DL_Spring_2019 國立交通大學電機工程學系，深度學習

Deep Learning Spring 2019 by Jen-Tzung Chien@ ECE NCTU Taiwan 授課教師：簡仁宗 (PS. This is my first time taking a Deep Learning class, the code might not be optimal, but I have tried my best to make it readable and clean)

Course website here 課程網站 點此

HW lists and details 作業解答以及各作業內容

HW score shown as below

• HW1, Score 98/100

  • Spec PDF
  • Handcraft a DNN classifier from scratch, without any assistence from python module, i.e. no import Keras, torch nor tenserflow.
  • An 0-1 classifier from the famous titanic dataset in Kaggle
  • Analyze the correlation-coeffiecnt and find out the principal column element.
  • Discuss the need of one-hot encoding.
  • LaTeX report
  • LaTeX report source code

• HW2, Score 100/100

  • Spec PDF
  • CNN
    • A traditional CNN classifier to classify the image of various types of animals (CUDA and NVIDIA GPU is required for accelerating the tasks.)
    • Background knowledge with image processing might help this homework.
    • PyTorch is allowed and used in this homework, but auto model build framework such as autoML is not allowed.
  • RNN / LSTM
    • Use RNN / LSTM for text analyzing and judge whether a paper with certain title will get accepted or not.
    • Word vector is important, which is the fundamental of NLP
    • Compare the performance b/w RNN and LSTM as well as discussing the reason behind it.
    • Try explain the reason of Gradient Vanishing or Gradient Exploding
  • LaTeX report
  • LaTeX report source code

• HW3, Score 95/100

  • Spec PDF
  • Variational Autoencoder
    • Use VAE to compress images and resconstruct them.
    • Background knowledge with probabilities and statistics might help this homework.
  • CycleGAN
    • Implement this paper
    • Use CycleGAN to fake generate some anime/cartoon character from the other style.
  • LaTeX report
  • LaTeX report source code

Final Project 期末專題

• A Style Transfer to make picture of someone younger / older.
• The original idea is taken from [here].(https://www.pytorchtutorial.com/pytorch-style-transfer/)
• Our team proposed 2 ways to implement, the CycleGAN from above and Neural Style Transfer from here.
• LaTeX report
• LaTeX report source code