NASH

This repository is the code of Paper: NASH: Neural Architecture Search for Hardware-Optimized Machine Learning Models

Training

The code in the resnet18 and resnet34 folders is the code for training the networks.
Brevitas is needed to train the models. https://github.com/Xilinx/brevitas
A10 GPU is used in this design.
The environment of training is as follows:
brevitas 0.8.0
certifi 2022.12.7
charset-normalizer 3.1.0
dependencies 2.0.1
future-annotations 1.0.0
idna 3.4
numpy 1.24.2
onnx 1.11.0
onnxoptimizer 0.2.0
packaging 23.1
Pillow 9.5.0
pip 20.0.2
pkg-resources 0.0.0
protobuf 3.19.5
requests 2.28.2
setuptools 44.0.0
tokenize-rt 5.0.0
torch 1.8.2+cu111
torchvision 0.9.2+cu111
typing-extensions 4.5.0
urllib3 1.26.15
The environment may needs to be changed based on the GPU you use. The code train_search.py needs to be run first to train the architecture of the model.
The next step is to run train_imagenet.py. There is already architecture we have trained for the networks, so it is also possible to run train_imagenet.py without running train_search.py.

Pre-trained models

The pre-trained models are in the models folder. These models are already under the architecture we trained.

FINN generation

FINN compiler needs to be used to generate the models into hardware. https://github.com/Xilinx/finn
The code build.py under the finn folder is needed to be run to generate the bit file.
Alveo FPGA board is the board that needs to be used.