This is PyTorch implementation of the paper: "An Efficient Lightweight Neural Network for Facial Expression". This work proposes a lightweight network that combines a high-performance lightweight backbone by applying ZenNAS [1] and a distillation feature part after backbone to refine features. Extensive experimental results demonstrate that our ZenNet-SA achieves new state-of-the-art results on three FER benchmark datasets, RAF-DB (88.53%) as well as AffectNet 7 class (65.23%) and 8 class(61.32%) in the light-weight models group with a model size less than 1M and computation costs approximately 450 M Flops. Our results demonstrate a noteworthy improvement in efficiency, with a decrease of about 97% in FLOPS and 91% in parameters when compared to POSTER V2 in terms of computational and resource expenses.
The performance of ours ZenNet-SA on RAF-DB dataset. (*) denoted the case using transfer-learning
| model | resolution | # params | FLOPs | Top-1 Acc |
|---|---|---|---|---|
| ZenNet-M-SA | 224 | 1.01M | 400.550M | 87.1% |
| ZenNet-R-SA | 224 | 0.33M | 446M | 87.26% |
| ZenNet-M-SA (*) | 224 | 1.01M | 400.550M | 88.53% |
| ZenNet-M-SA (*) | 224 | 0.33M | 446M | 87.39% |
The performance of ours ZenNet-SA on AffectNet-7cls dataset
| model | resolution | # params | FLOPs | Top-1 Acc |
|---|---|---|---|---|
| ZenNet-M-SA | 224 | 1.01M | 400.550M | 63.94% |
| ZenNet-R-SA | 224 | 0.33M | 446M | 65.23% |
| ZenNet-M-SA (*) | 224 | 1.01M | 400.550M | 64.57% |
| ZenNet-M-SA (*) | 224 | 0.33M | 446M | 64.77% |
The performance of ours ZenNet-SA on AffectNet-8cls dataset
| model | resolution | # params | FLOPs | Top-1 Acc |
|---|---|---|---|---|
| ZenNet-M-SA | 224 | 1.01M | 400.550M | 60.74% |
| ZenNet-R-SA | 224 | 0.33M | 446M | 61.32% |
| ZenNet-M-SA (*) | 224 | 1.01M | 400.550M | 61.40% |
| ZenNet-M-SA (*) | 224 | 0.33M | 446M | 60.82% |
To more deeper insight NAS, we highly recommend you go to the ZENNAS work.
scripts/ZenNAS-FER-M-searching.sh
scripts/ZenNAS-FER-R-searching.sh- RAF-DB
- AffectNet
'--use_lw_att' help to add distillation block into ZenNet backbone
if you want to inject SA block into all super-block in backbone, you can use the parameter '--use-sa-all'.
scripts/Train-ZenNet-SA-M-FER-rafdb.sh
scripts/Train-ZenNet-SA-M-FER-AffectNet.sh
scripts/Train-ZenNet-SA-R-FER-rafdb.sh
scripts/Train-ZenNet-SA-R-FER-AffectNet.shThe links of ours pretrain model here!.
Our code is based on the official source code of Zen-NAS for generated backbone
https://github.com/idstcv/ZenNAS
If you find this work useful, please cite the following paper:
@INPROCEEDINGS{10773246,
author={Nguyen, Trang and Nabi, Hamza Ghulam and Han, Ji-Hyeong},
booktitle={2024 24th International Conference on Control, Automation and Systems (ICCAS)},
title={ZenNet-SA: An Efficient Lightweight Neural Network with Shuffle Attention for Facial Expression Recognition},
year={2024},
volume={},
number={},
pages={55-60},
keywords={Human computer interaction;Deep learning;Costs;Face recognition;Computational modeling;Neural networks;Education;Medical services;Benchmark testing;Monitoring;Facial expression recognition;light-weight model},
doi={10.23919/ICCAS63016.2024.10773246}}