Deep Learning with pytorch

• Foundations of diverse neural networks, AI SCHOOL course.
• I write this to show what i have studied!

Week 1

---

• About AI
  • What is AI?
  • History of AI
  • Examples of using AI
• About Python
  • Character of Python
  • Data types of Python (Integer, float, character ...)
  • How to use List?
  • String Formatting
• Basic Theory of Deep Learning
  • Process of Learning

    1. Evaluation (Forward propagation)

       • softmax function

    2. Loss and Gradient (Backpropagation)

       • Loss Function(Mean Squared Error & Cross Entropy error)

    3. Update (Optimize)

  • What is the perceptron? (and Multi-layer perceptron)
• Basic of Pytorch

Assignment

1. Print some string using formatting
2. Sorting a list

Week 2

---

• About Python

  • Dictionary?
  • What is Function? & How to use Function?
  • Control Statement
    • if
    • for

• Basic theory of Deep Learning

  • Gradient decent & Learning rate
  • Back propagation
  • Optimizer
    • SGD
    • AdaGrad
    • RMSProp
    • Momentum
    • Adam

• Basic of Pytorch

  • Developement of handwriting recognizer
    • it has a code.

Assignment

1. Making a Code

num = 5 
if <num이 짝수라면>:
    print('짝수입니다.')
else: 
    print('홀수입니다.')

2. Calculation Back propagation

Week 3

---

• About Python
  • List nesting('for' Statement)
  • tuple
  • 'While' Statement
  • File I/O
• 'Deep Learning' Learning Methodology
  • Early Stopping
  • Weight Initialization
  • Dropout
  • Parameter Norm Penalties
  • Data Augmentation
• Development of handwriting recognizer with hyperparameter adjustment
  • Epoch, Batch Size, Iterations
  • Training, Test Set
  • What is MNITST Dataset

Assignment

1. Using 'While' to make below problem

*
**
***
****
*****

2. Implement the below code in one line using list nesting.

number = [1,2,3,4,5]
result = []
for n in numbers:
    if n % 2 == 0:
        result.append(n+2)

3. Read the given fileIO.txt file, Key is Last Name and Value is Allocate information to the dictionary name_age, which is the age, and then print it.

4. Set my own Hyper Pameter to to learn MNIST Dataset

Week 4

---

• About Python
  • Class 1
• Theory of Deep Learning
  • CNN(Convolution Neural Network)
• Pythorch
  • Classification of MNIST data using CNN model

Assignment

1. Create a Calculator class that performs four arithmetic operations

Week 5

---

• About Python

  • Web crawling
    • Using Beautiful Soap Package
    • Using IMDB Dataset

• Theory of Deep Learning

  • Representative model using CNN
    • *Just simple Theory
    • LeNet
    • AlexNet
    • GoogleNet
    • ResNet

• Pythorch

  • Classification of CIFAR-10 data using LeNet

Assignment

1. Dynamic crawl of 100 dates and save date.txt (feat. IMDB Dataset)
2. Target 74 % by adjusting the hyperparameter(feat. LeNet)

Week 6

---

• About Python
  • Class 2
    • Character of Class : Inheritance
• Theory of Deep Learning
  • Batch Normalization
• Pytorch Practice
  • Implement ResNet 18

Assignment

1. Make ScienCalculator class that inherits Calculator class

   • Add function - pow()_Calculate squared
   • Redefine 'class' by using method overriding
   • Print the result by using 5 methods after creating ScienCalculator class

2. Reimplement ResNet - 18 and then train the network using CIFAR 10 Dataset

3. Implement ResNet - 101

Week 7

---

• About Python
  • Numpy
  • NLTK
  • pre-processing
    • tokenize
    • Eliminate stopwords
    • stemming
• Theory of Deep Learning
  • CNN-based text classification
    • 김훈의 CNN
• Pytorch Practice
  • Movie review rating prediction using CNN model

Assignment

1. Find the similarity between the 3 given sentences.
   • step
     • pre-processing
     • Implement Cosine similarity function
     • Make Doc-word matrix
     • Finding the Similarity
2. Target 90 % by adjusting Hyper parameter (CNN Classifier)

Week 8

---

• Using python
  • Collect Instagram posts
    • Hashtag-based image and text collection
• Theory of Deep Learning
  • Basic Theory of Word2Vec 1
    • What is one-hot encoding
    • CBOW
    • Skip-Gram
    • Why does it occur Imbalance overhead?
• Pytorch Practice -practicing Word2vec - Training Word2Vec through gensim

Assignment

1. Crawling Instagram posting images based more than 5 hashtags
2. confirm the result by calculating more than 10 word analogy
3. Try to score by putting more than 10 similarity pairs
   • ex) 'Benz' & 'BMW'

