1, ia1_conv41_i, ia1_conv41_k], stddev = 0.1))
ia1_conv41_b = tf.Variable(tf.ones([ia1_conv41_k]) / 10)
ia1_conv41_o = tf.nn.conv2d(ia1_concat_i, ia1_conv41_w, strides = [1, 1, 1, 1], padding = 'SAME')
ia1_conv41_o = ia1_conv41_o + ia1_conv41_b;
# add signals and relu
ia1_conv41_o = tf.nn.relu(stem_conv6_o + ia1_conv41_o)
#---------- Reduction-A -----------------------------------------#
# previous layer (input 16*16*128)
# |
# _______|___________________________________
# / | \
# 2*2 max pooling 3*3 conv 64, stride=2 1*1 conv 32
# | | |
# | | 3*3 conv 32
# | | |
# | | 3*3 conv 64, stride=2
# |_________________________C_________________|
# |
# | threee signal concat
# |
# relu (output 8*8*256)
# Through the graph above, we can take this whole block as a pooling layer.
# path 1
ra1_pool11_o = tf.nn.max_pool(ia1_conv41_o, ksize= [1,2,2,1], strides= [1,2,2,1], padding='SAME')
# path 2
ra1_conv11_k = 64
ra1_conv11_w = tf.Variable(tf.truncated_normal([3, 3, ia1_conv41_k, ra1_conv11_k], stddev = 0.1))
ra1_conv11_b = tf.Variable(tf.ones([ra1_conv11_k]) / 10)
ra1_conv11_o = tf.nn.conv2d(ia1_conv41_o, ra1_conv11_w, strides = [1, 1, 1, 1], padding = 'SAME')
ra1_conv11_o = ra1_conv11_o + ra1_conv11_b;
ra1_conv11_o = tf.nn.max_pool(ra1_conv11_o, ksize= [1,2,2,1], strides= [1,2,2,1], padding='SAME')
# path 3
ra1_conv21_k = 32
ra1_conv21_w = tf.Variable(tf.truncated_normal([1, 1, ia1_conv41_k, ra1_conv21_k], stddev = 0.1))
ra1_conv21_b = tf.Variable(tf.ones([ra1_conv21_k]) / 10)
ra1_conv21_o = tf.nn.conv2d(ia1_conv41_o, ra1_conv21_w, strides = [1, 1, 1, 1], padding = 'SAME')
ra1_conv21_o = ra1_conv21_o + ra1_conv21_b;
ra1_conv22_k = 32
ra1_conv22_w = tf.Variable(tf.truncated_normal([3, 3, ra1_conv21_k, ra1_conv22_k], stddev = 0.1))
ra1_conv22_b = tf.Variable(tf.ones([ra1_conv22_k]) / 10)
ra1_conv22_o = tf.nn.conv2d(ra1_conv21_o, ra1_conv22_w, strides = [1, 1, 1, 1], padding = 'SAME')
ra1_conv22_o = ra1_conv22_o + ra1_conv22_b;
ra1_conv23_k = 64
ra1_conv23_w = tf.Variable(tf.truncated_normal([3, 3, ra1_conv22_k, ra1_conv23_k], stddev = 0.1))
ra1_conv23_b = tf.Variable(tf.ones([ra1_conv23_k]) / 10)
ra1_conv23_o = tf.nn.conv2d(ra1_conv22_o, ra1_conv23_w, strides = [1, 1, 1, 1], padding = 'SAME')
ra1_conv23_o = ra1_conv23_o + ra1_conv23_b;
ra1_conv23_o = tf.nn.max_pool(ra1_conv23_o, ksize= [1,2,2,1], strides= [1,2,2,1], padding='SAME')
# Out 8*8*256
ra1_concat_o = tf.concat(axis=3, values=[ra1_pool11_o, ra1_conv11_o, ra1_conv23_o])
# relu
ra1_concat_o = tf.nn.relu(ra1_concat_o)
