Readme.md

Facial-Expression-Recognition.Pytorch

A CNN based pytorch implementation on facial expression recognition (FER2013 and CK+), achieving 73.112% (state-of-the-art) in FER2013 and 94.64% in CK+ dataset

2021年4月10日15:33:17 更新

文档：Facial-Expression-Recognition.Pytorc... 链接：http://note.youdao.com/noteshare?id=0b97c224ba866d3cfc8911a3726cd33e&sub=9E9FD1BC60754F27A53085C55FDA737D

使用cpu

根据一位大佬的修改，我把他的代码抄到visualize_cpu.py 里面了，于是用下面这句话运行

python visualize_cpu.py --cpu

这位大佬（LittleWat ）把代码改成可以cpu运行了

Demos

Dependencies

文档：运行 WuJie1010_Facial-Expression-Rec... 链接：http://note.youdao.com/noteshare?id=dc290f5316d968af294bff094814c68b&sub=4D4656B3590B40F7921613D6E71EB700

• Python 2.7
• Pytorch >=0.2.0
• h5py (Preprocessing)
• sklearn (plot confusion matrix)

Visualize for a test image by a pre-trained model

• Firstly, download the pre-trained model from https://drive.google.com/open?id=1Oy_9YmpkSKX1Q8jkOhJbz3Mc7qjyISzU (or https://pan.baidu.com/s/1gCL0TlCwKctAy_5yhzHy5Q, key: g2d3) and then put it in the "FER2013_VGG19" folder; Next, Put the test image (rename as 1.jpg) into the "images" folder, then
• python visualize.py

FER2013 Dataset

• Dataset from https://www.kaggle.com/c/challenges-in-representation-learning-facial-expression-recognition-challenge/data Image Properties: 48 x 48 pixels (2304 bytes) labels: 0=Angry, 1=Disgust, 2=Fear, 3=Happy, 4=Sad, 5=Surprise, 6=Neutral The training set consists of 28,709 examples. The public test set consists of 3,589 examples. The private test set consists of another 3,589 examples.

Preprocessing Fer2013

• first download the dataset(fer2013.csv) then put it in the "data" folder, then
• python preprocess_fer2013.py

Train and Eval model

• python mainpro_FER.py --model VGG19 --bs 128 --lr 0.01

plot confusion matrix

• python plot_fer2013_confusion_matrix.py --model VGG19 --split PrivateTest

fer2013 Accurary

• Model： VGG19 ; PublicTest_acc： 71.496% ; PrivateTest_acc：73.112%
  
• Model： Resnet18 ; PublicTest_acc： 71.190% ; PrivateTest_acc：72.973%

CK+ Dataset

• The CK+ dataset is an extension of the CK dataset. It contains 327 labeled facial videos, We extracted the last three frames from each sequence in the CK+ dataset, which contains a total of 981 facial expressions. we use 10-fold Cross validation in the experiment.

Train and Eval model for a fold

• python mainpro_CK+.py --model VGG19 --bs 128 --lr 0.01 --fold 1

Train and Eval model for all 10 fold

• python k_fold_train.py

plot confusion matrix for all fold

• python plot_CK+_confusion_matrix.py --model VGG19

CK+ Accurary

• Model： VGG19 ; Test_acc： 94.646%
  
• Model： Resnet18 ; Test_acc： 94.040%