从字符串解析 graph_def 时出错

Question

我正在尝试运行一个非常简单的 Tensorflow 图保存为 .pb 文件，但在解析它时出现此错误：

Traceback (most recent call last):
  File "test_import_stripped_bm.py", line 28, in <module>
    graph_def.ParseFromString(fileContent)
  File "/usr/local/lib/python3.5/dist-packages/google/protobuf/message.py", line 185, in ParseFromString
    self.MergeFromString(serialized)
  File "/usr/local/lib/python3.5/dist-packages/google/protobuf/internal/python_message.py", line 1069, in MergeFromString
    if self._InternalParse(serialized, 0, length) != length:
  File "/usr/local/lib/python3.5/dist-packages/google/protobuf/internal/python_message.py", line 1105, in InternalParse
    pos = field_decoder(buffer, new_pos, end, self, field_dict)
  File "/usr/local/lib/python3.5/dist-packages/google/protobuf/internal/decoder.py", line 633, in DecodeField
    if value._InternalParse(buffer, pos, new_pos) != new_pos:
  File "/usr/local/lib/python3.5/dist-packages/google/protobuf/internal/python_message.py", line 1105, in InternalParse
    pos = field_decoder(buffer, new_pos, end, self, field_dict)
  File "/usr/local/lib/python3.5/dist-packages/google/protobuf/internal/decoder.py", line 612, in DecodeRepeatedField
    if value.add()._InternalParse(buffer, pos, new_pos) != new_pos:
  File "/usr/local/lib/python3.5/dist-packages/google/protobuf/internal/python_message.py", line 1105, in InternalParse
    pos = field_decoder(buffer, new_pos, end, self, field_dict)
  File "/usr/local/lib/python3.5/dist-packages/google/protobuf/internal/decoder.py", line 743, in DecodeMap
    if submsg._InternalParse(buffer, pos, new_pos) != new_pos:
  File "/usr/local/lib/python3.5/dist-packages/google/protobuf/internal/python_message.py", line 1095, in InternalParse
    new_pos = local_SkipField(buffer, new_pos, end, tag_bytes)
  File "/usr/local/lib/python3.5/dist-packages/google/protobuf/internal/decoder.py", line 850, in SkipField
    return WIRETYPE_TO_SKIPPER[wire_type](buffer, pos, end)
  File "/usr/local/lib/python3.5/dist-packages/google/protobuf/internal/decoder.py", line 799, in _SkipGroup
    new_pos = SkipField(buffer, pos, end, tag_bytes)
  File "/usr/local/lib/python3.5/dist-packages/google/protobuf/internal/decoder.py", line 850, in SkipField
    return WIRETYPE_TO_SKIPPER[wire_type](buffer, pos, end)
  File "/usr/local/lib/python3.5/dist-packages/google/protobuf/internal/decoder.py", line 814, in _SkipFixed32
    raise _DecodeError('Truncated message.')
google.protobuf.message.DecodeError: Truncated message.

这是我用来将其写入 .pb 的代码：

import tensorflow as tf

builder = tf.saved_model.builder.SavedModelBuilder('models/TEST-3')

w1 = tf.Variable(tf.random_normal((2,2)), name="w1")
w2 = tf.Variable(tf.random_normal((2,2)), name="w2")

sess = tf.Session()
sess.run(tf.global_variables_initializer())

builder.add_meta_graph_and_variables(sess, tags=[tf.saved_model.tag_constants.SERVING], clear_devices = True)

builder.save()
sess.close()

这是解析它的代码：

import tensorflow as tf
import os

model_path = os.path.join('models/TEST-3', 'saved_model.pb')
with open(model_path, mode='rb') as f:
    fileContent = f.read()
graph_def = tf.GraphDef()
graph_def.ParseFromString(fileContent)

要查看我必须做的确切错误

export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python

在运行它之前。 我也在 python 2 和 3 上用不同的 tensorflow 版本尝试过这个，我在 Ubuntu 16.04 上运行。在带有 tensorflow 0.9.0rc0 的 python 2.7 上，我设法得到了一个稍微不同的错误：