#---------- Inception-ResNet-B 1 --------------------------------#
# We only use 1 Inception-ResNet-A block. There are 10 Inception-ResNet-A block in origin paper.
# previous layer (input 8*8*256)
# |
# ___2___|______1____________________3_______
# / | \
# 1*1 conv 128 1*1 conv 128 |
# | | |
# 1*5 conv 128 | |
# | | |
# 5*1 conv 128 | |
# | | |
# |_______concat________| |
# | |
# 1*1 conv 256 |
# | |
# |______________+_____________|
# |
# | two signal add
# |
# relu (output 8*8*256)
# path 1
ib1_conv11_i = 256
ib1_conv11_k = 128
ib1_conv11_w = tf.Variable(tf.truncated_normal([1, 1, ib1_conv11_i, ib1_conv11_k], stddev = 0.1))
ib1_conv11_b = tf.Variable(tf.ones([ib1_conv11_k]) / 10)
ib1_conv11_o = tf.nn.conv2d(ra1_concat_o, ib1_conv11_w, strides = [1, 1, 1, 1], padding = 'SAME')
ib1_conv11_o = ib1_conv11_o + ib1_conv11_b;
# path 2
ib1_conv21_k = 128
ib1_conv21_w = tf.Variable(tf.truncated_normal([1, 1, ib1_conv11_i, ib1_conv21_k], stddev = 0.1))
ib1_conv21_b = tf.Variable(tf.ones([ib1_conv21_k]) / 10)
ib1_conv21_o = tf.nn.conv2d(ra1_concat_o, ib1_conv21_w, strides = [1, 1, 1, 1], padding = 'SAME')
ib1_conv21_o = ib1_conv21_o + ib1_conv21_b;
ib1_conv22_k = 128
ib1_conv22_w = tf.Variable(tf.truncated_normal([1, 5, ib1_conv21_k, ib1_conv22_k], stddev = 0.1))
ib1_conv22_b = tf.Variable(tf.ones([ib1_conv22_k]) / 10)
ib1_conv22_o = tf.nn.conv2d(ib1_conv21_o, ib1_conv22_w, strides = [1, 1, 1, 1], padding = 'SAME')
ib1_conv22_o = ib1_conv22_o + ib1_conv22_b;
ib1_conv23_k = 128
ib1_conv23_w = tf.Variable(tf.truncated_normal([5, 1, ib1_conv22_k, ib1_conv23_k], stddev = 0.1))
ib1_conv23_b = tf.Variable(tf.ones([ib1_conv23_k]) / 10)
ib1_conv23_o = tf.nn.conv2d(ib1_conv22_o, ib1_conv23_w, strides = [1, 1, 1, 1], padding = 'SAME')
ib1_conv23_o = ib1_conv23_o + ib1_conv23_b;
# concat path 1 - 2
ib1_concat_i = tf.concat(axis=3, values=[ib1_conv11_o, ib1_conv23_o])
# the conv after concat
ib1_conv31_i = ib1_conv11_k + ib1_conv23_k
ib1_conv31_k = 256
ib1_conv31_w = tf.Variable(tf.truncated_normal([1, 1, ib1_conv31_i, ib1_conv31_k], stddev = 0.1))
ib1_conv31_b = tf.Variable(tf.ones([ib1_conv31_k]) / 10)
ib1_conv31_o = tf.nn.conv2d(ib1_concat_i, ib1_conv31_w, strides = [1, 1, 1, 1], padding = 'SAME')
ib1_conv31_o = ib1_conv31_o + ib1_conv31_b;
# signal add
ib1_conv31_o = tf.nn.relu(ra1_concat_o + ib1_conv31_o)
#---------- Reduction-B -----------------------------------------#
# previous layer (input 8*8*256)
# |
# _______|________________________________________________________
# / | | \
# 2*2 max pooling 1*1 conv 64 1*1 conv 64 1*1 conv 64
# | | | |
# 3*3 conv 64,stride=2 3*3 conv 64,stride=2 3*3 conv 64
# | | | |
# | | | 3*3 conv 128,stride=2
# | | | |
# \_________________________|______________________|_______________/
# C
# | four signals concat
# |
# relu (output 4*4*512)
# path 1
rb1_pool11_o = tf.nn.max_pool(ib1_conv31_o, ksize= [1,2,2,1], strides= [1,2,2,1], padding='SAME')
