gRPC debug URL scheme support for Windows

[rvfischione]

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): Yes (See below)
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Windows 10
• TensorFlow installed from (source or binary): binary
• TensorFlow version (use command below): 1.6.0
• Python version: 3.6.1
• Bazel version (if compiling from source): N/A
• GCC/Compiler version (if compiling from source): N/A
• CUDA/cuDNN version: N/A
• GPU model and memory: N/A
• Exact command to reproduce: See below

Describe the problem

Incidentally, this is both a bug report and a feature request.

To reproduce the bug, run the following:

import tensorflow as tf
from tensorflow.python import debug as tf_debug

session = tf.Session()
session = tf_debug.TensorBoardDebugWrapperSession(session, "localhost:9898")

current_epoch = tf.Variable(0, trainable=False, name='current_epoch')
increment_epoch = tf.assign(current_epoch, current_epoch + 1)

start_epoch = current_epoch.eval(session)
print("start_epoch:", start_epoch)
for cur_epoch in range(start_epoch, 10):
    session.run(increment_epoch)

The following is the output:

Traceback (most recent call last):
  File "C:/workspace/test.py", line 12, in <module>
    for cur_epoch in range(start_epoch, 10):
TypeError: 'UnimplementedError' object cannot be interpreted as an integer
start_epoch: grpc:// debug URL scheme is not implemented on Windows yet.

It's clear that the error is being assigned to a variable instead of being raised. Furthermore, the behavior is only present when evaluating a variable; this is unexpected, as it has nothing to do with the session itself.

I came across this while trying to enable debugging via TensorBoard. After some digging, it appears that the error message originates from TensorFlow core code, so that's why this issue is here and not in the TensorBoard project. Is there any status on implementation of this feature? If not, could there be a log warning or some sort of feedback to reflect this? The TensorBoard modal at http://localhost:6006/#debugger gives no indication that it is platform-dependent.

[naefl]

Had same issue. Tried running it in WSL without problems, only difference was running file from WSL instead of windows 10

[dgiunchi]

Same issue for me, windows 10

[ClauCec]

try to replicate the tensorboard debugger tutorial with same issue on windows 10

[rvfischione]

@ClaCec The tfdbg tutorial (here for posterity) doesn't use gRPC, but there's a flag in debug_mnist.py for using TensorBoard, so I'll assume that's what you're referring to. I ran these:

tensorboard.exe --logdir . --debugger_port 1234

python .\debug_mnist.py --tensorboard_debug_address 'localhost:1234'

Here's the output:

33 ops no flops stats due to incomplete shapes.
Accuracy at step 0: grpc:// debug URL scheme is not implemented on Windows yet.
Accuracy at step 1: grpc:// debug URL scheme is not implemented on Windows yet.
Accuracy at step 2: grpc:// debug URL scheme is not implemented on Windows yet.
Accuracy at step 3: grpc:// debug URL scheme is not implemented on Windows yet.
Accuracy at step 4: grpc:// debug URL scheme is not implemented on Windows yet.
Accuracy at step 5: grpc:// debug URL scheme is not implemented on Windows yet.
Accuracy at step 6: grpc:// debug URL scheme is not implemented on Windows yet.
Accuracy at step 7: grpc:// debug URL scheme is not implemented on Windows yet.
Accuracy at step 8: grpc:// debug URL scheme is not implemented on Windows yet.
Accuracy at step 9: grpc:// debug URL scheme is not implemented on Windows yet.

[wjwka]

got the same issue

[ideenfix]

Briefly said: Currently Windows is not supported by Tensorflow 1.8 / Tensorboard 1.8 (see https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/debug/debug_io_utils.cc). The CC++ code block is labeled as #ifndef ... #else directive block for URLs starting with grpc:://

[rvfischione]

@ideenfix That code block was linked in the issue description. :)

I'd still like some consistency between the documentation, the Tensorboard UI and the C++ error.

[ghost]

@naefl

To bypass the Windows incompatibility and run the debugger in WSL, are you just running the command line in it (i.e. tensorboard --logdir= --port --debugger port ) or do you have everything reinstalled for linux (i.e. python, tensorflow, CUDA toolkit, CuDNN, etc.)?

