• State Resolved
  • Locked Locked
  • Replies 1 reply
  • Subscribers 98 subscribers
  • Views 304 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

AM3358: An error has been reported running tflite-runtime

Cherry Zhou
Mastermind 22235 points
Part Number: AM3358


Hi team,

Here's an issue from the customer may need your help:

Use BeagleBone Black to attempt to run the tensorflow lite2.5.0, but an illegal instruction is reported:

The customer found that this was due to an issue with the invoke() implementation because the tflite installation package uses neno-vfpv4, but AM3358 only supports neno-vfpv3.

However, according to docs, AM3x can run the tensorflow lite.

Running tflite-support on the BeagleBone Black was successful, but the python file in the library was not found. Could you help share this code?

In order to resolve vfpv3 issues, the customer tries to compile itself from the tensorflow source code, but the clog compilation fails, as shown below:

ADD_SUBDIRECTORY not given a binary directory but the given source
directory "/home/hxy/Desktop/clog-source" is not a subdirectory of
"/home/hxy/Desktop/cpuinfo". When specifying an out-of-tree source a
binary directory must be explicitly specified.

Could you help share the compiled code? Thanks.

Best Regards,

Cherry

  • +1
    TI__Guru 69957 points

    Not sure what your overall setup looks like but our official Linux SDK offering is the Processor SDK Linux for AM335x, see here: https://www.ti.com/tool/PROCESSOR-SDK-AM335X  This SDK contains some version of tensorflow lite (v1.15).

    If you want a newer/different version and perhaps different add-on packages (such as python3 support) you can customize your Yocto build accordingly by updating your local.conf file as below (example). Building everything through Yocto will ensure proper cross compilation of all components including making sure appropriate compile flags are set, typically avoiding any and all binary compatibility issues such as you reported.

    # Install some extra packages
    CORE_IMAGE_EXTRA_INSTALL:append = " tensorflow-lite python3-tensorflow-lite"

    For this to work properly you'd want to update your meta-processor-sdk layer to the latest 'dunfell' commit to get access to some of the new/updated deep learning related Yocto recipes that have gotten added since AM335x Processor SDK v8.02 was released (those packages area already available and integrated into later SDK releases for other platforms). For example, tensorflow lite was updated to v2.8.

    Regards, Andreas