Week 9

---

• Data structure
  • Stack
    • Concept of Stack
    • How to use Stack functions in python
      • ex)append(), pop() ..
  • Queue
    • Concept of Queue
    • How to use Queue functions in python
      • ex) append(), popleft(), appendleft(), pop() ..
• Theory of Deep Learngin
  • Basic Theory of Word2vec 2
    • About Distributional hypothesis
    • Solutions for Imbalance overhead
      1. Hierarchical softmax
      2. Negative sampling
    • Additional Performance Enhancement Techniques
      • Subsampling
• Pytorch Practice
  • Calculate Word analogy score by training several cases

Assignment

1. Target 35%
   • By increasing news dataset and iter, raise the score of word analogy task
2. Coding Test about stack & queue on 'programmers'

Week 10

---

• Practice Algorithms
  • practice 1
  • practice 2
• Theory of Deep Learning
  • Basic Theory of RNN
    • About sequential data
    • Concept of Recurrent Neural Network(RNN)
    • RNN applications
      • Image Captioning
      • Sentiment classification
      • Machine translation
      • Question answering
      • Language modeling
    • Bidirectional RNN
    • Limitation of RNN
    • How to classify Text using RNN
• Pytorch Practice
  • Movie review rating prediction using RNN model

Assignment

1. Target Accuracy 87% by adjusting hyper parameters of RNN

Week 11

---

• Practice Algorithms
  • practice 1
  • practice 2
• Theory of Deep Learngin
  • LSTM (Long Short Term Memory)
    • limitations of RNN
    • concept of LSTM
      • Cell state?
      • Gate Mechanism
        • Forget gate layer
        • Output gate layer
        • Input gate layer
    • Update Old cell state to new cell state
• Pytorch Practice
  • Text classification using LSTM
    • Movie review rating prediction using LSTM

Assignment

1. Target Accuracy 88% by adjusting hyper parameters (feat. LSTM)

Week 12

---

• Practice Algorithms
  • practice 1
  • practice 2
  • practice 3
• Seq2Seq
  • About Seq2Seq
  • Basic NMT (Neural Machine Translation)
    • Using methods in NMT
      1. Beam Search
      2. Greedy Search
• Pytorch Practice
  • Translation based LSTM

Week 13

---

• Using Python
  • Crawling of Youtube videos
    • Use Youtube class that is in Pytube
• Attention mechanism for NMT
  • There are three types of Attention mechanism for NMT
    1. Basic attention mechanism for NMT
    2. Multiplicative attention mechanism for NMT
    3. Additive attention mechanism for NMT
  • Simple concept how to image captioning by Attention mechanism
• Pytorch Practice
  • Attention mechanism NMT

Week 14

---

• GAN (Generative Adversarial Network)
  • About GAN
    • Origin of the name ['Generative', 'Adversarial', 'Network']
    • About D(Discriminator) & G(Generator)
      1. Discriminator
         • Classification network classified as 0 or 1 based on data
      2. Generator
         • Distribution approximation function network that similarly expresses the distribution of training data
    • How to make G and D fight hostile?
    • Equation?
      • 
      • If Excellent D, Try to make the equation maximum(0)
      • If Excellent G, Try to make the equation minimum(−∞)
    • There are several types of GAN
      1. Vanilla GAN
      • concept
        • Use default MLP
      2. Conditional GAN
      • concept
        • Insert The label that is the correct answer into D & G network together
        • Label becomes a condition in the process of achieving its purpose! (both G & D)
      3. DCGAN
      • concept
        • Adjust CNN into GAN structure
        • Generator uses Transpose Convolutional Network
        • Discriminator uses Convolutional Network
        • It has guideline for stable Deep Convolutional GANs
        • Operation can be performed using a random number(z)
      4. Conditional DCGAN
      • concept
        • Conditioanl GAN + Deep Convolutional GAN
• Pytorch Practice
  • Train all types of GAN and then confirm the results.

Week 15

---

• Action Recognition

  • concept

    • A task to find out where the actions in the video belong to among actions within a specified category.

  • I train several datasets.

    • UFC101
    • HMDB51
    • Kinetics
    • Sport - 1M

  • Before we study Action Recognition, we know what is the characters of video data.

    1. What is the video data?
       • Set of images with a certain time interval
    2. The existence of a time term can contain information that only one image does not have.
    3. The meaning of spatial information in one image and The meaning of temporal information in images must be well interpreted.

  • There are two basic thesis on Action Recognition.

    1. Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset

       • CVPR 2017
       • Google DeepMind
       • Link

    2. A Closer Look at Spatiotemporal Convolutions for Action Recognition

       • CVPR 2018
       • FAIR (Facebook AI Research)
       • Link

  • Models for Action Recognition

    • Just simple theory.
    • Conv2d + LSTM
      • Large-scale Video Classification with Convolutional Neural Networks
    • Conv3d -Learning Spatiotemporal Features with 3D Convolutional Networks
    • Two stream -Two-Stream Convolutional Networks for Action Recognition in Videos
    • 3D - Fused Two Stream
      • Convolutional two-stream network fusion for video action recognition
    • R(2+1)D -A Closer Look at Spatiotemporal Convolutions for Action Recognition