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python2.7/dist-packages/google/protobuf/message.py", line 185, in ParseFromString
    self.MergeFromString(serialized)
  File "/usr/local/lib/python2.7/dist-packages/google/protobuf/internal/python_message.py", line 1091, in MergeFromString
    if self._InternalParse(serialized, 0, length) != length:
  File "/usr/local/lib/python2.7/dist-packages/google/protobuf/internal/python_message.py", line 1127, in InternalParse
    pos = field_decoder(buffer, new_pos, end, self, field_dict)
  File "/usr/local/lib/python2.7/dist-packages/google/protobuf/internal/decoder.py", line 633, in DecodeField
    if value._InternalParse(buffer, pos, new_pos) != new_pos:
  File "/usr/local/lib/python2.7/dist-packages/google/protobuf/internal/python_message.py", line 1127, in InternalParse
    pos = field_decoder(buffer, new_pos, end, self, field_dict)
  File "/usr/local/lib/python2.7/dist-packages/google/protobuf/internal/decoder.py", line 612, in DecodeRepeatedField
    if value.add()._InternalParse(buffer, pos, new_pos) != new_pos:
  File "/usr/local/lib/python2.7/dist-packages/google/protobuf/internal/python_message.py", line 1127, in InternalParse
    pos = field_decoder(buffer, new_pos, end, self, field_dict)
  File "/usr/local/lib/python2.7/dist-packages/google/protobuf/internal/decoder.py", line 612, in DecodeRepeatedField
    if value.add()._InternalParse(buffer, pos, new_pos) != new_pos:
  File "/usr/local/lib/python2.7/dist-packages/google/protobuf/internal/python_message.py", line 1127, in InternalParse
    pos = field_decoder(buffer, new_pos, end, self, field_dict)
  File "/usr/local/lib/python2.7/dist-packages/google/protobuf/internal/decoder.py", line 489, in DecodeRepeatedField
    value.append(_ConvertToUnicode(buffer[pos:new_pos]))
  File "/usr/local/lib/python2.7/dist-packages/google/protobuf/internal/decoder.py", line 469, in _ConvertToUnicode
    return local_unicode(byte_str, 'utf-8')
UnicodeDecodeError: 'utf8' codec can't decode byte 0x80 in position 18: 'utf8' codec can't decode byte 0x80 in position 18: invalid start byte in field: tensorflow.FunctionDef.Node.ret

我可以使用此代码解析其他 .pb 图形，例如 https://github.com/taey16/tf/blob/master/imagenet/classify_image_graph_def.pb

提前致谢。

Answer 1

这里的问题是你试图解析一个SavedModel 协议缓冲区，就好像它是一个GraphDef。尽管SavedModel 包含GraphDef，但它们具有不同的二进制格式。以下代码，使用tf.saved_model.loader.load() 应该可以工作：

import tensorflow as tf

with tf.Session(graph=tf.Graph()) as sess:
    tf.saved_model.loader.load(
        sess, [tf.saved_model.tag_constants.SERVING], "models/TEST-3") 

Answer 2

graph_def = tf.GraphDef()
graph_def.ParseFromString(fileContent)

这里的 fileContent 应该是一个**“Frozen Graph”。 Tensorflow 也提供了相同的 api，参考Tensorflow freeze_graph API

另一种创建冻结图的方法是：

with tf.Session(graph=tf.Graph()) as sess:
      saver = tf.train.import_meta_graph(<.meta file>)
      saver.restore(sess, <checkpoint>)
      output_graph_def = tf.graph_util.convert_variables_to_constants(
                    sess,
                    tf.get_default_graph().as_graph_def(),
                    [comma separated output nodes name]
                ) 
      # Saving "output_graph_def " in a file and generate frozen graph.
      with tf.gfile.GFile('frozen_graph.pb', "wb") as f:
      f.write(output_graph_def.SerializeToString())

使用 freeze_graph.pb 作为

graph_def.ParseFromString("frozen_graph.pb")

所以首先使用 Saver 对象生成 .meta 和其他文件。完成后创建冻结图。