Stanford CS231n (HKUST COMP4901J) Assignment Repo

Score: Full mark + all bonus

Grade: A+

Introduction

This repository stores the assignments I have done in HKUST COMP4901J - Deep Learning for Computer Vision, which is a course adapted from Stanford CS231n - Convolutional Neural Networks for Visual Recoginition.

There are four sets of assignments in total, in which the first three are adapted from Stanford, and the fourth is created by HKUST.

Assignment 1

• knn.ipynb: k-NN with Cross Validation using pure Numpy.
• svm.ipynb: SVM classifier on CIFAR-10 using pure Numpy.
• softmax.ipynb: Softmax classifier on CIFAR-10 using pure Numpy.
• two_layer_net.ipynb: NN with 2 FC layers on CIFAR-10 using pure Numpy.
• feature.ipynb: Training classifiers on CIFAR-10 with image features (Histogram of Gradients, Color Histrogram) using pure Numpy.
• Bonus 1 (in feature.ipynb): Further training networks with different combinations of features.
• Bonus 2 (in feature.ipynb): Using a NN with 2 FC layers on MNIST using pure Numpy.

Assignment 2

• FullyConnectedNets.ipynb: Neural Network with n FC layers and ReLU activations on CIFAR-10 with various optimization algorithms (SGD, momentum, RMSProp and Adam) using pure Numpy.
• Dropout.ipynb: Dropout layer implementation in pure Numpy.
• BatchNormalization.ipynb: Batch Normalization layer implementation in pure Numpy.
• ConvolutionalNetworks.ipynb: CNN implementation in pure Numpy.
• Tensorflow.ipynb: Training networks on CIFAR-10 using tensorflow. Achieved 80% on test set.
• Bonus 1 (in FullyConnectedNets.ipynb): Early stopping in model training.
• Bonus 2 (in ConvolutionalNetworks.ipynb): Average Pooling layer implementation in pure Numpy.
• Bonus 3 (in Tensorflow.ipynb): CNN kernel visualizations.

Assignment 3

• RNN_Captioning.ipynb: Simple RNN implementation using pure Numpy on Coco dataset for image captioning.
• LSTM_Captioning.ipynb: LSTM implementation using pure Numpy on Coco dataset for image captioning.
• NetworkVisualization-Tensorflow.ipynb: Saliency Map, Fooling Images and Class Visualization on tensorflow.
• StyleTransfer-Tensorflow.ipynb: Style Transfer network implementation on tensorflow.
• GANs-Tensorflow.ipynb: Ordinary GAN, LSGAN, DCGAN implementations on tensorflow.
• Bonus 1 (in GANs-Tensorflow.ipynb): WGAN-GP implementation on tensorflow.
• Bonus 2 (in Bonus-CGAN.ipynb): CGAN implementation on tensorflow.
• Bonus 3 (in Bonus-InfoGAN.ipynb): InfoGAN implementation on tensorflow.
• Bonus 4 (in Bonus-SemiSupervisedLearning.ipynb): GAN for semi-supervised learning on MNIST. Result is not the greatest.

Assignment 4

• Q_learning_basic.ipynb: Q-table and Q-learning on gym using tensorflow.
• DQN_WorldNavigate.ipynb: DQN, Double/Dueling DQN on gym using tensorflow.
• PG_CartPole.ipynb: Policy Gradient on gym using tensorflow.
• Model_Policy_Network.ipynb: Model-based Policy Gradient on gym using tensorflow.
• Bonus 1 (in Bonus-DQRN.ipynb): DQRN implementation on gym using tensorflow.
• Bonus 2 (in Bonus-DQN_WorldNavigate-HyperparameterTuning.ipynb): DQN with tuned hyperparameters.
• Bonus 3 (in Bonus-Model_Policy_Network-HyperparameterTuning.ipynb): Model-based Policy Gradient network with tuned hyperparamenters.
• Bonus 4 (in Bonus-DQN_WorldNavigate-Breakout.ipynb): DQN on Atari Breakout. Result is not the best.
• Bonus 5 (in Bonus-Curiosity-MountainCar.ipynb): Curiosity driven RL implementation using tensorflow.