CS-420- Facial Expression Recogition Task

Zhang Haoqi 516030910393

Jin Kaiqi 516030910382

Classifier: Model Averaging of VGG19 and ResNet18

In the folder "Classifier"

In this part, we train the VGG19 and RenNet18 and use model averaging idea to get a improvement. You can download pretrained model in https://jbox.sjtu.edu.cn/l/iJFKfe (password:itxu)

Usage

Proprocess/Generate data

put dataset(fer2013.csv) in "data" folder, then

python preprocess.py

Training

python main.py --bs 128 --lr 0.01 --model VGG19

Testing

Use the idea of model averaging.

python model_average.py --model VGG19 --model ResNet18 --bs 128

Experiment

Model	PrivateTest_acc(%)
VGG19	72.834
ResNet18	73.391
Model Averaging(VGG19 + ResNet18)	73.725

Attack the classifier and Defense

In the folder "attack & defense"

We apply DeepFool to attack the trained model and put the attack samples into training dataset to make the model more robust after model retraining. You can download pretrained model in https://jbox.sjtu.edu.cn/l/EnSEzf (password:cath)

Usage

Attack

python test_deepfool.py --model VGG19 --split PrivateTest 

Defense

python defense.py --model ResNet19 --attmodel VGG19 

Experiment

Model	Attack Accuracy
VGG19-Before Retraining	13.5%
VGG19-After Retraining	17.7%