有没有办法让 tensorflow tf.Print 输出出现在 Jupyter Notebook 输出中

Question

我在 Jupyter 笔记本中使用 tf.Print 操作。它按要求工作，但只会将输出打印到控制台，而不是在笔记本中打印。有没有办法解决这个问题？

一个例子如下（在笔记本中）：

import tensorflow as tf

a = tf.constant(1.0)
a = tf.Print(a, [a], 'hi')
sess = tf.Session()
a.eval(session=sess)

该代码将在控制台中打印“hi[1]”，但在笔记本中没有任何内容。

Answer 1

2017 年 2 月 3 日更新 我已将其包装到 memory_util 包中。示例用法

# install memory util
import urllib.request
response = urllib.request.urlopen("https://raw.githubusercontent.com/yaroslavvb/memory_util/master/memory_util.py")
open("memory_util.py", "wb").write(response.read())

import memory_util

sess = tf.Session()
a = tf.random_uniform((1000,))
b = tf.random_uniform((1000,))
c = a + b
with memory_util.capture_stderr() as stderr:
    sess.run(c.op)

print(stderr.getvalue())

** 旧东西**

您可以从 IPython 核心重用 FD redirector。 （马克·桑德勒的想法）

import os
import sys

STDOUT = 1
STDERR = 2

class FDRedirector(object):
    """ Class to redirect output (stdout or stderr) at the OS level using
        file descriptors.
    """ 

    def __init__(self, fd=STDOUT):
        """ fd is the file descriptor of the outpout you want to capture.
            It can be STDOUT or STERR.
        """
        self.fd = fd
        self.started = False
        self.piper = None
        self.pipew = None

    def start(self):
        """ Setup the redirection.
        """
        if not self.started:
            self.oldhandle = os.dup(self.fd)
            self.piper, self.pipew = os.pipe()
            os.dup2(self.pipew, self.fd)
            os.close(self.pipew)

            self.started = True

    def flush(self):
        """ Flush the captured output, similar to the flush method of any
        stream.
        """
        if self.fd == STDOUT:
            sys.stdout.flush()
        elif self.fd == STDERR:
            sys.stderr.flush()

    def stop(self):
        """ Unset the redirection and return the captured output. 
        """
        if self.started:
            self.flush()
            os.dup2(self.oldhandle, self.fd)
            os.close(self.oldhandle)
            f = os.fdopen(self.piper, 'r')
            output = f.read()
            f.close()

            self.started = False
            return output
        else:
            return ''

    def getvalue(self):
        """ Return the output captured since the last getvalue, or the
        start of the redirection.
        """
        output = self.stop()
        self.start()
        return output

import tensorflow as tf
x = tf.constant([1,2,3])
a=tf.Print(x, [x])

redirect=FDRedirector(STDERR)

sess = tf.InteractiveSession()
redirect.start();
a.eval();
print "Result"
print redirect.stop()

Answer 2

我遇到了同样的问题，并通过在我的笔记本中使用这样的函数来解决它：

def tf_print(tensor, transform=None):

    # Insert a custom python operation into the graph that does nothing but print a tensors value 
    def print_tensor(x):
        # x is typically a numpy array here so you could do anything you want with it,
        # but adding a transformation of some kind usually makes the output more digestible
        print(x if transform is None else transform(x))
        return x
    log_op = tf.py_func(print_tensor, [tensor], [tensor.dtype])[0]
    with tf.control_dependencies([log_op]):
        res = tf.identity(tensor)

    # Return the given tensor
    return res


# Now define a tensor and use the tf_print function much like the tf.identity function
tensor = tf_print(tf.random_normal([100, 100]), transform=lambda x: [np.min(x), np.max(x)])

# This will print the transformed version of the tensors actual value 
# (which was summarized to just the min and max for brevity)
sess = tf.InteractiveSession()
sess.run([tensor])
sess.close()

仅供参考，在我的自定义函数中使用记录器而不是调用“print”对我来说很神奇，因为标准输出通常由 jupyter 缓冲，并且在“Loss is Nan”类型的错误之前没有显示——这就是重点就我而言，首先使用该功能。

Answer 3

您可以查看启动jupyter notebook 的终端以查看消息。

import tensorflow as tf

tf.InteractiveSession()

a = tf.constant(1)
b = tf.constant(2)

opt = a + b
opt = tf.Print(opt, [opt], message="1 + 2 = ")

opt.eval()

在终端，我可以看到：

2018-01-02 23:38:07.691808: I tensorflow/core/kernels/logging_ops.cc:79] 1 + 2 = [3]

Answer 4

一个简单的方法，在常规 python 中尝试过，但还不是 jupyter。 os.dup2(sys.stdout.fileno(), 1) os.dup2(sys.stdout.fileno(), 2)

解释在这里：In python, how to capture the stdout from a c++ shared library to a variable

Answer 5

我遇到的问题是无法在 Tensorflow Graph 中运行会话，例如在训练或评估中。 这就是为什么使用sess.run(opt) 或opt.eval() 的选项对我来说不是解决方案。 最好的办法是使用tf.Print() 并将日志重定向到外部文件。 我使用临时文件执行此操作，然后将其传输到这样的常规文件：

STDERR=2
import os
import sys
import tempfile

class captured:
    def __init__(self, fd=STDERR):
        self.fd = fd
        self.prevfd = None
    def __enter__(self):
        t = tempfile.NamedTemporaryFile()
        self.prevfd = os.dup(self.fd)
        os.dup2(t.fileno(), self.fd)
        return t
    def __exit__(self, exc_type, exc_value, traceback):
        os.dup2(self.prevfd, self.fd)

with captured(fd=STDERR) as tmp:
    ...
    classifier.evaluate(input_fn=input_fn, steps=100)

with open('log.txt', 'w') as f:
    print(open(tmp.name).read(), file=f)

然后在我的评估中：

a = tf.constant(1)
a = tf.Print(a, [a], message="a: ")