rb1_conv11_i = 256
rb1_conv11_k = 64
rb1_conv11_w = tf.Variable(tf.truncated_normal([1, 1, rb1_conv11_i, rb1_conv11_k], stddev = 0.1))
rb1_conv11_b = tf.Variable(tf.ones([rb1_conv11_k]) / 10)
rb1_conv11_o = tf.nn.conv2d(ib1_conv31_o, rb1_conv11_w, strides = [1, 1, 1, 1], padding = 'SAME')
rb1_conv11_o = rb1_conv11_o + rb1_conv11_b;
rb1_conv12_k = 64
rb1_conv12_w = tf.Variable(tf.truncated_normal([3, 3, rb1_conv11_k, rb1_conv12_k], stddev = 0.1))
rb1_conv12_b = tf.Variable(tf.ones([rb1_conv12_k]) / 10)
rb1_conv12_o = tf.nn.conv2d(rb1_conv11_o, rb1_conv12_w, strides = [1, 1, 1, 1], padding = 'SAME')
rb1_conv12_o = rb1_conv12_o + rb1_conv12_b;
rb1_conv12_o = tf.nn.max_pool(rb1_conv12_o, ksize= [1,2,2,1], strides= [1,2,2,1], padding='SAME')
# path 2
rb1_conv21_k = 64
rb1_conv21_w = tf.Variable(tf.truncated_normal([1, 1, rb1_conv11_i, rb1_conv21_k], stddev = 0.1))
rb1_conv21_b = tf.Variable(tf.ones([rb1_conv21_k]) / 10)
rb1_conv21_o = tf.nn.conv2d(ib1_conv31_o, rb1_conv21_w, strides = [1, 1, 1, 1], padding = 'SAME')
rb1_conv21_o = rb1_conv21_o + rb1_conv21_b;
rb1_conv22_k = 64
rb1_conv22_w = tf.Variable(tf.truncated_normal([3, 3, rb1_conv21_k, rb1_conv22_k], stddev = 0.1))
rb1_conv22_b = tf.Variable(tf.ones([rb1_conv22_k]) / 10)
rb1_conv22_o = tf.nn.conv2d(rb1_conv21_o, rb1_conv22_w, strides = [1, 1, 1, 1], padding = 'SAME')
rb1_conv22_o = rb1_conv22_o + rb1_conv22_b;
rb1_conv22_o = tf.nn.max_pool(rb1_conv22_o, ksize= [1,2,2,1], strides= [1,2,2,1], padding='SAME')
# path 3
rb1_conv31_k = 64
rb1_conv31_w = tf.Variable(tf.truncated_normal([1, 1, rb1_conv11_i, rb1_conv31_k], stddev = 0.1))
rb1_conv31_b = tf.Variable(tf.ones([rb1_conv31_k]) / 10)
rb1_conv31_o = tf.nn.conv2d(ib1_conv31_o, rb1_conv31_w, strides = [1, 1, 1, 1], padding = 'SAME')
rb1_conv31_o = rb1_conv31_o + rb1_conv31_b;
rb1_conv32_k = 64
rb1_conv32_w = tf.Variable(tf.truncated_normal([3, 3, rb1_conv31_k, rb1_conv32_k], stddev = 0.1))
rb1_conv32_b = tf.Variable(tf.ones([rb1_conv32_k]) / 10)
rb1_conv32_o = tf.nn.conv2d(rb1_conv31_o, rb1_conv32_w, strides = [1, 1, 1, 1], padding = 'SAME')
rb1_conv32_o = rb1_conv32_o + rb1_conv32_b;
rb1_conv33_k = 128
rb1_conv33_w = tf.Variable(tf.truncated_normal([3, 3, rb1_conv32_k, rb1_conv33_k], stddev = 0.1))
rb1_conv33_b = tf.Variable(tf.ones([rb1_conv33_k]) / 10)
rb1_conv33_o = tf.nn.conv2d(rb1_conv32_o, rb1_conv33_w, strides = [1, 1, 1, 1], padding = 'SAME')
rb1_conv33_o = rb1_conv33_o + rb1_conv33_b;
rb1_conv33_o = tf.nn.max_pool(rb1_conv33_o, ksize= [1,2,2,1], strides= [1,2,2,1], padding='SAME')
# concat
rb1_concat_o = tf.concat(axis=3, values=[rb1_pool11_o, rb1_conv12_o, rb1_conv22_o, rb1_conv33_o])
rb1_concat_o = tf.nn.relu(rb1_concat_o)
