Tensorflow为张量创建Protobuf 我想有一个Python脚本，它将NoPy数组转换成TensorFlow Tensors，在C++中我可以重新加载它们。这可以通过计算图来完成，如

我在TensorFlow Python API中发现了以下函数和特性

C++具有相应的

---

你能给我一个将TF张量序列化为Protobuf二进制并返回的例子吗？

我会在找到答案后发布，这样也许有人可以参与解决方案的其余部分

Python

张量->Protobuf二进制

>>> import tensorflow as tf
>>> with tf.Graph().as_default():
...     s = tf.constant([1.2, 3.4, 5.6, 7.8])._op.node_def.attr['value'].SerializeToString()
...
>>> s
'B\x1a\x08\x01\x12\x04\x12\x02\x08\x04"\x10\x9a\x99\x99?\x9a\x99Y@33\xb3@\x9a\x99\xf9@'

Protobuf二进制->张量

>>> import tensorflow as tf
>>> s = 'B\x1a\x08\x01\x12\x04\x12\x02\x08\x04"\x10\x9a\x99\x99?\x9a\x99Y@33\xb3@\x9a\x99\xf9@'
>>> with tf.Graph().as_default():
...     c = tf.constant(1)
...     c._op.node_def.attr['value'].ParseFromString(s)
...     c._op.node_def.attr['dtype'].type = c._op.node_def.attr['value'].tensor.dtype
...     print c.eval(session=tf.Session())
... 
28
[ 1.20000005  3.4000001   5.5999999   7.80000019]

基准

   Array Elements  from_string [us]  to_string [us]
0              10         10.273593        2.308139
1             100         10.450414        2.291126
2            1000         10.540897        2.359392
3           10000         12.175265        2.734819
4          100000         31.460438        7.349958

基准脚本

import tensorflow as tf
import pandas as pd
import numpy as np
import timeit
import matplotlib.pyplot as plt

def to_string(shape):
    with tf.Graph().as_default():
        s = tf.constant(np.empty(shape))._op.node_def.attr['value'].SerializeToString()
    return s


def from_string(s):
    with tf.Graph().as_default():
        c = tf.constant(1)
        c._op.node_def.attr['value'].ParseFromString(s)
        c._op.node_def.attr['dtype'].type = c._op.node_def.attr['value'].tensor.dtype
        c.eval(session=tf.Session())

NUM_RUNS = 10000
MAX_POW = 6

print "Collecting to_string stats"
to_string_results = np.array([[N, timeit.timeit('to_string((%d,))' % N,
                                                setup="from __main__ import to_string",
                                                number=NUM_RUNS)]
                              for N in 10**np.arange(1, MAX_POW)]).T

print "Collecting from_string stats"
strings = {N:to_string((N,)) for N in 10**np.arange(1, MAX_POW)}
from_string_results = np.array([[N, timeit.timeit('from_string(strings[%d])' % N,
                                                  setup="from __main__ import from_string, strings",
                                                  number=NUM_RUNS)]
                                for N in 10**np.arange(1, MAX_POW)]).T

df = pd.DataFrame.from_dict({"Array Elements": to_string_results[0],
                             "to_string [us]": to_string_results[1],
                             "from_string [us]": from_string_results[1]})

print df
df.to_csv('benchmark.csv')

plt.subplot(2, 1, 1)
plt.loglog(to_string_results[0], to_string_results[1])
plt.title('to_string')
plt.ylabel('microseconds')
plt.xlabel('Array size')
plt.subplot(2, 1, 2)
plt.loglog(from_string_results[0], from_string_results[1])
plt.title('from_string')
plt.ylabel('microseconds')
plt.xlabel('Array size')
plt.show()

