Compiler/PROCESSOR-SDK-DRA8X-TDA4X: TIDL to import YOLOV3

Lola Gorochana Addepalli

Part Number: PROCESSOR-SDK-DRA8X-TDA4X

Tool/software: TI C/C++ Compiler

Hi all,

I want to import my YOLOV3 model using TIDL for PC simulation.

1)I need assistance in defining pipeline configuration file for YoloV3 model----> I can successfully import the Darknet backbone architecture but need help for the detection part of "prediction from scales".

2)I have came across few e2e threads where it is mentioned " we have not validated YOLO models but will do for it soon"---->So when can we expect for a YOLO validated example.

3)I need more information on the approach below on how to implement or the way to use the feature of DLR-TVM-->

"Enabling heterogeneous xNN compute. where, the nodes that cannot run on TIDL would run on ARM. In other words, it will be possible to create subgraphs that can run on TIDL accelerators and the rest on ARM.

Current PSDK6.2 has already the option to do full offload to ARM or TIDL by using DLR-TVM.. So, if a model is completely supported by TIDL it would run in TIDL. But, If there is any unsupported layer(s), then the model would be compiled for ARM only."

Any help would be appreciated

Thanks in advance!

A. Lola Gorochana

over 4 years ago

0 Manu Mathew over 4 years ago

TI__Genius 11376 points

Hi,

We have added object detection scripts/examples with support for model export. The exported model can be used in TIDL. Currently SSD and RetinaNet models are supported.

YOLOv3 support is also experimentally added with the following limitations:

(1) Training of YOLOv3 is not working.

(2) Inference accuracy may not be optimal in TIDL and we are trying to analyze it.

Please see more details here:

https://github.com/TexasInstruments/jacinto-ai-devkit

https://git.ti.com/cgit/jacinto-ai/pytorch-mmdetection/about/

https://git.ti.com/cgit/jacinto-ai/pytorch-mmdetection/about/docs/det_modelzoo.md

We have listed the YOLOv3 complexity (GigaMACs) in the above link. As you can see YOOv3 is too complex - not at all recommended. There are other configuration that give similar accuracy at lower complexity. YOLOv3 is also not fully working in our training code and in TIDL.

We recommend you start with one of our RegNetX based SSD or RetinaNet configurations.

0 Lola Gorochana Addepalli over 4 years ago in reply to Manu Mathew

Intellectual 310 points

Thanks Manu for the reply!

1) I am training YOLOV3 independently and then trying to infer so I dont think the training part will be a problem.

2) When can we expect a fully validated(working) example for YOLOV3 from TIDL.?

3) As of now I imported onnx format of model does .pb format input will result in output?

4) Regarding the approach to "run all the layers on TIDL except the last layer which can be executed later on DSP/EVE as a C function (Detections to Bounding boxes)."---> can you provide with any links which can be helpful.

I have already tried for SSD based networks and able to infer but now I want to try for YOLO, thats the very reason i want information on it.

0 Manu Mathew over 4 years ago in reply to Lola Gorochana Addepalli

TI__Genius 11376 points

Hi,

YOLOv3 support is planned for the release being planned for November 2020. With that release, you will be able to infer the model given here:

https://github.com/onnx/models/tree/master/vision/object_detection_segmentation/yolov3

(yolov3-10.onnx)

You will need to provide the onnx file and a .prototxt file similar to the way SSD inference is done using onnx file and .prototxt file. In this method, all the layers will run on TIDL except the last layer which will be executed later on DSP/EVE as a C function - you don't need to do anything special for it. Did this answer your question?

If it did not answer your question, please explain how did you do inference of SSD - what were the files that you gave as input - was it onnx+prototxt or something else. Also, why did you mention .pb file? Are you doing training in tensorflow?

0 Lola Gorochana Addepalli over 4 years ago in reply to Manu Mathew

Intellectual 310 points

HI, Thanks for the update. So i will try inferencing my model once he release is done on November.

Yes my .onnx format is same as the one mentioned in github link.

Yes, the answer u provides justifies my query.

No, I do training with PyTorch then i convert .pth to different model formats through .onnx.(So .pth-->.onnx-->.pb)

0 kumar.desappan over 4 years ago in reply to Lola Gorochana Addepalli

TI__Mastermind 22145 points

Closing the thread for now.

Processors

Processors forum

Compiler/PROCESSOR-SDK-DRA8X-TDA4X: TIDL to import YOLOV3