AM5708: TIDL version

Hi,

Can you please elaborate your requirements better?  We only support running the inference on Sitara processor. CUDA is not applicable to Sitara processor side. 

Are you asking which framework and corresponding version to pick to train the network model so that it can run the inference on Sitara devices successfully? Do you want to run the inference of your network model on DSP or ARM core? Do you already have network model available to train on?  Can you share the layers/operators types used in the model? What's your performance need for the inference? I will be better able to assist you if I know these information. 

We support following versions of Caffe and TF frameworks trained model to run on DSP and EVE accelerators.

• Caffe-Jacinto, is fork of 0.17 caffe (nvidia), pretty much latest Caffe (which is more or less in dormant state)
• TF support, is, I think, aligned with 1.11 (last TF1 is 1.14)

Network models can as well be run on Arm core of Sitara processor using ARMNN. Currently we are supporting ARMNN/ACL version 19.05 in our Processor SDK Linux. Please check the documentation of 19.05 ARMNN version to understand the TF and Caffe models supported by it. 

There are certain restrictions on the layers/operators that we support to run the inference on DSP/EVE. Please check the documentation here on supported layers/operators type.

http://software-dl.ti.com/processor-sdk-linux/esd/docs/latest/linux/Foundational_Components_TIDL.html#neural-network-layers-supported-by-tidl

Regards,

Manisha