SK-TDA4VM: SK-TDA4VM

Cherry Zhou

Part Number: SK-TDA4VM

Hi team,

Here's 2 questions may need your help:

1. the customer found that the struct of sTIDL_LayerParams_t has no tanh function, is this supported in tidl?

2. the csutomer would like to use yolov5 model with tidl, and the output of the model is not only cls_conf, cls_id,box, but also with other information following box(e.g cls_conf, cls_id, xywh, theta1, theta2). Then if they want to infer the model with tidl-type, what should they do?

Could you help resolve this case? Thanks.

Best Regards,

Cherry

over 4 years ago

0 Anshu Jain over 4 years ago

TI__Guru 56820 points

Hi Cherry,

Cherry Zhou said:
1. the customer found that the struct of sTIDL_LayerParams_t has no tanh function, is this supported in tidl?

We do support operators beyond the supported on C7x (list) on ARM with our open source run time offering (TFLite, ONNX RT and TVM DLR)
For most of the open source exchange formats such as Tensorflow, ONNX, tflite, etc., you can use our OSRT offering and get such operators executed on ARM

We do have some plan to support tanh operator on c7x but as per current road map its planned for SDK 9.1 release ( expected to be by end of next year).

Cherry Zhou said:
2. the csutomer would like to use yolov5 model with tidl, and the output of the model is not only cls_conf, cls_id,box, but also with other information following box(e.g cls_conf, cls_id, xywh, theta1, theta2). Then if they want to infer the model with tidl-type, what should they do

Sorry I cound't understand the part about inferring the model with tidl-type? What do you mean by tidl-type here?

Regards,

Anshu

0 yuanfu duan over 4 years ago in reply to Anshu Jain

Prodigy 186 points

the meaning of tidl-type is a model we get by the tools of tidlModelImport.out. we use a model following the output with[x,y,w,h,conf, cls_id, cos_value](cos_value is a value of cos ).in the Tl Deep Learning product User Guide, wo only find the use of keypoints with postprocess, but we only want to get a cos-value without postprocess, what steps should I take?

0 Anshu Jain over 4 years ago in reply to yuanfu duan

TI__Guru 56820 points

Hi,

I am still not completely clear here, if you are looking for TIDL output without post processing then it can be done by using outDataNamesList variable in import config file. Here you can specify all the outputs before post processing so TIDL will not parse the post processing part. You can find the documentation for the same here:

https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/08_00_00_12/exports/docs/tidl_j7_08_00_00_10/ti_dl/docs/user_guide_html/md_tidl_model_import.html#tidl_model_import_basic

Let me know if this is what you are looking for?

Regards,

Anshu

0 Debapriya Maji over 4 years ago in reply to Anshu Jain

TI__Intellectual 2406 points

Hi Yuanfu,

I think you meant for each box, there are some extra parameters: {theta1, theta2}. You want these extra parameters as it is without any post-processing that is they are output of the final convolution without any changes. Let me know if my understanding is correct. We can help you based on that.

Regards, Debapriya

0 yuanfu duan over 4 years ago in reply to Debapriya Maji

Prodigy 186 points

YES! It is exactly what I want to say!

0 Debapriya Maji over 4 years ago in reply to yuanfu duan

TI__Intellectual 2406 points

Hi Duan,

As you mentioned, we support key-points detections only right now and these keypoints are offsetted w.r.t the center of the corresponding anchor during post-processing.

We don't have support for passing a parameter without any post-processing. If you can share little more details about the parameters you are trying to predict that will help us. If you are trying to do human pose estimation, we are planning to add support for that in 8.2 release.

Regards, Debapriya