#---------- Inception-ResNet-C 1 --------------------------------#
# We only use 1 Inception-ResNet-A block. There are 5 Inception-ResNet-A block in origin paper.
# previous layer (input 4*4*512)
# |
# ___2___|______1____________________3_______
# / | \
# 1*1 conv 256 1*1 conv 256 |
# | | |
# 1*3 conv 256 | |
# | | |
# 3*1 conv 256 | |
# | | |
# |_______concat________| |
# | |
# 1*1 conv 512 |
# | |
# |______________+_____________|
# |
# | two signal add
# |
# relu (output 4*4*512)
# path 1
ic1_conv11_i = 512
ic1_conv11_k = 256
ic1_conv11_w = tf.Variable(tf.truncated_normal([1, 1, ic1_conv11_i, ic1_conv11_k], stddev = 0.1))
ic1_conv11_b = tf.Variable(tf.ones([ic1_conv11_k]) / 10)
ic1_conv11_o = tf.nn.conv2d(rb1_concat_o, ic1_conv11_w, strides = [1, 1, 1, 1], padding = 'SAME')
ic1_conv11_o = ic1_conv11_o + ic1_conv11_b;
# path 2
ic1_conv21_k = 256
ic1_conv21_w = tf.Variable(tf.truncated_normal([1, 1, ic1_conv11_i, ic1_conv21_k], stddev = 0.1))
ic1_conv21_b = tf.Variable(tf.ones([ic1_conv21_k]) / 10)
ic1_conv21_o = tf.nn.conv2d(rb1_concat_o, ic1_conv21_w, strides = [1, 1, 1, 1], padding = 'SAME')
ic1_conv21_o = ic1_conv21_o + ic1_conv21_b;
ic1_conv22_k = 256
ic1_conv22_w = tf.Variable(tf.truncated_normal([1, 3, ic1_conv21_k, ic1_conv22_k], stddev = 0.1))
ic1_conv22_b = tf.Variable(tf.ones([ic1_conv22_k]) / 10)
ic1_conv22_o = tf.nn.conv2d(ic1_conv21_o, ic1_conv22_w, strides = [1, 1, 1, 1], padding = 'SAME')
ic1_conv22_o = ic1_conv22_o + ic1_conv22_b;
ic1_conv23_k = 256
ic1_conv23_w = tf.Variable(tf.truncated_normal([3, 1, ic1_conv22_k, ic1_conv23_k], stddev = 0.1))
ic1_conv23_b = tf.Variable(tf.ones([ic1_conv23_k]) / 10)
ic1_conv23_o = tf.nn.conv2d(ic1_conv22_o, ic1_conv23_w, strides = [1, 1, 1, 1], padding = 'SAME')
ic1_conv23_o = ic1_conv23_o + ic1_conv23_b;
# concat
ic1_concat_i = tf.concat(axis=3, values=[ic1_conv11_o, ic1_conv23_o])
ic1_conv31_i = ic1_conv11_k + ic1_conv23_k
# conv after concat
ic1_conv31_k = 512
ic1_conv31_w = tf.Variable(tf.truncated_normal([1, 1, ic1_conv31_i, ic1_conv31_k], stddev = 0.1))
ic1_conv31_b = tf.Variable(tf.ones([ic1_conv31_k]) / 10)
ic1_conv31_o = tf.nn.conv2d(ic1_concat_i, ic1_conv31_w, strides = [1, 1, 1, 1], padding = 'SAME')
ic1_conv31_o = ic1_conv31_o + ic1_conv31_b;
# signal add and relu
ic1_conv31_o = tf.nn.relu(rb1_concat_o + ic1_conv31_o)对于Inception-ResNet-V1网络的Average Pooling，Dropout，Softmax，我用了三个full connection来替代。
#---------- Average Pooling / Dropout / Softmax -----------------#
# here we use full connection layers instead the left 3 models in Incpetion-ResNet-V1
# define the number of neruals in full connection layers
Full1_Num = 1024
Full2_Num = 1024
Full3_Num = CLASS_NUM
# full connection 1 w and b
Full1W = tf.Variable(tf.truncated_normal([4*4*ic1_conv31_k, Full1_Num], stddev=0.1))
Full1B = tf.Variable(tf.ones([Full1_Num]) / 10)
# full connection 2 w and b
Full2W = tf.Variable(tf.truncated_normal([Full1_Num, Full2_Num], stddev=0.1))