C++

---

正在研究它

我会在找到答案后发布答案，这样也许有人可以参与解决方案的其余部分

Python

张量->Protobuf二进制

>>> import tensorflow as tf
>>> with tf.Graph().as_default():
...     s = tf.constant([1.2, 3.4, 5.6, 7.8])._op.node_def.attr['value'].SerializeToString()
...
>>> s
'B\x1a\x08\x01\x12\x04\x12\x02\x08\x04"\x10\x9a\x99\x99?\x9a\x99Y@33\xb3@\x9a\x99\xf9@'

Protobuf二进制->张量

>>> import tensorflow as tf
>>> s = 'B\x1a\x08\x01\x12\x04\x12\x02\x08\x04"\x10\x9a\x99\x99?\x9a\x99Y@33\xb3@\x9a\x99\xf9@'
>>> with tf.Graph().as_default():
...     c = tf.constant(1)
...     c._op.node_def.attr['value'].ParseFromString(s)
...     c._op.node_def.attr['dtype'].type = c._op.node_def.attr['value'].tensor.dtype
...     print c.eval(session=tf.Session())
... 
28
[ 1.20000005  3.4000001   5.5999999   7.80000019]

基准

   Array Elements  from_string [us]  to_string [us]
0              10         10.273593        2.308139
1             100         10.450414        2.291126
2            1000         10.540897        2.359392
3           10000         12.175265        2.734819
4          100000         31.460438        7.349958

基准脚本

import tensorflow as tf
import pandas as pd
import numpy as np
import timeit
import matplotlib.pyplot as plt

def to_string(shape):
    with tf.Graph().as_default():
        s = tf.constant(np.empty(shape))._op.node_def.attr['value'].SerializeToString()
    return s


def from_string(s):
    with tf.Graph().as_default():
        c = tf.constant(1)
        c._op.node_def.attr['value'].ParseFromString(s)
        c._op.node_def.attr['dtype'].type = c._op.node_def.attr['value'].tensor.dtype
        c.eval(session=tf.Session())

NUM_RUNS = 10000
MAX_POW = 6

print "Collecting to_string stats"
to_string_results = np.array([[N, timeit.timeit('to_string((%d,))' % N,
                                                setup="from __main__ import to_string",
                                                number=NUM_RUNS)]
                              for N in 10**np.arange(1, MAX_POW)]).T

print "Collecting from_string stats"
strings = {N:to_string((N,)) for N in 10**np.arange(1, MAX_POW)}
from_string_results = np.array([[N, timeit.timeit('from_string(strings[%d])' % N,
                                                  setup="from __main__ import from_string, strings",
                                                  number=NUM_RUNS)]
                                for N in 10**np.arange(1, MAX_POW)]).T

df = pd.DataFrame.from_dict({"Array Elements": to_string_results[0],
                             "to_string [us]": to_string_results[1],
                             "from_string [us]": from_string_results[1]})

print df
df.to_csv('benchmark.csv')

plt.subplot(2, 1, 1)
plt.loglog(to_string_results[0], to_string_results[1])
plt.title('to_string')
plt.ylabel('microseconds')
plt.xlabel('Array size')
plt.subplot(2, 1, 2)
plt.loglog(from_string_results[0], from_string_results[1])
plt.title('from_string')
plt.ylabel('microseconds')
plt.xlabel('Array size')
plt.show()

C++

---

使用它

我已经测试了from_string函数，它使用TensorFlow r1.7工作，但使用r1.9不再工作

ParseFromString

函数似乎不再更新张量。遵循我的变通方法

def from_string(s):
    with tf.Graph().as_default():
        c_tpb = tf.constant(1)._op.node_def.attr['value'].FromString(s)
        c = tf.constant(c_tpb.tensor)
        c.eval(session=tf.Session())

---

我已经测试了来自的

from_string

函数，它使用TensorFlow r1.7工作，但使用r1.9不再工作

ParseFromString

函数似乎不再更新张量。遵循我的变通方法

def from_string(s):
    with tf.Graph().as_default():
        c_tpb = tf.constant(1)._op.node_def.attr['value'].FromString(s)
        c = tf.constant(c_tpb.tensor)
        c.eval(session=tf.Session())