ssd_inception_v2 model and tensorflow on Ultra96v2 (petalinux 2020.2, Vitis AI-1.3)

Pradeep,

This is an excellent question, representing what I run into when I try to run one of the models in python.

Xilinx only provides a few examples for the Python API, so we are left to figure this out ourselves.

I have observed that I often find the answers in the python scripts included with the source archive for the model archive.

Based on your feedback, it sounds like this is not a possibility for the tensor flow models ...

Are you willing to share your python code, so that we can help find the solution ?

Regards,

Mario.

Hello Mario,

Thanks for the response.

1. I am trying to modify the python code what you have developed for facedetction unsing densobox model (face_py_vart example). I'll share the code once it is in a day or 2, once it is in some shape. I am trying to use the mode ssd_inceptionv2 (vitis-ai/models/AI-Model-Zoo/model-list/tf_ssdinceptionv2_coco_300_300_9.62G_1.3).

2. Any clue on why I am not getting any detection in the C++ application provided in sample/tfssd ? I tried with the compiled model provided in the off-the-shelf image given for V-1.3 for Ultra96V2, there are no detection boxes.

3. Is there any way by which we can get the tensorflow up and running on Ultra96V2 (or any other Xilinx board with Petalinux)?

Regards

Pradeep

Pradeep,

1. Understood, I will wait for your code

2. Yes, I "may" be able to explain that.  Another user identified an issue with my Vitis-AI 1.3 images, where I omitted some important {model}_officialcfg.prototxt files for the SSD models, which resulted in less detections.  Please refer back to the hackster project, where I describe this issue (in "Known Issues – Missing {model}_officialcfg.prototxt files for certain SSD models" section), and provide a workaround (archive in the "Code" section).

  http://avnet.me/vitis-ai-1.3-project

3. I have not tried to install tensorflow on Ultra96-V2.  Have you tried to install with "pip3 install tensorflow" ?

Regards,

Mario.

Hello Mario,

  1. In-Progress

  2. Yes. It is now showing the bounding boxes with lot of fliker . Any idea on how to show the class labels with this application?

  3. Yes I tried and It throws the following error:

root@u96v2-sbc-base-2020-2:~# pip3 install tensorflow

Collecting tensorflow

  ERROR: Could not find a version that satisfies the requirement tensorflow (from versions: none)

ERROR: No matching distribution found for tensorflow

root@u96v2-sbc-base-2020-2:~#

Also I did not see any aarch64 file @ https://pypi.org/project/tensorflow/#files

Regards

Pradeep

There is a Tensorflow install Guide on the 96Boards site.  This might help, or not.  However, having done something similar on an ODROID-C4, this could take quite a long time to compile.

https://www.96boards.org/documentation/consumer/guides/tensorflow-installation.md.html

Hello Mario,

   Sorry, it took long time for me to get back to you. Was involved in something else.

I have updated the face detection script of yours and I am stuck in-between.

The problem is:

Actually to get the detection boxes coordinate, I need the initial anchor boxes( priors), which I dont know how to get?

From DPU I am getting 2 tensors:

1) detection boxes with class (91 classes, as this is trained on COCO) confidences : shape: (1, 1917, 91,1)

2) detection boxes coordinates or something I dont understand (total 1917 boxes with each having 4 values) (1, 1917, 1, 4)

Let me know how to upload the scripts, I could not find any attachment link here.

Regards

Pradeep

Thanks. Will try it and update.

Regards

Pradeep

Hello Mario,

   Sorry, it took long time for me to get back to you. Was involved in something else.

I have updated the face detection script of yours and I am stuck in-between.

As I understood, I need to do the following to run SSD in Python (without tensorflow)

1) Do multiclass softmax on detection boxes score:: DONE

2) Convert the detection boxes to the actual size.:: STUCK

3) Do the multi-class Non-max suppression without tensorflow : NOT STARTED

The problem is:

Actually to get the detection boxes coordinate, I need the initial anchor boxes( priors), which I dont know how to get?

From DPU I am getting 2 tensors:

1) detection boxes with class (91 classes, as this is trained on COCO) confidences : shape: (1, 1917, 91,1)

2) detection boxes coordinates or something I dont understand (total 1917 boxes with each having 4 values) (1, 1917, 1, 4)

Let me know how to upload the scripts, I could not find any attachment link here.

Regards

Pradeep

Hi, did you make any progress on this? I'm implementing SSD MobileNet v2 (quite similar) & is now overwhelm with the amount of effort I may have to put in order build a python demo. The original source code depends entirely on TensorFlow 1.

Hello,

   Yes I was able to complete the conversion and run on Board. Due to Covid related curbs couldn't update the thread with the code. I'll be happy to share it