coral_usb_ros icon indicating copy to clipboard operation
coral_usb_ros copied to clipboard

Published 20 hours ago • verified publisher icon jsk-ros-pkg

tensorflow v2 model support

Open Petros626 opened this issue 2 years ago • 6 comments

Hey,

I've read that only ssd models are supported for TFLite1 and TFLite2, so the hardware accelerator (Edge TPU) supports only ssd models until now. I saw that in the example the ssd_mobilenet_v2 i used for inference. But would the newer one (SSD MobileNet V2 FPNLite 640x640) of TensorFlow 2 also with this node?

Petros626 avatar May 25 '22 12:05 Petros626

Only tflite1 is supported, I think. Ive never tried tensorflow v2 before.

knorth55 avatar May 25 '22 13:05 knorth55

Okay thank you. So you dont know it 100%. Is there the option to ask the guy, who added the support for ssd_mobilenet_v2?

Petros626 avatar May 25 '22 13:05 Petros626

Sorry, but I dont understand your question. Mobilenet V2 is the model name, and it doesnt mean the model is for tensorflow 2

knorth55 avatar May 25 '22 13:05 knorth55

Ok maybe I explain it wrong. The user: Kanazawanaoaki added: "Add ssd to bounding box sample" to the jsk_recognition package right? So this is implemented in this package, because you said it's connected?

The other point was, that the model: ssd_mobilenet_v2 is only supported by TensorFlow 1 and the newer one (I asked for) SSD MobileNet V2 FPNLite only by TensorFlow 2. The question is, if I try to run your node with a tflite model, which bases on the newer one would it be compatible or I must stick to the ssd model from TensorFlow 1 :)

Petros626 avatar May 25 '22 14:05 Petros626

I understood.

I wrote almost whole part of this repo (you can check it in commit log), and, at that time, tensorflow2 is not common. So I wrote in tensorflow1, and all models are in tensorflow1. If you want to use tensorflow2 model, I have no idea. Please try by yourself first. If there is no backward compatibility, you need to change this repo's code or hate tensorflow because of the lack of backward compatibility.

This is open source repo, and not a product. I will help you some but I want you to help us, too. If I need to use tensorflow2 model, I will implement, but there is no need in my lab now, so I won't do it soon.

Do not hesitate to change the code. Please open a pull request. Do not completely expect to or rely on others' work. Now you get a chance to learn tensorflow2.

knorth55 avatar May 25 '22 15:05 knorth55

Yeah thank you for the detailed explanation, if I have the time I would do so, but I'm working on my bachelor degree, so for first use it should only run on TensorFlow 1 model.

Hahah I think the guys from TensorFlow do their best :D

Petros626 avatar May 28 '22 12:05 Petros626