Code for [SenSys'24] PieBridge: Fast and Parameter-Efficient On-Device Training via Proxy Networks.

Project structure

PieBridge
|–– datasets/
|   |–– ... # datasets for on-device training
|–– doc/
|   |–– ... # documents
|–– proxynetworks/ # original NN, proxy networks, etc. 
|   |–– models/
|   |   |––jit/
|   |   |   |––...
|   |––pth/
|   |   |––...
|–– res/ # training configs and logs
|   |––train_log/ 
|–– scripts/
|   |––run_e2e.sh
|   |––...
|–– src/
|   |––training.py
|   |––...

Installation

Hardware configuration

To run the code, you must prepare an edge device with Pytorch support, e.g. Jetson TX2 or Jetson Orin.

By default, we use a Jetson TX2 with 8GB RAM and 16 GB swap size. We attach an extra 900 GB NVME SSD to TX2 since the born disk size is insufficient.

sudo systemctl disable nvzramconfig
sudo fallocate -l 16G /mnt/16GB.swap
sudo mkswap /mnt/16GB.swap
sudo swapon /mnt/16GB.swap

Maximize the power mode of your device to get the highest performance (e.g., GPU frequency).

# check the current power mode
$ sudo nvpmodel -q
NV Power Mode: MODE_30W
2

# set it to mode 0 (typically the highest)
$ sudo nvpmodel -m 0

# reboot if necessary, and confirm the changes
$ sudo nvpmodel -q
NV Power Mode: MAXN
0

Software environments

The basic environment of our device is

Jetpack 4.5.1 [L4T 32.5.2] with

Distribution: Ubuntu 18.04
Release: 4.9.201-tegra
Python: 3.6.9
CUDA: 10.2.89

For the direct dependencies of PieBridge, we recommend you directly using the docker image, or checking the software version details in the docker image and manually installing them.

docker pull yinwangsong2000/piebridge_ae

Downloading datasets and weights

The datasets and model weights are pre-uploaded in Google Drive. You can download them manually by

gdown <file-id>

and put them in the corresponding path of ./datasets/* and ./proxynetworks/*.

We also provide an automated script for downloading in ./scripts/downloading.sh.

Running

Run the end-to-end experiments:

cd PieBridge
bash ./scripts/run_e2e.sh 0

Run per-dataset experiments:

cd PieBridge
bash ./scripts/run_standalone_datasets/run_caltech101.sh 0

The results will be shown in ./res/*/log.txt.

Artifact evaluation

We provide a docker image of reproducing PieBridge and its baselines on Jetson TX2.

See doc/ae.pdf for details.