Python API for converting tensorflow models to DLA support

Hi
I want to run tensorflow models on jetson xavier nx on DLA accelerator.
1- How I can convert the frozen model to DLA support? Is there sample codes for python api?
2- DLA accelerator is hardware like decoder/encoder HW? Isn’t like GPU?
3- For running and converting the models to TensorRT for its GPU, Does work the same codes of jetson nano for xavier nx GPU?

Hi,

DLA is an extra hardware on Xavier which is independent to the GPU.
Currently, the only API to control DLA is TensorRT so you will need to deploy your model with TensorRT first.

TensorRT do have python interface and you can choose which hardware to deploy via setting the DeviceType:
https://docs.nvidia.com/deeplearning/tensorrt/api/python_api/infer/Core/NetworkConfig.html?highlight=devicetype#tensorrt.IBuilderConfig.set_device_type

Some python example can be found in this folder:

$ /usr/src/tensorrt/samples/python/

Thanks.

Thanks so much, @AastaLLL
1- Only version TensorRT 7 support DLA? Does support Trt 5,6 DLA also?
2- DLAs only support INT8? or FP16?
3- What about GPU? Can be support INT8?
4- DLAs indeed are Tensor Cores? I see the xavier nx has Tensor Core? Tensor Core related to GPU or DLA?

Hi,

1. YES, TenosrRT v5 and v6 also support DLA.

2 DLA only support INT8 and FP16 data format.

3 Xavier GPU can support FP32, FP16 and INT8.
Please see our support matrix here for the details:
https://docs.nvidia.com/deeplearning/tensorrt/support-matrix/index.html

4 No. Tensor Core is part of GPU.
DLA are the extra hardware on the Xavier.
You can find the spec of XavierNX here:

Thanks.