Deep-Learning-and-Practice (DLP)

Teachers: Wen-Hsiao Peng, Tong-Shen Chen, I-Chen Wu

This repo contains the lab code in the 2022 Deep-Learning-and-Practice course. For the detail of each lab please refer to its report.

Lab02

Keyword : Back propagation

Implement the back propagation from scratch and train the network to classify the linear and non-linear data.

-Linear data

accuracy: 99%

-Non-linear XOR data

accuracy:99%

---

Lab03

Keyword : Reinforcement learning, Q-learning, N-tuple network

Mastering the 2048 game by reinforcement-learning and N-tuple network.

Score in 10000 times playing the 2048 game after training:

---

Lab04-1

Keyword : EEGNet, activation function

Implement the EEGNet and DeepConvNet to classify the BCI dataset. In this lab, different activation function were tested for the preformance.

-EEGNet

accuracy: 87.4% with LeakyReLU

-DeepConvNet

accuracy:81.76 with ELU

---

Lab04-2

Keyword : Resnet

Trained the classifier with Resnet18 and Resnet50 to detect the Diabetic Retinopathy.

-Resnet18

accuracy: 82.47% with pretrained weight

-Resnet50

accuracy: 83.15% with pretrained weight

---

Lab05

Keyword : VAE

Inplement the conditional variational autoencoder(VAE) to predict the motion frame of robot end effector.

-Ground truth

-Predict result

psnr : 23.6 dB

---

Lab06

Keyword : Deep-reinforcement-learning

Master the LunarLander game with reinforcement-learning, using DQN, DDPG, DDQN and TD3.

-Result

LunarLander score of each RL model

---

Lab07

Keyword : GAN, WGAN, WGAN-GP

Implenet the conditional generative adversarial network(GAN) to generate the specific synthetic images.

-Generated images: (from only words such as 'red cylinder, blue cube... etc')

---

Final

Keyword : Self-supervised learning, unsupervised learning

We dicussed the to pretrained a classifiaction model. Three pretrained weights were generated from autoencoder, K-means and simCLR seperately. Then the weights were used to fine-tune and detects on covid-19 detection task. We discuss that the variety of the pretraining dataset is more important to the performance of down-stream task.

-Result