accepted at CADL - ECCV2024 https://arxiv.org/abs/2410.00050

Before proceeding with the training or fine-tuning processes, it is essential to set up the environment correctly. Please use the following Docker image to ensure compatibility and reproducibility of results:

docker pull 1208977/random_cyclebnn

This Docker image contains all necessary dependencies and environments pre-configured for your convenience.

Below are detailed commands for training the ResNet18 architecture on different datasets. Please ensure that the data directories contain the correctly formatted datasets before proceeding.

To train a ResNet18 model on the ImageNet dataset, use the following command:

python main.py -save_folder ./logs --arch imagenet_resnet_18 --workers 5 --datadir ./imagenet2012 --is_cyclic_precision --cyclic_num_bits_schedule 2 6 --cyclic_num_grad_bits_schedule 8 8 --num_cyclic_period 8 --lr 0.001 --num_cyclic_annealing 1 --iters 9000000 --batch_size 16 --optimizer adamw --lr_schedule cosine --clamping_tau no --dataset imagenet --eval_every 1250

For training the ResNet18 model on the CIFAR-10 dataset, execute:

python main.py -save_folder ./logs --arch cifar10_resnet_18 --workers 5 --datadir ./data --is_cyclic_precision --cyclic_num_bits_schedule 2 6 --cyclic_num_grad_bits_schedule 8 8 --num_cyclic_period 8 --lr 0.001 --num_cyclic_annealing 1 --iters 187500 --batch_size 16 --optimizer adamw --lr_schedule cosine --clamping_tau no --dataset imagenet --eval_every 1250

To fine-tune a ResNet18 model specifically for the Pascal VOC dataset, please use the script provided:

./run.sh

This script will automatically adjust the model parameters for optimal performance on the Pascal VOC dataset.

After fine-tuning, you can test the performance of the ResNet18 model on the Pascal VOC dataset with:

./run_test.sh

This will evaluate the model against the dataset and provide performance metrics.

This work is based on contributions from several outstanding repositories. We extend our gratitude to the authors of the following repositories for their groundbreaking work in the field of efficient computing and binary neural networks:

• Efficient Computing - AdaBin by Huawei Noah's Ark Lab
• ReCU by Z. Xu
• BiDet by Ziwei Wang
• CPT by GATECH-EIC

Their contributions have significantly influenced the development of our project, and we highly recommend referring to their work for additional insights into binary neural networks and efficient computing.