还是要从下图开始。 到目前为止，我实际上还不知道。
固态硬盘的网络构成流程如下图所示。
SSD共有11个块，与以前的VGG16相比，改变了第5块的第4层，去除了第6、7、8个卷积层，分别增加了红框、黑框、黄框、蓝框。
tensorflow代码如下：
withTF.variable_scope(scope，' ssd_300_vgg '，[inputs]，reuse=reuse ) 3360#originalvgg-16blocks.net 3 scope='conv1 ' ) end _ points [ ' block1' ]=netnet=slim.max _ pool 2d (net，[ 2，2 ]，scope='pool1' ) ) scope='conv2 ' ) end _ points [ ' block2' ]=netnet=slim.max _ pool 2d (net，[ 2，2 ]，scope=slim.conv2d，2d ) scope=' con v3 ' ] end _ points [ ' block3' ]=netnet=slim.max _ pool 2d ([ net，[2] scope='pool3' ) # block 4 scope='conv4 ' ) end _ pool [ 2，2 ]，scope='pool4' ) #block5.net=slim.repeat(net，3，slim.conv2d，512 scope=#='conv5 ' ) end_points['block5']=net #注意事项net=slim.max_pool2d(net，[ 3，3 ]，stride=1，scope scope='conv6 ' ) end_points['block6']=netnet=TF.)。training=is _ training (# block 733601 x1 conv.becauseuses scope='conv7 ' ) end _ points [ ' block7' ]=netnet=TF.layers.dropout (net，rate=dropout_keep_prob， training=is _ training (# block8/9/10/11:1 x1 an d3x3convolution sstride2(except lasts ).end _ point=' block83365306; scope='conv1x1 ' )注意事项：实际上相当于在下一个卷积操作中填充了net=custom_layers.pad220) net=slim.conv2d(net，512，[ padding='VALID ' ) end _ points [ end _ point ]=netend _ point=' block9' withtf.variable _ scope (end _ point soint ) 注意事项：实际上在下一个卷积操作中填充的net=custom_layers.pad2d(net，pad=(1，1 ) ) net=slim.conv2d ) ) net，pad=(1，1 ) padding='VALID ' ) end _ points [ end _ point ]=netend _ point=' block 10 ' withtf.variable 128，[ 1，1 ]，scope padding='VALID ' ) end _ points [ end _ point ]=netend _ point=' block 11 ' withtf.variable _ scope (end _ point )