Caffe模型转TensorFlow模型并预测

在实际的生产环境中，有时候我们需要caffe模型转换为TensorFlow模型用于部署，甚至支持手机端（主要由于tensorflow有c++、python 的api支持手机端）。本文主要是介绍开源库caffe-tensorflow的使用。
python2版本：https://github.com/ethereon/caffe-tensorflow
python3版本：https://github.com/dhaase-de/caffe-tensorflow-python3

环境

• Python 2.7 (required by caffe-tensorflow)
• TensorFlow 1.x installed (tested with TensorFlow 1.9)
• caffe(no need，but required if convert the mean files)

转换过程

可以参考examples/mnist中转换示例，本文以转换一个年龄预测的模型为例：
1、convert a prototxt model to tensorflow code::

python ./convert.py /path/to/age.prototxt --code-output-path=age_net.py

生成的age_net.py如下：

from network import Network

class AgeNet(Network):
    def setup(self):
        (self.feed('data')
             .conv(3, 3, 64, 1, 1, name='conv1_1')
             .conv(3, 3, 64, 1, 1, name='conv1_2')
             .max_pool(2, 2, 2, 2, name='pool1')
             .conv(3, 3, 128, 1, 1, name='conv2_1')
             .conv(3, 3, 128, 1, 1, name='conv2_2')
             .max_pool(2, 2, 2, 2, name='pool2')
             .conv(3, 3, 256, 1, 1, name='conv3_1')
             .conv(3, 3, 256, 1, 1, name='conv3_2')
             .conv(3, 3, 256, 1, 1, name='conv3_3')
             .max_pool(2, 2, 2, 2, name='pool3')
             .conv(3, 3, 512, 1, 1, name='conv4_1')
             .conv(3, 3, 512, 1, 1, name='conv4_2')
             .conv(3, 3, 512, 1, 1, name='conv4_3')
             .max_pool(2, 2, 2, 2, name='pool4')
             .conv(3, 3, 512, 1, 1, name='conv5_1')
             .conv(3, 3, 512, 1, 1, name='conv5_2')
             .conv(3, 3, 512, 1, 1, name='conv5_3')
             .max_pool(2, 2, 2, 2, name='pool5')
             .fc(4096, name='fc6')
             .fc(4096, name='fc7')
             .fc(101, relu=False, name='fc8-101')
             .softmax(name='prob'))

2、convert .caffemodel weight files to .npy files

python ./convert.py /path/to/age.prototxt --caffemodel /path/to/age_iter_300000.caffemodel --data-output-path=age_model.npy

3、Convert the model and the mean file
此步骤需要caffe环境，我的实验没有这步骤，如果有需要可以参考示例如下：
来源参考：https://blog.wildcat.io/2018/04/a-simple-tutorial-about-caffe-tensorflow-model-conversion/

import caffe
import numpy as np
import sys

if len(sys.argv) != 3:
    print("Usage: python convert_protomean.py proto.mean out.npy")

blob = caffe.proto.caffe_pb2.BlobProto()
data = open('./original-caffe-models/DB_train_w32_5.binaryproto', 'rb').read()
blob.ParseFromString(data)
arr = np.array(caffe.io.blobproto_to_array(blob))
out = arr[0]
np.save('mean.npy', out)

预测阶段

在预测阶段只需要四个文件即可
age_model.npy（生成的model文件）
network.py（来源于caffe-tensorflow/kaffe/tensorflow/network.py）
age_net.py（生成的网络文件）
age_predict.py（预测文件，可以参考examples/imagenet/classify.py，亦可参考代码如下）

#coding=utf-8
import os,cv2,time
from age_net import AgeNet
import numpy as np
import tensorflow as tf

class AgePredict:
	model_data_path = r'face_age/age_model.npy'

	def __init__(self):
		self.input_node = tf.placeholder(tf.float32,shape=(None, 224, 224, 3))
		self.net = AgeNet({'data': self.input_node})

		self.sess = tf.Session()
		with self.sess.as_default():
			# Load the converted parameters
			print('Loading the model')
			self.net.load(self.model_data_path, self.sess)

	def predict(self, im):
		img = cv2.resize(im,(224,224),cv2.INTER_NEAREST)
		img = img.astype(np.float32, copy=True)
		img -= np.array([[[104, 117, 123]]])

		output_prob = self.sess.run(self.net.get_output(), feed_dict={self.input_node: [img]})
		age = 0
		for idx, prob in enumerate(output_prob[0]):
			age += idx * prob
		return int(age)

	def main(self):
		pic_path = 'data/test_image.jpg'
		img = cv2.imread(pic_path)
		print(self.predict(img))

if __name__ == '__main__':
    agePredict = AgePredict()
    agePredict.main()

参考：https://my.oschina.net/yilian/blog/672135