ARM6/RPI: Executor failed to create kernel _FusedConv2D

[sdeoras]

Please make sure that this is a bug. As per our GitHub Policy, we only address code/doc bugs, performance issues, feature requests and build/installation issues on GitHub. tag:bug_template

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow):
  yes

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04):
  ARM 6 (Raspberry PI Zero W)

• Mobile device (e.g. iPhone 8, Pixel 2, Samsung Galaxy) if the issue happens on mobile device:
  No

• TensorFlow installed from (source or binary):
  Source at 92b598a

• TensorFlow version (use command below):
  n/a

• Python version:
  n/a

• 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

You can collect some of this information using our environment capture script
You can also obtain the TensorFlow version with
python -c "import tensorflow as tf; print(tf.GIT_VERSION, tf.VERSION)"
n/a

Describe the current behavior
My use case is running TensorFlow as a C-library using Go interface. I am performing an image inferencing task using a model (.pb file). I see that my libtensorflow.so build works fine on ARM7 but fails on ARM6 with the following error:

2019-01-07 04:07:27.533445: E tensorflow/core/common_runtime/executor.cc:634] Executor failed to create kernel. Not found: No registered '_FusedConv2D' OpKernel for CPU devices compatible with node {{node module_apply_default/InceptionV3/InceptionV3/Conv2d_1a_3x3/Relu}}
	.  Registered:  <no registered kernels>

	 [[module_apply_default/InceptionV3/InceptionV3/Conv2d_1a_3x3/Relu]]
2019/01/07 04:07:27 session run:No registered '_FusedConv2D' OpKernel for CPU devices compatible with node {{node module_apply_default/InceptionV3/InceptionV3/Conv2d_1a_3x3/Relu}}
	.  Registered:  <no registered kernels>

	 [[module_apply_default/InceptionV3/InceptionV3/Conv2d_1a_3x3/Relu]]

Describe the expected behavior
expecting ARM6 build of libtensorflow.so and libtensorflow_framework.so to work the same as ARM7 build.

Code to reproduce the issue
Provide a reproducible test case that is the bare minimum necessary to generate the problem.

Other info / logs
Include any logs or source code that would be helpful to diagnose the problem. If including tracebacks, please include the full traceback. Large logs and files should be attached.

[msymp]

Hello @sdeoras , Although your question is interesting, it is more of a systems query regarding ARM6 vs ARM7 builds, hence is best addressed on Stack Overflow at:
Stack Overflow
We shall close this issue now. Thanks.

[sdeoras]

@msymp, np. i'll try to explore this issue on stack overflow. I could not tell from the error mesg if was code issue or ARM6 issue :)

[gunan]

@petewarden may know

[ghost]

Hello @sdeoras I have exactly the same problem on a armv7l board (cortex A9) with tensorflow 1.12.0. Do you solve this problem on StackOverflow (i didn't find any relevant questions on SO) ?

Regards.

[sdeoras]

@Nemo-Point: i saw the issue on arm 6 but things worked fine on arm 7. I tried master though. no, i have not found a solution so far.

[petewarden]

I'm not sure what's happening here unfortunately! I don't think I'll have a chance soon to debug, so adding contributions welcome.

[maliksyria]

hello
same error here
have you solve it ?

[njbuch]

Yep, same problem here on Raspberry 3+

[ghost]

I solved my problem (Nitrogen Quad board - iMX6 - cortex A9, armv7l) by changing some compilation features following the notifications from ARM:

Using the build script provided by tensorflow for RASPBERRY boards, I used the following compilation line in tensorflow/tools/ci_build/pi/build_raspberry_pi.sh :

PI_COPTS="--copt=-march=armv7-a --copt=-mfpu=vfpv3
  --copt=-std=gnu11 --copt=-DS_IREAD=S_IRUSR --copt=-DS_IWRITE=S_IWUSR
  --copt=-O3
  --copt=-U__GCC_HAVE_SYNC_COMPARE_AND_SWAP_1
  --copt=-U__GCC_HAVE_SYNC_COMPARE_AND_SWAP_2
  --copt=-U__GCC_HAVE_SYNC_COMPARE_AND_SWAP_8"
  WHEEL_ARCH=linux_armv7l

After cross-compiling and installing the wheel on the target board, everything was OK : training, testing, load & save models with a simple CNN on MNIST (classical example of Keras).

Regards

[mhurliman]

Any TF staff guidance on an ARM6l fix/workaround? It seems Tensorflow is keen to fuse the Conv2D & ReLU operations even if I make them separate ops - in contrast to using the activation parameter in tf.keras.layers.Conv2D.

[gunan]

@petewarden may have a comment on this.

[dudasdavid]

It seems fused Conv2D and ReLU operations can work on ARMv6 if I disable biasing in Conv2D operations (use_bias=False).

[github-actions]

This issue is stale because it has been open for 180 days with no activity. It will be closed if no further activity occurs. Thank you.

[pkgoogle]

Hi,

Thank you for opening this issue. Since this issue has been open for a long time, the code/debug information for this issue may not be relevant with the current state of the code base.

The TFLite team is constantly improving the framework by fixing bugs and adding new features. We suggest you try the latest TensorFlow version with the latest compatible hardware configuration which could potentially resolve the issue. If you are still facing the issue, please create a new GitHub issue with your latest findings, with all the debugging information which could help us investigate.

Please follow the release notes to stay up to date with the latest developments which are happening in the TFLite space.

[github-actions]

This issue is stale because it has been open for 7 days with no activity. It will be closed if no further activity occurs. Thank you.

[github-actions]

This issue was closed because it has been inactive for 7 days since being marked as stale. Please reopen if you'd like to work on this further.

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No