有时候,我们需要将TensorFlow的模型导出为单个文件(同时包含模型架构定义与权重),方便在其他地方使用(如在c++中部署网络)。利用tf.train.write_graph()默认情况下只导出了网络的定义(没有权重),而利用tf.train.Saver().save()导出的文件graph_def与权重是分离的,因此需要采用别的方法。
我们知道,graph_def文件中没有包含网络中的Variable值(通常情况存储了权重),但是却包含了constant值,所以如果我们能把Variable转换为constant,即可达到使用一个文件同时存储网络架构与权重的目标。
我们可以采用以下方式冻结权重并保存网络:
| 12
 3
 4
 5
 6
 7
 8
 9
 10
 11
 12
 13
 14
 15
 
 | import tensorflow as tffrom tensorflow.python.framework.graph_util import convert_variables_to_constants
 
 
 a = tf.Variable([[3],[4]], dtype=tf.float32, name='a')
 b = tf.Variable(4, dtype=tf.float32, name='b')
 
 output = tf.add(a, b, name='out')
 
 
 with tf.Session() as sess:
 sess.run(tf.global_variables_initializer())
 
 graph = convert_variables_to_constants(sess, sess.graph_def, ["out"])
 tf.train.write_graph(graph, '.', 'graph.pb', as_text=False)
 
 | 
当恢复网络时,可以使用如下方式:
| 12
 3
 4
 5
 6
 7
 
 | import tensorflow as tfwith tf.Session() as sess:
 with open('./graph.pb', 'rb') as f:
 graph_def = tf.GraphDef()
 graph_def.ParseFromString(f.read())
 output = tf.import_graph_def(graph_def, return_elements=['out:0'])
 print(sess.run(output))
 
 | 
输出结果为:
| 12
 
 | [array([[ 7.],[ 8.]], dtype=float32)]
 
 | 
可以看到之前的权重确实保存了下来!!
问题来了,我们的网络需要能有一个输入自定义数据的接口啊!不然这玩意有什么用。。别急,当然有办法。
| 12
 3
 4
 5
 6
 7
 8
 9
 10
 11
 12
 
 | import tensorflow as tffrom tensorflow.python.framework.graph_util import convert_variables_to_constants
 
 a = tf.Variable([[3],[4]], dtype=tf.float32, name='a')
 b = tf.Variable(4, dtype=tf.float32, name='b')
 input_tensor = tf.placeholder(tf.float32, name='input')
 output = tf.add((a+b), input_tensor, name='out')
 
 with tf.Session() as sess:
 sess.run(tf.global_variables_initializer())
 graph = convert_variables_to_constants(sess, sess.graph_def, ["out"])
 tf.train.write_graph(graph, '.', 'graph.pb', as_text=False)
 
 | 
用上述代码重新保存网络至graph.pb,这次我们有了一个输入placeholder,下面来看看怎么恢复网络并输入自定义数据。
| 12
 3
 4
 5
 6
 7
 8
 
 | import tensorflow as tf
 with tf.Session() as sess:
 with open('./graph.pb', 'rb') as f:
 graph_def = tf.GraphDef()
 graph_def.ParseFromString(f.read())
 output = tf.import_graph_def(graph_def, input_map={'input:0':4.}, return_elements=['out:0'], name='a')
 print(sess.run(output))
 
 | 
输出结果为:
| 12
 
 | [array([[ 11.],[ 12.]], dtype=float32)]
 
 | 
可以看到结果没有问题,当然在input_map那里可以替换为新的自定义的placeholder,如下所示:
| 12
 3
 4
 5
 6
 7
 8
 9
 10
 
 | import tensorflow as tf
 new_input = tf.placeholder(tf.float32, shape=())
 
 with tf.Session() as sess:
 with open('./graph.pb', 'rb') as f:
 graph_def = tf.GraphDef()
 graph_def.ParseFromString(f.read())
 output = tf.import_graph_def(graph_def, input_map={'input:0':new_input}, return_elements=['out:0'], name='a')
 print(sess.run(output, feed_dict={new_input:4}))
 
 | 
看看输出,同样没有问题。
| 12
 
 | [array([[ 11.],[ 12.]], dtype=float32)]
 
 | 
另外需要说明的一点是,在利用tf.train.write_graph写网络架构的时候,如果令as_text=True了,则在导入网络的时候,需要做一点小修改。
| 12
 3
 4
 5
 6
 7
 8
 9
 10
 11
 
 | import tensorflow as tffrom google.protobuf import text_format
 
 with tf.Session() as sess:
 
 with open('./graph.pb', 'r') as f:
 graph_def = tf.GraphDef()
 
 text_format.Merge(f.read(), graph_def)
 output = tf.import_graph_def(graph_def, return_elements=['out:0'])
 print(sess.run(output))
 
 | 
参考资料
Is
there an example on how to generate protobuf files holding trained
Tensorflow graphs