This repository is the official implementation of Efficiera Residual Networks (ERNs). ERNs is a fully w1a2 image classification model including the input and output layer. ERNs achieves competitive performance compared to the state-of-the-art ultra-low-bit quantized models.
ImageNet accuracy of the pretrained models. Note that the accuracy may depend on the environment such as GPU architecture.
| Model | Model Size (Mbytes) | Top1 w/o TTA | Top5 w/o TTA | Top1 with TTA | Top5 with TTA |
|---|---|---|---|---|---|
| ERNs18-0.75x | 0.98 | 60.7 | 83.6 | 63.6 | 85.5 |
| ERNs18 | 1.4 | 62.7 | 84.9 | 65.3 | 86.8 |
| ERNs34 | 2.6 | 66.7 | 87.8 | 70.0 | 89.7 |
| ERNs50 | 3.1 | 70.3 | 90.0 | 72.5 | 91.3 |
| ERNs101 | 5.6 | 72.2 | 91.1 | 73.8 | 92.1 |
We provide the library as a python wheel file (efficiera_residual_networks_library-1.0.0-py3-none-any.whl).
We confirmed that the evaluation of the models work under the following conditions
- python 3.8.0
- NVIDIA DGX Server Version 4.7.0
- NVIDIA-SMI 515.86.01
- CUDA Version 11.7
python3.8 -m venv ern_venv
source ./ern_venv/bin/activate
git clone [email protected]:LeapMind/ERN.git
cd ERN
pip install pip==23.3.2 --upgrade # the installation of the wheel package may fail with version 24.0 and higher.
pip install efficiera_residual_networks_library-1.0.0-py3-none-any.whl
To train and evaluate models, ImageNet dataset is required.
root parameter in configs/dataset/imagenet.yaml needs to be modified.
The model can be trained from scratch using train.py.
python train.py +experiment=ERNs18x075_imagenet_best
The experiment is managed by hydra. The experiment settings are managed in configs/experiment.
The best experiment settings for each models are as follows.
- ERNs18x075:
configs/experiment/ERNs18x075_imagenet_best.yaml - ERNs18:
configs/experiment/ERNs18_imagenet_best.yaml - ERNs34:
configs/experiment/ERNs34_imagenet_best.yaml - ERNs50:
configs/experiment/ERNs50_imagenet_best.yaml - ERNs101:
configs/experiment/ERNs101_imagenet_best.yaml
The trained model can be evaluated using evaluate.py
python evaluate.py +experiment=ERNs18x075_imagenet_best checkpoint_filepath=</path/to/checkpoint_file.ckpt>
The results of Table 2 in our paper are confirmed with the following commands. Note that the accuracy may be slightly different depending on the environment such as GPU architecture.
python evaluate.py +experiment=ERNs18x075_imagenet_best checkpoint_filepath=./Efficiera_Residual_Networks_Checkpoints/ern18x075/checkpoints/last.ckpt input_image_sizes="[[308, 308]]" pl_module.tencrop_evaluation=true pl_module.tencrop_size=288 training.batch_size=25
python evaluate.py +experiment=ERNs18_imagenet_best checkpoint_filepath=./Efficiera_Residual_Networks_Checkpoints/ern18/checkpoints/last.ckpt input_image_sizes="[[308, 308]]" pl_module.tencrop_evaluation=true pl_module.tencrop_size=288 training.batch_size=5
python evaluate.py +experiment=ERNs34_imagenet_best checkpoint_filepath=./Efficiera_Residual_Networks_Checkpoints/ern34/checkpoints/last.ckpt input_image_sizes="[[306, 306]]" pl_module.tencrop_evaluation=true pl_module.tencrop_size=288 training.batch_size=25
python evaluate.py +experiment=ERNs50_imagenet_best checkpoint_filepath=./Efficiera_Residual_Networks_Checkpoints/ern50/checkpoints/last.ckpt input_image_sizes="[[308, 308]]" pl_module.tencrop_evaluation=true pl_module.tencrop_size=288 training.batch_size=25
python evaluate.py +experiment=ERNs101_imagenet_best checkpoint_filepath=./Efficiera_Residual_Networks_Checkpoints/ern101/checkpoints/last.ckpt input_image_sizes="[[308, 308]]" pl_module.tencrop_evaluation=true pl_module.tencrop_size=288 training.batch_size=25