Project structure

Python modules:

• LocalDNAS.py: impementation of both the SuperNet method and cost-aware regularizers
• alternatives.py: alternative modules considered in supermodules
• utils.py: utility module

Jupyter notebooks:

• SuperNet_MBNV3.ipynb: training strategy for MBNV3 on ImageNet-100 and application of SuperNet to MBNV3 targeting SE modules
• SuperNet_regularizers.ipynb: cost-aware optimizations of SuperNet on MBNV3 by considering cost-aware regularizers on number of parameters and FLOPs
• model_comparison.ipynb: comparison between standard MBNV3 and models achieved with the application of SuperNet

Run experiments

1 - Create conda enviroment

1. Create and activate the conda environment:

conda create -n cv-project python==3.11 -y
conda activate cv-project

2. Install requirements.txt file:

pip install -r requirements.txt

2 - Download ImageNet100 dataset in the project folder

1. Move to the project folder and download the dataset using Kaggle:

kaggle datasets download ambityga/imagenet100

2. Unzip imagenet100.zip:

mkdir ImageNet
unzip imagenet100.zip -d ImageNet