我使用的caffe模型：https://github.com/BVLC/caffe/tree/ea455eb29393ebe6de9f14e88bfce9eae74edf6d/models/bvlc_alexnet
其中，需要下载deploy prototxt文件和caffmodel以供转换模型用。
首先将caffe模型转换为TensorFlow，参考：https://github.com/ethereon/caffe-tensorflow
但是将其生产pb模型参考了如下代码，见https://github.com/ethereon/caffe-tensorflow/issues/23
from googlenet import GoogleNet # the output python script of caffe2tensorflow import tensorflow as tf from freeze_graph import freeze_graph # tensorflow comes up with a tool allowing freeze graph x = tf.placeholder(tf.float32, shape=[1, 224, 224, 3]) y = tf.placeholder(tf.float32, shape=[1, 1000]) net = GoogleNet({'data': x}) sess = tf.InteractiveSession() sess.run(tf.initialize_all_variables()) net.load('googlenet.npy', sess) saver = tf.train.Saver() saver.save(sess, 'chkpt', global_step=0, latest_filename='chkpt_state') tf.train.write_graph(sess.graph.as_graph_def(), './', googlenet.pb', False) input_saver_def_path = '' input_binary=True input_checkpoint_path = 'chkpt-0' input_graph_path = 'googlenet.pb' output_graph_path = 'googlenet.pb' output_node_names = 'prob' restore_op_name = "save/restore_all" filename_tensor_name = "save/Const:0" clear_devices = True freeze_graph(input_graph_path, input_saver_def_path, input_binary, input_checkpoint_path, output_node_names, restore_op_name, filename_tensor_name, output_graph_path, clear_devices, "")
注意，input node name改变了，自己print graph就知道。
另外一个博客里 见 https://ndres.me/post/convert-caffe-to-tensorflow/ ，提到自己fork的caffe to TensorFlow可以生成pb，但是我自己尝试发现针对alexnet抛出了异常。因此就没有使用。
还有就是 https://www.cs.toronto.edu/~guerzhoy/tf_alexnet/ 里提到有现成的模型，原文如下：
AlexNet implementation + weights in TensorFlow
This is a quick and dirty AlexNet implementation in TensorFlow. You may also be interested in Davi Frossard's VGG16 code/weights.
- myalexnet_forward.py -- the implementation itself + testing code for versions of TensorFlow current in 2017 (Python 3).
- myalexnet_forward.py -- for older versions of TensorFlowm in Python 2(See this version for a variable rather than placeholder input; you probably want the myalexnet_forward.py version if you want to fine-tune the networks.)
- bvlc_alexnet.npy -- the weights; they need to be in the working directory
- caffe_classes.py -- the classes, in the same order as the outputs of the network
- poodle.png, laska.png, dog.png, dog2.png, quail227.JPEG -- test images (images should be 227x227x3)
Credits:
- The model and weights are from Caffe
- The weights were converted using caffe-tensorflow, and code was taken from there as well
但是我自己使用的时候发现会抛出异常，可能是因为我使用的TensorFlow 1.3版本问题。因此最后也没有使用。
其次，就是模型压缩，命令为：
~/tensorflow-master/transform_graph/tensorflow/tools/graph_transforms/transform_graph --in_graph=alexnet.pb --outputs="prob" --out_graph=quantized_alexnet.pb --transforms='quantize_weights'
最后，导入pb到andriod项目里，我参考了：http://blog.csdn.net/u014432647/article/details/74743332和https://github.com/ppplinday/alexnet-tensorflow-android-withoutfc67
注意我是自己bazel compile TensorFlow 1.0生成的arm64-v8a的inferance so和jar文件。
值得一提的是，在https://github.com/ppplinday/alexnet-tensorflow-android-withoutfc67里提到，其使用的模型在android红米机器上inference时间是0.713s，华为mate10上稍微快点，是0.6s。
贴下http://blog.csdn.net/u014432647/article/details/74743332 原文内容：
首先先说明一下这个alexnet是进过修改的网络，由于模型太大，我把alexnet的全连接层6和全连接7给删除了，测试手机是红米Note 4X，cpu是高通骁龙625。
分别放出测试代码的github
arm compute library：
https://github.com/ppplinday/alexnet-armcomputelibrary-android-withoutfc67
tensorflow：
https://github.com/ppplinday/alexnet-tensorflow-android-withoutfc67
测试结果为：
arm compute lilrary：0.422s
tensorflow：0.713s
arm compute library比tensorflow快了差不多百分之40。arm这个也是有一定的优化，但是现在还是有挺多bug，而且接口暂时只提供C++，可以等之后成熟了再使用。
大致过程记录在此，以备忘。