This is the course final project for CS420, which is a classification test on the facial expression dataset fer2013. The soruce code are in the src folder.
The project report has been also uploaded (Jan, 2020). The whole following information can be found in the report.
- Bohong Wu 吴薄鸿 516030910365
- Hao Sun 孙豪 516030910362
- Numpy
- matplotlib
- scikit-learn
- PyTorch
- Advertorch
You can split the data into training.csv, valdata.csv, and testdata.csv. Or (recommanded) you can download the data on this link with password zcy0
Before training any model, first make sure the relative paths are correct. We have provided our splited data in data folder.
- To train the basic VGG model,
python train_vgg.py - To train the HOG-embedded model,
python train_vgghog.py - To train a model with data augmented by Gaussian Noise,
python train_with_gaussian.py
The validation is provided with the training scripts.
The pretrained models are available on this link with no password.
The accuracy curves.
The loss curves.
| Type | Basic | HOG-embedded |
|---|---|---|
| Private accuracy | 72.67% | 72.56% |
| Public accuracy | 70.44% | 70.97% |
The confusion matrix.
LinfPGDA attack.
FGSM attack
| Method | Raw VGG | Add Gaussian | Add DAE | DAE embedded |
|---|---|---|---|---|
| Original | 65.28% | 68.32% | 49.04% | 65.12% |
| Adversarial | 49.04% | 54.86% | 46.04% | 57.15% |
To validate the results on the defenses methods, change the filepath in /src/acc_on_attackand modified the relative function. python acc_on_attack.py