search

Xilinx / Vitis-AI

Vitis AI is Xilinx’s development stack for AI inference on Xilinx hardware platforms, including both edge devices and Alveo cards.
https://www.xilinx.com/ai
Apache License 2.0
1.44k stars 624 forks source link

No models available in Model zoo for Alveo U200 under tensorflow 2 #334

Open daivikbhatia opened 3 years ago

daivikbhatia commented 3 years ago

Hi, I want to test tensorflow 2 for Alveo U200 and I want to start with a pre built model. Unfortunately the .xmodel file is missing for alveo U200 under tensorflow 2 in the .ymal file. Can someone please tell when the examples will be ready for alveo U200?

sumitn-xilinx commented 3 years ago

Check this area: https://github.com/Xilinx/Vitis-AI/tree/master/examples/DPUCADX8G/tensorflow

daivikbhatia commented 3 years ago

this is for tensorflow 1. I need examples for tensorflow 2

bryanloz-xilinx commented 2 years ago

@paolodalberto, @woinck ... is it true we don't have tf2 support for u200 / u250 in the model zoo?

For instance: https://github.com/Xilinx/Vitis-AI/blob/master/models/AI-Model-Zoo/model-list/tf2_resnet50_imagenet_224_224_7.76G_2.0/model.yaml

I notice this doesn't have an entry for u200 / u250.

paolodalberto commented 2 years ago

I can answer the question only partially. The system support tensorflow2 (TF2)-> xmodel. Thus we should be able to compile and execute TF2 models for DPUV3INT8 and U2**.

So although there is no tutorial entry or example entry, any one could try to generate such an entry and the system should work: especially for the same networks that are in the model zoo TF1 (i.e., resnet50)

A more specific question is whether the tutorial examples uses TF2 native or already quantized xmodel from TF2.

Looking at the contents of one file:

(dv3int8xrm) paolod@xsjfislx22:/wrk/hdstaff/paolod/perforce/RDI_paolod_Dev_work/src/DeepLearning/xilinx/gitlab/vai-toolchain/TEMP$ tar zxvf resnet50_tf2-zcu102_zcu104_kv260-r2.0.0.tar.gz 
resnet50_tf2/
resnet50_tf2/resnet50_tf2.prototxt
resnet50_tf2/resnet50_tf2.xmodel
resnet50_tf2/md5sum.txt
resnet50_tf2_acc/
resnet50_tf2_acc/resnet50_tf2_acc.prototxt
resnet50_tf2_acc/resnet50_tf2_acc.xmodel

it seems the latter, It seems that you are going to work with the xmodel directly.

Does this make sense to you @daivikbhatia @bryanloz-xilinx ?