search

Microchip-Vectorblox / VectorBlox-SDK

18 stars 8 forks source link

[tutorials/tensorflow/mobilenet-v1-1.0-224-tf/] #6

Closed agnyde closed 2 years ago

agnyde commented 2 years ago

Hi, my set up is a VM Ubuntu 20.04x64 (kernel 5.13.0-28-generic). VirtualBox version is 6.1.32.
I try to run some tutorials in vbx environment, but I always have the same error. I followed the guide instructions, I run install_dependencies.sh and source setup_vars.sh and then in the tutorial folder I try to run mobilenet-v1-1.0-224-tf.sh and I have these two warnings and the core dumped error:
immagine So no results are generated, what could be the problem? Best regard

LiiNen commented 2 years ago

It seems like dependencies are not successfully installed

Is there a file named 'converter' in path below? {your working directory}/VectorBlox-SDK/vbx_env/bin/converter

if the converter file not exists, consider installing the dependencies again

best regard,

cirqueit commented 2 years ago

This should be simplified in release 1.3 (coming out this week). We will be using openvino-dev which provides mo (previously converter) and omz_downloader (previously downloader).

Please try 1.3 when released, and confirm it works in your VM.

agnyde commented 2 years ago

Hello, thanks for your answers.

llegal instruction (core dumped) python ../../saved_model_signature.py

Could the problem be related to some features of the host machine? Are there any other requirements? For example in the screenshot below there are proc/cpuinfo of my machine: immagine

Best regard!

agnyde commented 2 years ago

@LiiNen @cirqueit @vanjoe Hello, are there any updates regarding the issue above ? If it could be useful, I show you more precisely a case where python fails again with the message: llegal instruction (core dumped) this often happens when some python code is used. For example running the tutorial: tensorflow/yolo-v4-tiny-tf execution crashs at line 28 where there is this command: python $DOWNLOAD_DIR/keras-YOLOv3-model-set/tools/model_converter/convert.py $DOWNLOAD_DIR/keras-YOLOv3-model-set/cfg/yolov4-tiny.cfg $DOWNLOAD_DIR/yolov4-tiny.weights yolo-v4-tiny.h5 the error is always: Illegal instruction (core dumped) I am attaching a screenshot about some errors details too:

immagine

cirqueit commented 2 years ago

This is likely due to tensorflow not being able to run on your machine (often due to AVX instructions?). You may need to build tensorflow from source.

Run python -c "import tensorflow" to see if thats your issue

LiiNen commented 2 years ago

@agnyde Hi, sorry for late response of your mention.

I'm not sure about your problem, so VectorBlox developers might know. One thing I can advise you is using the previous version. Although the problems were not same, I also had trouble when I testing the board. When I use the previous version of source code, that problems were solved.

The commit hash which I used is below: 30b9c2815dae7c098e057a57c2656809d3ded03e

If you struggle yet, how about trying the previous version? This may not be the solution, but I think it worth it. Best regards,

agnyde commented 2 years ago

Hello, thanks for your answers. @LiiNen tensorflow core-dumped problem persists in all versions of source code. @cirqueit as you correctly pointed out the problem is the import of tensorflow, because AVX instructions are not supported on my machine. I tried to build tensorflow from source following this tutorial: https://gist.github.com/kmhofmann/e368a2ebba05f807fa1a90b3bf9a1e03 without success. So at the moment it seems that tensorflow cannot be used on vbx_enx if AVX istructions are not supported. Do you have any other suggestions? Thanks

cirqueit commented 2 years ago

Tensorflow is required when converting TF graphs into OpenVINO. If TF graphs are important, I would either stay trying to get a version of tensorflow built and installed in your env or enable AVX instructions in your virtual box VM as discussed here

The rest of the flow does not use TF.