Full2B = tf.Variable(tf.ones([Full2_Num]) / 10)
# full connection 3 w and b
Full3W = tf.Variable(tf.truncated_normal([Full2_Num, Full3_Num], stddev=0.1))
Full3B = tf.Variable(tf.ones([Full3_Num]) / 10)
# full connection
fc1 = tf.reshape(ic1_conv31_o, shape=[-1, 4*4*ic1_conv31_k])
re1 = tf.nn.relu(tf.matmul(fc1, Full1W) + Full1B)
dp1 = tf.nn.dropout(re1, keep_prob)
re2 = tf.nn.relu(tf.matmul(dp1, Full2W) + Full2B)
dp2 = tf.nn.dropout(re2, keep_prob)
# full connection
Ylogits = tf.matmul(dp2, Full3W) + Full3B
Y = tf.nn.softmax(Ylogits)
# cross entropy
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=Ylogits, labels=Y_)
cross_entropy = tf.reduce_mean(cross_entropy) * 100
# accuracy of the trained model, between 0 (worst) and 1 (best)
cls_true = tf.argmax(Y_, 1)
cls_eavl = tf.argmax(Y, 1)
correct_prediction = tf.equal(cls_eavl, cls_true)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
# training step, the learning rate is a placeholder
train_step = tf.train.AdamOptimizer(lr).minimize(cross_entropy)
# init
init = tf.global_variables_initializer()
# save model
Saver = tf.train.Saver(max_to_keep = 1) # defaults to saving all variables
ModelSavePath = './storage/'
""" computer the learning rate through a number
"""
def calculate_lr(INDEX):
max_learning_rate = 0.005
min_learning_rate = 1e-4
decay_speed = 400
if (INDEX >= 0):
learning_rate = min_learning_rate + \
(max_learning_rate - min_learning_rate) * \
math.exp(-INDEX / decay_speed)
return learning_rate
else:
return min_learning_rate
"""一个辅助函数，用来计算一批样本中，分对的比例和分错的比例
"""
def compute_accu(eav, tru):
N = len(eav)
correct_num = 0
false_positive_num = 0
false_negitive_num = 0
for i in range(N):
if (eav[i] == tru[i]):
correct_num = correct_num + 1
else:
if tru[i] == 1:
false_negitive_num = false_negitive_num + 1
else:
false_positive_num = false_positive_num + 1
return correct_num*1.0 / N, false_positive_num*1.0 / N, false_negitive_num*1.0 / N训练：自己准备样本集
def count_parameters():
print ('[INFO ] total parameters: %d'%np.sum([np.prod(v.get_shape().as_list()) for v in tf.trainable_variables()]))
"""训练，用tfrecords格式的数据
"""
def train_use_tfrecords_data():
with tf.Session() as sess:
sess.run(init)
ckpt = tf.train.get_checkpoint_state(ModelSavePath)
if ckpt and ckpt.model_checkpoint_path:
Saver.restore(sess, ckpt.model_checkpoint_path)
print '[INFO ] restore model successfully'
else:
print '[INFO ] no model so far'
count_parameters()
# prepare train data
batch_size = 100
image_tra_cls0, label_tra_cls0 = samples_to_tfrecord.read_tfrecord( \
['/home/raintai/z_tf/train0_class0.tfrecords', \
'/home/raintai/z_tf/train1_class0.tfrecords', \
'/home/raintai/z_tf/train2_class0.tfrecords', \
'/home/raintai/z_tf/train3_class0.tfrecords'],\
batch_size = batch_size)
image_tra_cls1, label_tra_cls1 = samples_to_tfrecord.read_tfrecord( \
['/home/raintai/z_tf/train0_class1.tfrecords', \
