This is an implementation for the first project of CS385 Machine Learning at Shanghai Jiao Tong University, instructed by Prof. Quanshi Zhang and Dr. Xu Cheng. I implemented some linear, kernel-based models as well as CNNs to do face classification, and also a sliding-window-based face detector.
The major dependencies of this project include:
- Python 3.6
- Numpy 1.14.0 +
- Scipy 0.19.0 +
- Sklearn
- Skimage
- h5py
- cv2
- ThunderSVM (The GPU version of SVM library, please refer to official document for installation)
- Numba (Used for LDA calculation acceleration)
- PyTorch 1.0 +
Notice: The SVMs are trained with ThunderSVM library, which is a little bit difficult to install. You may change my code in main.py to use a CPU version of SVM with sklearn or directly contact me.
You should download the FDDB dataset from here(579 M) and the annotations from here(161 K) and place them in the data folder under the root folder.
Then, you should extract the files with
tar -zxvf originalPics.tar.gz
tar -zxvf FDDB-filds.tgz
After doing that, you should make sure that the data folder looks like:
- data
---- fddb
-------- 2002
-------- 2003
-------- FDDB-folds
------------ FDDB-fold-01-ellipseList.txt
Then, please go to the data_processing folder. There's a file gen_bbox.py. You should modify the base_dir global parameter on line 9 to adapt to your own computer. After that please run:
python gen_bbox.py
We will automatically generate hdf5 datasets for training/evaluation. Please prepare around 800 MB disk space to hold the generated hdf5 files.
You can run main.py to evaluate our results. There are several flags in this file:
- model: The model to use. Please choose between "logistic", "svm", "lda" and "cnn".
- svm: The SVM kernel to use. Enabled only when model = "svm". Please choose between "linear", "rbf" and "poly".
- detection: Whether to run detection. Please choose between "True" and "False".
- vishog: Whether to visualize HOG features. Please choose between "True" and "False".
- vissv: Whether to visualize supporting vectors. Enabled only when model == "svm". Please choose between "True" and "False".
- train: Whether to train the model. For SVMs, you must train the model; for others, we have prepared the pretrained checkpoint under the root folder. Please choose between "True" and "False".
A sample code for running:
python main.py --model logistic --detection False --vishog True --vissv False --train False
You may modify it as you wish. If you choose detection = True, the results will be saved under the root folder.
You may run python visualize.py to visualize the distribution of HOG features (PCA/t-SNE). You can also run python visualize_cnn.py to get the t-SNE visualization for CNN features.
If you have any questions, please contace me through email: kentang AT sjtu DOT edu DOT cn.