Jongkuk Lim

KD currently does not support multi-gpu. We adopted KD method from "End-to-end semi-supervised object dection with soft teacher" and generating teacher's feature(I would say it's a guide feature for student)...

> is there a way to get the frozen_graph.pb? I haven't tried exporting a frozen TF model but will try on my free time 👍

I found related issue here [@56](https://github.com/Livox-SDK/livox_ros_driver2/issues/56)

Same situation here. I've tested on Jetson Xavier NX and RPi4 on ROS1 Noetic. NX was about 5\~10Hz and RPi was about 7\~10Hz. Also I have tried to increase `publish_freq`...

I believe I found a temporary solution. It's caused by checking publishing interval in millisecond. In the 161 ~ 163 lines of `pub_handler.cpp`, `recent_time_ms` is checked if it is modded...

Sorry for the late response. I don't think they have released any training code. I wish I had free time to write training code. I'm already tied up until December...

It's fully tested yet but if you want to try MobileViT, you can find out the sample yaml file at [tests/test_configs/yolo_mobilevit_sample.yaml](https://github.com/JeiKeiLim/kindle/blob/main/tests/test_configs/yolo_mobilevit_sample.yaml)

We use different class name for the head part which is `YOLOHead`. You can check this here https://limjk.ai/kindle/tutorial.html#5-make-object-detectiong-model-using-yolohead Also, we have separate MaxPool class too. Although it might work almost...

We may better switch MKApi to other auto API generator.