Code for Stress-Testing USB Accelerators for Efficient Edge Inference

Code and results for our corresponding research paper, which was published and presented at the 2024 IEEE/ACM Symposium on Edge Computing (SEC).

Explore our experimental results interactively

To view our experimental results, check out the publicly available webpage of our framework for Sustainable and Trustworthy Reporting, which offers the EdgeAcc results alongside several other evaluation databases. Note that we continue to advance our software - it is work in progress and subject to change, so you might encounter delays, off-times, and slight differences to our paper.

If you want to run the exploration tool locally, make sure you have the the following python packages installed: numpy, pandas, pint, scipy, dash, dash_bootstrap_components, Pillow, reportlab, fitz, frontend, plotly Then start our app via main.py and open the webpage. Besides the interactive plots, you can also inspect the PDF files in our paper_results directory.

Citing

If you appreciate our work and code, please cite our paper as given by Springer:

A. Van Der Staay, R. Fischer and S. Buschjäger, "Stress-Testing USB Accelerators for Efficient Edge Inference", 2024 IEEE/ACM Symposium on Edge Computing (SEC), Rome, Italy, 2024, pp. 1-14, doi: 10.1109/SEC62691.2024.00015.

or using the bibkey below:

@INPROCEEDINGS{10818191,
  author={Van Der Staay, Alexander and Fischer, Raphael and Buschjäger, Sebastian},
  booktitle={2024 IEEE/ACM Symposium on Edge Computing (SEC)}, 
  title={Stress-Testing USB Accelerators for Efficient Edge Inference}, 
  year={2024},
  volume={},
  number={},
  pages={1-14},
  doi={10.1109/SEC62691.2024.00015}
}

Models that we considered:

Models compared for Imagenet Classification:

'DenseNet121' 'DenseNet169' 'DenseNet201' 'EfficientNetB0' 'EfficientNetB1' 'EfficientNetB2' 'EfficientNetB3' 'EfficientNetB4' 'EfficientNetB5' 'EfficientNetB6' 'EfficientNetB7' 'EfficientNetV2B0' 'EfficientNetV2B1' 'EfficientNetV2B2' 'EfficientNetV2B3' 'EfficientNetV2L' 'EfficientNetV2M' 'EfficientNetV2S' 'InceptionResNetV2' 'InceptionV3' 'MobileNet' 'MobileNetV2' 'NASNetLarge' 'NASNetMobile' 'ResNet101' 'ResNet152' 'ResNet50' 'ResNet101V2' 'ResNet152V2' 'ResNet50V2' 'VGG16' 'VGG19' 'Xception' 'MobileNetV3Large' 'MobileNetV3Small'

Models compared for Imagenet Segmentation:

'yolov8s', 'yolov8n', 'yolov8m', 'yolov8x', 'yolov8l'

Repository structure:

Directories

• batchsize_comparison_results contains the tables that we used to compare the performance of different batch sizes on our three host systems.
• classification_database and segmentation_database contain the results and configurations of our experimental results. These two directories are accessed by main.py to create the paper results.
• creator_scripts contain the python scripts used to create the datasets used for the experiment as well as the model files for the different model formats required by the accelerators TPU and NCS.
• helper_scripts contain miscelaneous scripts that are used by other scripts or that can be used to clean up experimental result files. Model metadata that we use for analysis is collected with the corresponding scripts. Our batch size comparison that determines the batch size of used in our CPU experiments is included here.
• paper_results contain our result graphs as PDF files.
• result_databases contain the pickled pandas dataframes created by load_experiment_logs.py that are further merged with the merge_all_databases.py script.
• strep contains scripts used to create our interactive model results.
• Make sure to include a directory that holds the model input data as well as all model files using the scripts from creator_scripts.

