English | 简体中文

[TOC]

一、Introduction

An implementation of the face (emotion and gender) classification models (MiniXception。SimpleCNN) proposed in paper Real-time Convolutional Neural Networks for Emotion and Gender Classification with PaddlePaddle. SimpleCNN is a standard fully-convolutional neural network composed of 9 convolution layers, ReLUs, Batch Normalization and Global Average Pooling. MiniXception replaces the convolution layers with depth-wise separable convolutions and residual modules.

二、Accuracy

We trained the models with imdb_crop dataset in gender classification task and each model obtains an accuracy of 96%.

Model	Accuracy	Input shape
SimpleCNN	96.00%	(48, 48, 3)
MiniXception	96.01%	(64, 64, 1)

三、Dataset

We trained and tested models on dataset imdb_crop (the password is mu2h). The dataset can be also download from here. First, download and uncompress the dataset. Then, edit configuration files config/simple_conf.yaml and config/min_conf.yaml of models SimpleCNN and MiniXception. Set imdb_dir to be the path to the dataset. imdb_dir s should be the same in training and test stages. You don't need to split the dataset into training and test set, because the python scripts will do that. The dataset will be split in the manner of that proposed in the paper. That is, sorts the images by file names and considers the front 80% as training set and the rear 20% as test set.

四、Environment

scipy==1.2.1
paddlepaddle==2.1.2
numpy==1.20.1
opencv-python==3.4.10.37
pyyaml~=5.4.1
visualdl~=2.2.0
tqdm~=4.62.0

五、Quick Start

Step1: Clone

# clone this repo
git clone https://github.com/wapping/FaceClassification.git
cd FaceClassification

Step2: Train

Edit the configuration file for your own and run the command like

python train.py -c path_to_conf

For example

python train.py -c ./config/simple_conf.yaml

Step3: Test

Edit the configuration file for your own and run the command like

python eval.py -c path_to_conf

Just wait for the results.

六、Code Structure and Explanation

6.1 Code Structure

|____config
| |____conf.yaml
| |____confg.py
| |____simple_conf.yaml
| |____mini_conf.yaml
|____data
| |____dataset.py
|____models
| |____simple_cnn.py
| |____mini_xception.py
|____train.py
|____eval.py

6.2 Parameter Explanation

• train.py

  --conf_path: optional, the path to the configuration file, config/conf.yaml by default.

  --model_name: optional, the model name. If given, it will replace model name in the configuration file.

• eval.py --conf_path: optional, the path to the configuration file, config/conf.yaml by default.

  --model_name: optional, the model name. If given, it will replace model name in the configuration file. --model_state_dict: optional, the path to the model. If given, it will replace model_state_dict in the configuration file.

七、Model Infomation

Field	Content
Author	Huaping Li、Xiaoqian Song
Date	2021.09
Framework version	paddlepaddle 2.1.2
Application scenarios	Face classification
Supported hardware	CPU、GPU