tf.print can't print log in jupyter notebook.

[lyhue1991]

When I tried to use tf.print() or get the log_device_placement information in google colabratory. I just can't get the print output. I tried in my own jupyter, and I find the print log is in the cmd.exe windows. This behavior is very strange and I think it should be fixed. The code is like below.

import tensorflow as tf
print(tf.GIT_VERSION, tf.VERSION)
a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[2, 3], name='a')
b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], shape=[3, 2], name='b')
c = tf.matmul(a, b)
sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))
print(sess.run(c))

the print output is just like this:

v1.12.0-0-ga6d8ffae09 1.12.0
[[22. 28.]
[49. 64.]]

but it should be with the log_device_placement information.

[ymodak]

By adding tf.logging.set_verbosity(tf.logging.INFO) in your script you should be able to see the logging in the ipython console.

[lyhue1991]

By adding tf.logging.set_verbosity(tf.logging.INFO) in your script you should be able to see the logging in the ipython console.

Thank you very much. But I tried this, it does't work. The reason of this problem maybe in the cooperation of tensorflow and Jupyter notebook. Because I can see the log information in the cmd.exe window which runs jupyter.

[facaiy]

I think you're right, @lyhue1991 . It's a known limitation of tf.print, see RFC:

https://github.com/tensorflow/community/pull/14/files#diff-6dc73e00aed7c8a7fbc0f53d7981f296R365

[lyhue1991]

@facaiy Thank you very much. This limitation is really painful.

[facaiy]

@tomerk Hi, Tomer. Could you answer the question?

[tomerk]

Hi @facaiy what question still needs answering?
As you mentioned this is still a known limitation, because Jupyter notebooks don't capture C++ output.

We've decided to not add python-specific logic at this time (which we could do by running py_funcs internally) because it would prevent universal exporting of models that use tf.print

I've bumped the relevant issue on the ipython kernel's github:
ipython/ipykernel#110

For anyone who needs a workaround there's a number of scattered code on the internet that tries to redirect C++ outputs to ipython, e.g.
https://github.com/minrk/wurlitzer

[facaiy]

Thanks for your detailed explanation, and workaround :-) So can we close the issue, @lyhue1991 ?