Skip to content

Labs for 5003 Deep Learning Practice course in summer term 2021 at NYCU.

Notifications You must be signed in to change notification settings

joycenerd/Deep_Learning_Practice_labs

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

75 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Deep_Learning_Practice_labs

Labs for 5003 Deep Learning Practice course in summer term 2021 at NYCU.

Back-propagation

Understand and implement simple neural networks with forwarding pass and backward propagation using only two hidden layers.

  • Linear data accuracy: 99%
  • XOR data accuracy: 100%

📄 Report 💻 Code

EEG Classification

Implementing two EEG classification model which are EEGNet and DeepConvNet. Also, we are changing the activation function (ELU, ReLU, LeakyReLU) in the model and see the difference.

  • EEGNet accuracy: 87.87%
  • DeepConvNet accuracy: 76.48%

📄 Report 💻 Code

Diabetic Retinopathy Detection

Analyze diabetic retinopathy using ResNet, compare the results of ResNet18, ResNet50, and both networks with pre-trained weights.

  • ResNet50 with pretrained accuracy: 82.18%
  • ResNet18 with pretrained accuracy: 79.24%
  • ResNet50 w/o pretrained accuracy: 73.55%
  • ResNet18 w/o pretrained accuracy: 73.35%

📄 Report 💻 Code

Conditional sequence-to-sequence VAE

Implementing conditional seq2seq VAE for English tense conversion and generation.

monotonic KL annealing

  • BLEU score: 0.8312
  • Gaussian score: 0.472

cyclical KL annealing

  • BLEU score: 0.9527
  • Gaussian score: 0.452

📄 Report 💻 Code

Let's Play GANs

Implement conditional GAN to generate synthetic images in multi-label conditions.

  • SAGAN acc: 77.22%
  • cDCGAN acc: 68.89%
  • WGAN acc: 60.56%

📄 Report 💻 Code

Deep Q-Network and Deep Deterministic Policy Gradient

Implement DQN and DDQN to solve LunarLander-v2. Implement DDPG to solve LunarLanderContinuous-v2.

  • DQN average reward: 277
  • DDPG average reward: 273
  • DDQN average reward: 256

📄 Report 💻 Code