Scripts

• main.pycan be executed to create the interactive model results based on the content of classification_database and segmentation_database. It uses the paper_results.py script where the graph specifications are coded.
• load_experiment_logs.py merges an experiments monitoring directory into a pandas dataframe and then pickles it for further use.
• merge_all_databases.py merges all of the chosen dataframes form result_databases into the classification_database and segmentation_database directories for the final results.
• pycoral_classification.py, pycoral_segmentation.py and pycoral_classification_raspi.py are the scripts that run the actual experiments. They are run by the run_all scripts.
• The run_all scripts define one run of an experiment depending on the used environments. We ran these expermients multimple times for our paper results.

How to setup your own experimental runs:

Google Coral EdgeTPU Setup

• Follow guide: https://coral.ai/docs/accelerator/get-started/
• Setup EdgeTPU Compiler for TPU Conversion: https://coral.ai/docs/edgetpu/compiler/
• Setup pycoral: https://coral.ai/docs/edgetpu/tflite-python/#update-existing-tf-lite-code-for-the-edge-tpu

Intel NCS Setup

• Follow guide: https://docs.openvino.ai/2022.3/openvino_docs_install_guides_installing_openvino_from_archive_linux.html
• Step 2 (environment config) needs to be performed in each session!
• udev Regeln angepasst werden und dein Nutzer muss der users Gruppe hinzugefigt werden. https://docs.openvino.ai/2022.3/openvino_docs_install_guides_configurations_for_ncs2.html#ncs-guide
• pip install openvino

1. Setup Directories and Download Raw Data

• Make sure you created your model and data directory in our scripts, we use the directory name mnt_data.
• In mnt_data, we have the subfolders unpacked and staay. In unpacked, we downloaded the 1% imagenet database. (https://www.image-net.org/)
• In staay, we have the imagenet_data directory where our preprocessed data is saved as well as the models directory where the models are saved in the corresponding subdirectories edgetpu_models,openVINO,saved_models,tflite_models. These directories get filled by using the scripts in the creator_scripts directory.
• Also in staay the files coco.yaml and coco128-seg.yaml should be included for segmentation with YOLO models. (Follow : https://docs.ultralytics.com/datasets/detect/coco/#applications)
• If you choose to change the naming of your files, make sure to adjust all occurences of mnt_data in this repositories scripts.

2. Fill Directories using Creator Scripts

• Using the create_imagenet_dataset.py script, create the preprocessed imagenet data of the models that you want to compare. You may choose from any of the Tensorflow2 models listed in this readme. It saves the model-individualized preprocessed datasets as numpy arrays in the imagenet_data file.
• Use the export_tflite_models.py scripts to export the models into Tensorflow Lite models. These can then converted into TPU-compatible models using the export_edgetpu_models.py script. The latter script uses the EdgeTPU Compiler.
• Export the saved_models into NCS compatible models using export_NCS_models.py. This uses the model optimizer mo provided by openVINO2022.3. Make sure to save the created models in the correct directory (models/openVINO).
• Execute the create_YOLO_models.py script to create YOLO models as saved_model, edgeTPU compatible model and openVINO model. Save accordingly.

3. Run Experiments

• Now you can execute singular runs with the pycoral_ scripts. Adjust all flags to your liking.
• You can even test all of your created models by using the run_all scripts adjusting the flags to your liking.
• This will create a directory with the experiments logging. (Monitoring directory)

4. Create database from monitoring directories

• Adjust the path at the end of load_experiment_logs.py to your monitoring directory and execute the script. This will merge the directory into one dataframe that is saved in result_databases
• In merge_all_databases.py, adjust all of the databases you want to include from result_databases. This will create the final databases in the classification_database and segmentation_database directories.
• Now main.py can be run with the newly included experiment results.

Troubleshooting:

TPU ERROR: ValueError: Failed to load delegate from libedgetpu.so.1

• de- and reconnect edgeTPU
• try different USB cable!

Copyright (c) 2025 Raphael Fischer, Alexander van der Staay, Sebastian Buschjäger