CoTeD

CoTeD targets UAV–edge cooperative video object detection scenarios where a DNN is split between a resource-constrained UAV and a edge server over a private 5G network. It proposes a goal-oriented framework (as the figure below shows) that, without model retraining, can dynamically prunes, compresses, and reconstructs DNN activation tensors explicitly trading off radio bandwidth consumption against inference accuracy and latency in real time.

This repository contains the implementation and descriptions of the CoTeD framework and testbed presented in our publication:

Yenchia Yu, Matteo Mendula, Marco Levorato, Marina Papatriantafilou, and Carla Fabiana Chiasserini. Efficient Tensor Compression and Reconstruction for Edge-Based Object Detection. {Publication info to be added}

How to use this Repository

Testbed setup

CoTeD is designed and validated on a UAV-5G-Edge testbed. For the testbed architecture and setup, please see the description in ./testbed/.

Running experiment

In this repository, we provide the example codes to run the experiment with YOLOv3-tiny model (under folder ./coted_framework/yolov3_tiny_experiment/) and YOLOv3 model (under folder ./coted_framework/yolov3_experiment/).

We provide the find-tuned YOLOv3-tiny model checkpoints in folder ./coted_framework/ckpt/. For testing with other datasets, please check the description in ./pytorchyolo/ to perform fine-tuning.

Running the validation experiment of CoTeD framework with the following steps (use YOLOv3-tiny model as example),

1 . Update the edge server IP address and port according to the testbed setup:

# In ./coted_framework/yolov3_tiny_experiment/experiment_edge.py
...
cherrypy.server.socket_host = "10.0.1.34"
cherrypy.config.update({'server.socket_port':8092 })
...

# In ./coted_framework/yolov3_tiny_experiment/experiment_client.py
...
service_uri = "http://10.0.1.34:8092/tensor"
reset_uri = "http://10.0.1.34:8092/reset"
...

2. Start the CoTeD edge and DNN tail at the edge server side with,

python3 experiment_edge.py <DNN_split_layer>
# For YOLOv3-tiny, we split at the 8th layer. For YOLOv3 model, we split at the 10th layer.

3. Run the CoTeD mobile and DNN head at the UAV computation unit (i.e., the Nvidia Jetson Orin Nano) with,

python3 experiment_client.py <DNN_split_layer>```
# For YOLOv3-tiny, we split at the 8th layer. For YOLOv3 model, we split at the 10th layer.

4. Find the experiment measurements under folder ./coted_framework/measurements/.

Test with different datasets

To test the CoTeD framework with different datasets, the following configurations should be updated before running the experiment (use YOLOv3-tiny model as example),

1. Fine-tuned YOLOv3-tiny model checkpoints:

# In ./coted_framework/yolov3_tiny_experiment/experiment_edge.py
cfg_path = "../configs/yolov3_tiny.cfg"
model_path = "../ckpt/stmarc_tiny.pth" # to the fine-tuned model checkpoint

# In ./coted_framework/yolov3_tiny_experiment/experiment_client.py
cfg_path = "../configs/yolov3_tiny.cfg"
model_path = "../ckpt/stmarc_tiny.pth" # to the fine-tuned model checkpoint

2. The SNR to mAP/Sensitivity drop curve parameters:

# In ./coted_framework/manager/manager_full.py

# STMARC [k, h, b]
self.map_curve = [0.062, 0.917, 0]
self.sen_curve = [0.067, 0.654, 2]

3. The test video frame sequences and label names:

# In ./coted_framework/yolov3_tiny_experiment/experiment_client.py
testdata_path = "../../datasets/St_Marc_dataset/test_30_fps_long_cleaned.txt"
class_name_path = "../../datasets/St_Marc_dataset/coco.names"