I haven't seen anything that shows a fix for this would be (or would not be) coming to Windows , but the TensorBoard Debugger site doesn't have a mention of Windows 10 incompatibility.

[naefl]

@faust22 Just running Tensorboard from CL, WSL has no support for GPU, so no cuDNN/CUDA. Tf, Tb and all other dependencies are installed in WSL. From: Faust22 Sent: Thursday, May 31, 8:35 AM Subject: Re: [tensorflow/tensorflow] gRPC debug URL scheme support for Windows (#runningL To: tensorflow/tensorflow Cc: LucaNaef, Mention @naefl To bypass the Windows incompatibility and run the debugger in WSL, are you just running the command line in it (i.e. tensorboard --logdir= --port --debugger port ) or do you have everything reinstalled for linux (i.e. python, tensorflow, CUDA toolkit, CuDNN, etc.)? I haven't seen anything that shows a fix for this would be (or would not be) coming to Windows , but the TensorBoard Debugger site doesn't have a mention of Windows 10 incompatibility. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.

[ghost]

@naefl

I see, thanks. It might be a bit rough without GPU support. I mainly want the 'health' section for NAN and INF value analysis, but there appears to be other ways to do this, although TensorBoard seemed good for this.

@caisq

Is there potential to see Windows support for TensorBoard debugger in the near future? (i.e. 1.9 or further off?)

[LevanSanadiradze]

Same issue, on windows 10 with Tensorflow 1.9.
Has anyone found a way to fix this?

[khashei]

I have the same issue, on windows 10 with Tensorflow 1.9.

[rtygbwwwerr]

same issue on windows 7 tensorflow1.10

[LEChaney]

Same issue on Windows 10. Guess its not supported on windows....

[rvfischione]

@FurryFur As of this comment, it is explicitly not supported on Windows - that's why it's marked as a feature request.

[baogorek]

For what it's worth, the tensorboard debugger plugin appears to work pretty well with the Ubuntu 18.04 app on the Windows app store. I gave up on the original Ubuntu app over a year ago due to incredibly slow install times, but things have changed and it's much faster. I dropped miniconda on it, installed tensorflow, brought up tensorboard and everything is seeming to work in the browser. Figures crossed it keeps working without any surprises. Props to Microsoft for this sweet little app.

[pournami123]

I have tried tensorflow for deep learning in windows 10. However, facing the same issue "1 ops no flops stats due to incomplete shapes" with TypeError: 'UnimplementedError' object is not iterable, while trying to use Tensorflow debugger. Is there any solution?

[Efaq]

Is there any status on this one? I'm also receiving
tensorflow.python.framework.errors_impl.UnimplementedError: grpc:// debug URL scheme is not implemented on Windows yet. when trying to run an Estimator with debugging enabled (hooks = [tf_debug.TensorBoardDebugHook(_DEBUGGER_LOCATION)]) .
Tensorboard 1.14.0

[anandh51295]

Still this Problem is not solved?

[tilakrayal]

@rvfischione,
Is this still an issue?
Could you update TensorFlow to the latest stable version v.2.9 and let us know if you are facing the same error. Thank you!

[google-ml-butler]

This issue has been automatically marked as stale because it has no recent activity. It will be closed if no further activity occurs. Thank you.

[rvfischione]

@tilakrayal Substituting some of the calls in the above reproduction with their tf.compat.v1 counterparts (and disabling eager execution), I get

grpc._channel._InactiveRpcError: <_InactiveRpcError of RPC that terminated with:
        status = StatusCode.UNAVAILABLE
        details = "failed to connect to all addresses"
        debug_error_string = "{"created":"@1654740831.153000000","description":"Failed to pick subchannel","file":"src/core/ext/filters/client_channel/client_channel.cc","file_line":3218,"referenced_errors":[{"created":"@1654740831.153000000","description":"failed to connect to all addresses","file":"src/core/lib/transport/error_utils.cc","file_line":165,"grpc_status":14}]}"

I have no idea how much of the user base still uses V1 API, but I doubt this issue is still relevant.