Fix for #12537

[jzju]

I have built the fix on x86 linux no gpu and it's working.

There are some android specific code in logging that I ignored.
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/platform/default/logging.cc#L35-L75

[tensorflow-jenkins]

Can one of the admins verify this patch?

[googlebot]

Thanks for your pull request. It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

📝 Please visit https://cla.developers.google.com/ to sign.

Once you've signed, please reply here (e.g. I signed it!) and we'll verify. Thanks.

---

• If you've already signed a CLA, it's possible we don't have your GitHub username or you're using a different email address. Check your existing CLA data and verify that your email is set on your git commits.
• If your company signed a CLA, they designated a Point of Contact who decides which employees are authorized to participate. You may need to contact the Point of Contact for your company and ask to be added to the group of authorized contributors. If you don't know who your Point of Contact is, direct the project maintainer to go/cla#troubleshoot.
• In order to pass this check, please resolve this problem and have the pull request author add another comment and the bot will run again.

[jzju]

I signed it!

[googlebot]

CLAs look good, thanks!

[vrv]

I am going to defer to API review first, and possibly someone else who works on TF core team should be a real reviewer.

[vrv]

The concerns I have are; API bloat, and internally our logging infrastructure under the LOG macros do a lot of extra helpful things relevant to Google-infrastructure that std:cerr would lose. I don't know how to rectify these things easily, even though I like the spirit of your change.

Perhaps this is a candidate for a custom op; it's not clear that we can make "Print" work generically for all environments in the world, and trying will just make Print more complicated than it needs to be.

[jzju]

Another solution would be to have an environment variable like TF_CPP_MIN_LOG_LEVEL that modifies the underlying fprintf

tensorflow/tensorflow/core/platform/default/logging.cc

Lines 88 to 91 in 4ad1204

	// TODO(jeff,sanjay): Replace this with something that logs through the env.
	fprintf(stderr, "%s.%06d: %c %s:%d] %s\n", time_buffer, micros_remainder,
	"IWEF"[severity_], fname_, line_, str().c_str());
	}

Also is it a bug or feature that TF_CPP_MIN_LOG_LEVEL can disable tf.Print?

[vrv]

@asimshankar and I chatted offline, there may be a simpler solution to this, but he'll have to pass it by the API review folks first.

It's a feature to allow users to completely disable all logging via LOG(INFO), and perhaps a bug that tf.Print uses LOG(INFO) in the OSS build.

[asimshankar]

@Johanju : Thanks for the contribution. We feel that instead of adding a new API, it might be okay to just change the implementation of the Print kernel to always use std::cerr instead of LOG(INFO).

Thoughts?

[jzju]

@asimshankar This is my first PR so I will just follow your lead, It won't be any problem the Google-infrastructure angel that @vrv talked about? What do you think is better between std::cerr and std::cout in this case?

[asimshankar]

@Johanju : I think it is fine, @vrv 's concerns are valid, but I believe we have a workaround to integrate better with our internal logging system.

For now I'd stick with std::cerr, since the current behavior with LOG(INFO) is to write to stderr as well (IIUC).

Thanks very much!

[asimshankar]

Thanks!