'/home/raintai/z_tf/train1_class1.tfrecords', \
'/home/raintai/z_tf/train2_class1.tfrecords'],\
batch_size = batch_size)
# prepare test data
image_tes_cls0, label_tes_cls0 = samples_to_tfrecord.read_tfrecord( \
['/home/raintai/z_tf/test0_class0.tfrecords', \
'/home/raintai/z_tf/test1_class0.tfrecords'],\
batch_size = batch_size)
image_tes_cls1, label_tes_cls1 = samples_to_tfrecord.read_tfrecord( \
['/home/raintai/z_tf/test0_class1.tfrecords', \
'/home/raintai/z_tf/test1_class1.tfrecords'],\
batch_size = batch_size)
# start queue
threads = tf.train.start_queue_runners(sess=sess)
# one hot encode
label_tra_cls0 = tf.one_hot(label_tra_cls0, CLASS_NUM, dtype = tf.int32)
label_tra_cls1 = tf.one_hot(label_tra_cls1, CLASS_NUM, dtype = tf.int32)
label_tes_cls0 = tf.one_hot(label_tes_cls0, CLASS_NUM, dtype = tf.int32)
label_tes_cls1 = tf.one_hot(label_tes_cls1, CLASS_NUM, dtype = tf.int32)
iterNum = 1000000000L
for i in xrange(1, iterNum):
batch_X_tra_cls0, batch_Y_tra_cls0 = sess.run([image_tra_cls0, label_tra_cls0])
batch_X_tra_cls1, batch_Y_tra_cls1 = sess.run([image_tra_cls1, label_tra_cls1])
batch_X = np.append(batch_X_tra_cls0, batch_X_tra_cls1, axis = 0)
batch_Y = np.append(batch_Y_tra_cls0, batch_Y_tra_cls1, axis = 0)
#print(batch_X.shape, batch_Y.shape)
cur_lr = calculate_lr(INDEX = i)
#cur_lr = 1e-4
eav, tru, cor, a, c, s = sess.run( \
[cls_eavl, cls_true, correct_prediction, accuracy, cross_entropy, train_step], \
{X: batch_X, Y_: batch_Y, lr: cur_lr, keep_prob: 0.6})
correct_num, false_positive_num, false_negitive_num = compute_accu(eav, tru)
if (i + 1) % 10 == 0:
saveFg = 0
if (i + 1) % 100 == 0:
# save tensorflow model
Saver.save(sess, ModelSavePath + "model", global_step = (i + 1))
saveFg = 1
# test
batch_X_tes_cls0, batch_Y_tes_cls0 = sess.run([image_tes_cls0, label_tes_cls0])
batch_X_tes_cls1, batch_Y_tes_cls1 = sess.run([image_tes_cls1, label_tes_cls1])
test_batch_x = np.append(batch_X_tes_cls0, batch_X_tes_cls1, axis = 0)
test_batch_y = np.append(batch_Y_tes_cls0, batch_Y_tes_cls1, axis = 0)
eav1, tru1, cor1, test_cross, test_accu = sess.run( \
[cls_eavl, cls_true, correct_prediction, cross_entropy, accuracy], \
{X: test_batch_x, Y_: test_batch_y, keep_prob: 1.0})
correct_num1, false_positive_num1, false_negitive_num1 = compute_accu(eav1, tru1)
if not saveFg:
print \
'%6d'%(i + 1), \
'%12.4f(cross) '%c, \
'%3.2f(accu) -%3.3f-%3.3f'%(a, false_positive_num, false_negitive_num), \
'%9.4f(cross) '%test_cross, \
'%3.2f(accu) -%3.3f-%3.3f'%(test_accu, false_positive_num1, false_negitive_num1), \
'%10.5f(lr) '%cur_lr
else:
print \
'%6d'%(i + 1), \
'%12.4f(cross) '%c, \
'%3.2f(accu) -%3.3f-%3.3f'%(a, false_positive_num, false_negitive_num), \
'%9.4f(cross) '%test_cross, \
'%3.2f(accu) -%3.3f-%3.3f'%(test_accu, false_positive_num1, false_negitive_num1), \
'%10.5f(lr) '%cur_lr, \
'save'
if __name__ == "__main__":
train_use_tfrecords_data()