• State Answer Suggested
  • Locked Locked
  • Replies 3 replies
  • Answers 2 answers
  • Subscribers 97 subscribers
  • Views 484 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.

TDA2PXEVM: Problems with the TIDL library import tool

Ahmed Anwar
Intellectual 260 points

Part Number: TDA2PXEVM
Other Parts Discussed in Thread: TDA2

Hello,

I am trying to use TIDL library on TDA2PX EVM to run DNNs on EVE and DSP cores. I am using TensorFlow as a framework. I am following the user guide with the title " TI Deep learning Library on TDAx " - November 2019.

The user guide states that " TIDL supports slim based tensorflow models ". Does TIDL library supports only slim based TensorFlow models ? Does it mean that I can not import models trained in TensorFlow, not TensorFlow slim?

I have tried to import MobileNet V2 from TensorFlow Keras applications using the TIDL library, but I get error mesages saying that "Pad Layer is not supported by TIDL and cannot be merged into any TIDL layer".

Moreover, I need to know where I can find the file "optimize_for_inference.py".

Thanks,

Ahmed Anwar

  • 0
    TI__Genius 17580 points

    Hi Ahmed,

    >> Does it mean that I can not import models trained in TensorFlow, not TensorFlow slim?

    Yes, TIDL on TDA2 can import only TensorFlow slim models. On TDA4 we support TensorFlow models also.

    >>  but I get error mesages saying that "Pad Layer is not supported by TIDL and cannot be merged into any TIDL layer"

    This is because of Pad layer is not supported, refer to datasheet for all the supported layers

    >>  I need to know where I can find the file "optimize_for_inference.py"

    This is available in "tensorflow/python/tools" folder

    Thanks,

    Praveen

  • 0 in reply to Praveen Eppa1
    Intellectual 260 points

    Hello Praveen,

    Thank you for your reply.

    I have a question regarding the TensorFlow slim models. In the user guide, there is an example of a Keras/TensorFlow model which can be found here. This example uses Keras from TensorFlow, not TensorFlow slim, to build a simple CNN, which can be used as an input to the TIDL import tool. Does this mean that I can use Keras from TensorFlow, not TensorFlow slim, to build a CNN and use it as an input to the TIDL import tool ?

    I just need to understand this point better, as most of my development is already in Keras/TensorFlow, and the example that I mentioned also uses Keras/TensorFlow. However, based on your reply, "TIDL on TDA2 can import only TensorFlow slim models". Would you please illustrate how this example uses Keras/TensorFlow if only TensorFlow slim models are supported on TDA2 ?

    Thanks,

    Ahmed Anwar

  • 0 in reply to Ahmed Anwar
    TI__Genius 17580 points

    Yes, We have limited layers support for Tesnsoflow in TDA2, where as most of the layers are supported with Caffe/Tensorflow slim models. Please refer to below comment in the user guide section 3.6.5. 

    " We have developed/defined TIDL library layers based on the layer types Caffe framework. Most of our test cases (Layer level and network level) and demos are based caffe framework. With respect to Tensorflow, we have validated two pre-trained models from tensorflow github (Slim based Mobilenet V1 and Googlenet/inceptionetV1), this covers most of the CNN layers (Convolution, Max Pooling , Average pooling, Batch norm, Fully connected layer, softmax, Relu, Relu6, concate etc). "

    Thanks,

    Praveen