Assignments done throughout Ph. D. course at Yonsei University. Codes may not be perfect according to the description. Reports are available upon request.
NN_Assignment_1.ipynb- Drop wave cosine function approximation using linear regression.NN_Assignment_2.ipynb- Convolutional Neural Networks (CNN) training implementation on CIFAR-10 dataset.NN_Assignment2_1.ipynb- Transfer learning using CIFAR-10 pretrained model on face dataset. Face dataset is inface.zip.NN_Assignment2_2.ipynb- LSTM implementation to predict bitcoin closing price. The dataset is inbtc.csv.STDL_Assignment_1.ipynb- Develop 3 types of customized CNN, namely a normal self-defined CNN, CNN with 1x1 convolution, CNN with depth-wise separable convolution.STDL_Assignment_2.ipynb- Develop a 1D Generative Adversarial Network (GAN) to generate new datasets, given a set of old datasets. The implementation is based on this tutorial from this blog.STDL_Assignment2_1.ipynb- Knowledge Distillation (KD) from a teacher network to a student network. VGG16 is used for the teacher network (using Keras sequential), while the student network is self-defined (as inSTDL_Assignment_1.ipynb). The distillation class is defined based on Keras documentation. CIFAR-10 is used to train the networks. The distillation works but the results are still not satisfactory. (There are future plans to implement a TensorFlow version of this question).STDL_Assignment2_2.ipynb- Implementation of 3 continuous learning methods: Feature Extraction, Fine Tuning, Learning Without Forgetting (LWF), to see performance of the network on old task (CIFAR-10 dataset as inSTDL_Assignment2_1.ipynb) and new task (face dataset, usingface.zip). The results are not good, due to implementation issues halfway. (Future plans to implement a TensorFlow version of this question, similar toSTDL_Assignment2_1.ipynb).
Homework_1.ipynb- Image Discrete Fourier Transform (DFT) in 8x8 blocks, then performing subsampling and interpolation.Homework_2.ipynb- Discrete Cosine Transform (DCT), Singular Value Decomposition (SVD), Karhunen-Loeve Transform (KLT).Homework_3.ipynb- Image Differential Pulse Code Modulation (DPCM) and Lloyd-Max Quantization. (Huffman Coding Part is incomplete).Homework_4.ipynb- Implementation of median filter and FIR filter, specifically using Hamming Window to denoise images with different types of noise.Homework_5.ipynb- Image Gamma Correction, Histogram Equalization, conversion to YCrCb format.Homework_6.ipynb- Inverse filter, Wiener filter, constrained matrix inversion filter to deblur cylindrical boundary blurred and Gaussian blurred images.Homework_7.ipynb- Image segmentation using mean values and Otsu thresholding, object and edge detection using Sobel and Canny filters.
llgc.ipynb- Implementation of Learning with Local and Global Consistency (LLGC) on two-spirals dataset. The LLGC paper is in this website, while the code implementation is based on this Git.
transform_theory.ipynb- Design a simple self-defined 2-channel filter bank with perfect reconstruction to perform image compression (with 3 level octave filter banks). PyWavelets is used to perform Discrete Wavelet Transform (DWT) of the designed filter on the images to perform the compression.
machine_learning.ipynb- Implement a simple CNN for classification on CIFAR-10 dataset.
STPR_Homework_1.ipynb- Implementation of Linear and Multinomial regressions on a two-class multivariate Gaussian dataset. The boundary line for the multinomial regression (5th order) is not optimal, and weight dimension reduction has to be performed.STPR_Homework_2.ipynb- Implementation of LSE-Linear, LSE-RM (as in Homework 1#), and TER-RM classification on Iris, Mushroom, and Optical Recognition of Handwritten Digits dataset. The data preprocessing are extracted manually using Python scripts (though it is simpler to do so using Pandas library). The TER-RM performs classification on positive and negative (subject to that certain) classes, and relies on one-versus-all classification technique.