ckpt模型转pb是项目落地的关键，如果你仅仅是想在服务器部署，也可忽略此步。人生的艰难有很多坑，不从坑里爬出来体会不到生命中的乐趣所在。简单的说，原来ckpt模型有4个文件（分别为checkpoint,index,data,meta），转成pb后就一个文件，至少方便了很多。然而其大小并无明显变化，甚至没有变化，想要有明显变化，还需要转tf lite，但首先第一步是转成pb格式。
pb格式的转化其实也有很多参考，然后不同模型的inference写起来可能不是那么简单，在其中都是要定义一个入口和一个出口，也就是必需input和output，这样图结构才可以方便的拿起来用。本人先前所见的都是分类模型的pb转化，未见有回归模型的pb转化，有人说这不都一样吗？你别瞎逼逼，你先甩给我个回归的链接，我跪谢你。
其中的难点在于回归的inference中output不好定义，而分类的只是定义一个softmax就OK了。目前只有发挥我的聪明才智，或者广泛搜索了。
人生不易，请你珍惜眼前的我。随时更新，欢迎关注。
补充一：寡人得到了模型的节点名字，但是也不知道该怎么定义输出，因为模型中的输出结果并不是想要的直接输出，还要经过运算，而这种运算不是tf定义的，是不是先定义下tf的加减乘除为好？
tensor_name: beta1_power
tensor_name: beta2_power
tensor_name: net/rnn/multi_rnn_cell/cell_0/gru_cell/candidate/bias/Adam
tensor_name: net/rnn/multi_rnn_cell/cell_0/gru_cell/candidate/bias/Adam_1
tensor_name: net/rnn/multi_rnn_cell/cell_0/gru_cell/candidate/kernel/Adam
tensor_name: net/rnn/multi_rnn_cell/cell_0/gru_cell/candidate/kernel/Adam_1
tensor_name: net/y_hat_src2/kernel/Adam
tensor_name: net/y_hat_src2/bias/Adam
tensor_name: net/rnn/multi_rnn_cell/cell_1/gru_cell/gates/kernel/Adam
tensor_name: net/rnn/multi_rnn_cell/cell_1/gru_cell/gates/bias/Adam
tensor_name: global_step
tensor_name: net/rnn/multi_rnn_cell/cell_0/gru_cell/candidate/bias
tensor_name: net/rnn/multi_rnn_cell/cell_0/gru_cell/candidate/kernel
tensor_name: net/rnn/multi_rnn_cell/cell_2/gru_cell/gates/kernel/Adam_1
tensor_name: net/rnn/multi_rnn_cell/cell_2/gru_cell/gates/kernel/Adam
tensor_name: net/rnn/multi_rnn_cell/cell_0/gru_cell/gates/bias
tensor_name: net/rnn/multi_rnn_cell/cell_0/gru_cell/gates/bias/Adam
tensor_name: net/rnn/multi_rnn_cell/cell_2/gru_cell/gates/kernel
tensor_name: net/rnn/multi_rnn_cell/cell_0/gru_cell/gates/bias/Adam_1
tensor_name: net/rnn/multi_rnn_cell/cell_0/gru_cell/gates/kernel
tensor_name: net/rnn/multi_rnn_cell/cell_0/gru_cell/gates/kernel/Adam
tensor_name: net/rnn/multi_rnn_cell/cell_0/gru_cell/gates/kernel/Adam_1
tensor_name: net/rnn/multi_rnn_cell/cell_1/gru_cell/candidate/kernel
tensor_name: net/rnn/multi_rnn_cell/cell_1/gru_cell/candidate/bias
tensor_name: net/y_hat_src1/kernel/Adam_1
tensor_name: net/rnn/multi_rnn_cell/cell_1/gru_cell/candidate/kernel/Adam
tensor_name: net/rnn/multi_rnn_cell/cell_1/gru_cell/candidate/bias/Adam
tensor_name: net/y_hat_src1/kernel/Adam
tensor_name: net/rnn/multi_rnn_cell/cell_1/gru_cell/candidate/kernel/Adam_1
tensor_name: net/rnn/multi_rnn_cell/cell_1/gru_cell/candidate/bias/Adam_1
tensor_name: net/y_hat_src2/kernel
tensor_name: net/y_hat_src2/bias
tensor_name: net/rnn/multi_rnn_cell/cell_1/gru_cell/gates/kernel
tensor_name: net/rnn/multi_rnn_cell/cell_1/gru_cell/gates/bias
tensor_name: net/y_hat_src2/kernel/Adam_1
tensor_name: net/y_hat_src2/bias/Adam_1
tensor_name: net/rnn/multi_rnn_cell/cell_1/gru_cell/gates/kernel/Adam_1
tensor_name: net/rnn/multi_rnn_cell/cell_1/gru_cell/gates/bias/Adam_1
tensor_name: net/rnn/multi_rnn_cell/cell_2/gru_cell/candidate/bias
tensor_name: net/rnn/multi_rnn_cell/cell_2/gru_cell/candidate/bias/Adam
tensor_name: net/rnn/multi_rnn_cell/cell_2/gru_cell/candidate/bias/Adam_1
tensor_name: net/rnn/multi_rnn_cell/cell_2/gru_cell/candidate/kernel
tensor_name: net/rnn/multi_rnn_cell/cell_2/gru_cell/candidate/kernel/Adam
tensor_name: net/rnn/multi_rnn_cell/cell_2/gru_cell/candidate/kernel/Adam_1
tensor_name: net/y_hat_src1/bias/Adam_1
tensor_name: net/rnn/multi_rnn_cell/cell_2/gru_cell/gates/bias
tensor_name: net/rnn/multi_rnn_cell/cell_2/gru_cell/gates/bias/Adam
tensor_name: net/y_hat_src1/bias
tensor_name: net/rnn/multi_rnn_cell/cell_2/gru_cell/gates/bias/Adam_1
tensor_name: net/y_hat_src1/bias/Adam
tensor_name: net/y_hat_src1/kernel
节点名字应该重新用tf定义。
补充二：刚发现，input的定义也不是简单的直接输入数据就对了，而应该是fft后的结果，整个网络的处理是针对fft来做的，因此输出的也是fft，中间要保留相位信息，用于恢复之用，掩蔽在网络结构外做。现在考虑的是到底是直接用tf来求fft还是直接传入fft后的结果？不妨都试试看了。这个真是不简单啊。
补充三：苦苦寻觅，就是找不到出口在哪里？如何定义真是麻烦。网络中似乎在训练时就缺少了某些节点，而且也没默认节点，因为不全是tf定义的运算，应该重新从网络设置上找结果，或许应该重新写一下网络。不妨试试看。
另外有相关问题可以加入QQ群讨论，不设微信群
QQ群：868373192
语音深度学习群