Deep Learning Practice

Basically neural network based implementation and corresponding notes.

More "general" machine learning notes will be noted in my Machine Learning repository.

If you want to clone this repository, please use the following command:

GIT_LFS_SKIP_SMUDGE=1 git clone https://github.com/daviddwlee84/DeepLearningPractice.git

The notes of this repository haven't updated for a long time, I will update it once I organize my local notes.

Environment

• Using Python 3

Dependencies

• tensorflow
  • github
  • Brief Notes - Placeholder, Graph, Session
  • TensorFlow 2.0 Notes
  • Model Save and Restore Notes - ckpt, transfer learning
  • Data Manipulating Notes - TFRecord, Iterator
  • Multi-thread Notes
  • High-level API Notes - tf.keras, tf.layer
  • simple demos with maybe jupyter notebook?!
• keras
  • github
  • Brief Notes
• pytorch
  • github
  • Brief Notes
  • torch friends
  • tensorboardX - tensorboard for pytorch (and chainer, mxnet, numpy, ...)
  • pytorch-lightning - The lightweight PyTorch wrapper for ML researchers. Scale your models. Write less boilerplate
  • tnt - is torchnet for pytorch, supplying you with different metrics (such as accuracy) and abstraction of the train loop
  • inferno and torchsample - attempt to model things very similar to Keras and provide some tools for validation
  • skorch - is a scikit-learn wrapper for pytorch that lets you use all the tools and metrics from sklearn

Project

PKU Courses and Some side projects

• Basically based on TensorFlow 1.x and Keras
• Begin with the most basic model > CV > NLP

Subject	Technique	Framework	Complexity	Remark
Perceptron Practice	SLP, MLP	Numpy	○○●●●	Truth Table (AND, OR, XOR) and Iris Dataset (simulate Keras API)
Softmax Derivation	FCNN	Numpy	○○○●●	Backpropagation of Softmax with Cross Entropy Loss
MNIST Handwriting Digit	FCNN	Tensorflow (and tf.keras)	○○●●●	Implement by different ways
Semeion Handwritten Digit	FCNN	Tensorflow	○○○●●	Made a Tensorflow like Dataset Class
CIFAR-10	FCNN, CNN	Tensorflow	○○●●●	Comparison of FCNN and CNN
Chinese Named Entity Recognizer	RNN, LSTM	Tensorflow	○●●●●	TODO: LSTM testing
Flowers	CNN	Tensorflow	○○●●●	Transfer Learning
Fruits	CNN	Tensorflow (and tf.layer)	○○●●●	Multi-thread training and TFRecord TODO: Try more complex model
Trigonometric Function Prediction	RNN	Tensorflow	○○○○●	Predict sine, cosine using LSTM
Penn TreeBank	RNN, LSTM	Tensorflow	○○●●●	Language corpus preprocessing and training
Chinese Neural Machine Translation	RNN, Attention	Tensorflow	○●●●●	A practice of Seq2Seq and Attention TODO: Multi-graph, Try transformer
Dogs!	CNN	Keras	○○●●●	Using images from ImageNet, Keras Transfer learning and Data augmentation
2048	FCNN with Policy Gradient	Tensorflow	●●●●●	Reinforcement Learning
Text Relation Classification	Multiple Models	Multiple Libraries	●●●●●	SemEval2018 Task 7 Semantic Relation Extraction and Classification in Scientific Papers
Medical Corpus	Human Labor	Naked Eyes	●●●●●	From Chinese word segmentation to POS tagging to NER
Word Sense Induction	Multiple Models	Multiple Libraries	●●●●●	SemEval2013 Task 13 Word Sense Induction for Graded and Non-Graded Senses
Chinese WS/POS/(NER)	RNN, CRF	TansorFlow	●●●●●	The "from scratch" version of the previous project ("Medical Corpus") (paper)
Toxicity Classification	BiLSTM	Keras	●●●●●	Jigsaw Unintended Bias in Toxicity Classification - Detect toxicity across a diverse range of conversations
CWS/NER	RNN, CRF	TensorFlow	●●●●●	The sequence labeling model on the classic Chinese NLP task

NLP PyTorch

• Basically based on PyTorch and most of the contents are NLP

Subject	Technique	Framework	Complexity	Remark
Machine Translation	RNN, Transformer	PyTorch	●●●●●	Machine translation model from Chinese to English based on WMT17 corpus (use result of CS224n)
Sentence Similarity	RNN	PyTorch	●●●●●	Enhanced-RCNN and other baseline models on some sentence similarity dataset

Other Projects

NCTU DL Course

Subject	Technique	Framework	Complexity	Remark

Deep Learning Categories

TODO: Tasks, Subtasks, Structure, General Architecture, Elements, State-of-the-art model

• General Architecture (DNN, CNN, RNNs, Atteniton, Transformer)
• Categorized by Learning (supervised, ...)
• Categorized by Tasks (NMT, NER, RE, ...)
• Categorized by Structure (Seq2seq, Siamese)
• Categorized by Learning Framework (GAN ?!)
• State-of-the-art models and papers (BERT, ...)

Technique / Network Structure

• Feedforward Neural Network
  • Multilayer Perceptron (MLP)
• Fully Connected Neural Network (FCNN) - And an overview of neural network training process including forward and back propagation
  • Dense Neural Network (DNN)

Image Learning

• Convolusion Neural Network (CNN)

Sequence Learning

Basic Block for Sequence Model!

• Recurrent Neural Network (RNN) - Basis of Sequence model
• Long Short Term Memory (LSTM) - Imporvement of "memory" (brief introduce other regular RNN block)
• Gated Recurrent Units (GRUs)

Reinforcement Learning (RL)

• Q Learning
• Policy Gradient Methods (PG)

Uncategorized

• Generative Adversarial Network (GAN)
• Variational Autoencoder (VAE)
• Self-Organizing Map (SOM)

Learning Framework / Model

Object Detection

• You Only Look Once (YOLO)

Text and Sequence

• Sequence-to-Sequence (seq-to-seq) (Encoder-Decoder) Architecture - Overview of sequence models
  • Bidirectional RNN (BRNN) - RNN-Based seq-to-seq
  • Convolution-based seq-to-seq
  • Attention Model - Transformer-based seq-to-seq
  • Transformer - Attention Is All You Need - Transformer-based multi-headed self-attention
• Word Piece Model (WPM) aka. SentencePiece

Transfer Learning in NLP

"Pre-training in NLP" ≈ "Embedding"

• ELMo
• BERT
• GPT
  • openai/gpt-2
  • OpenAI GPT-2: An Almost Too Good Text Generator - YouTube
• XLNet

Others

• Neural Architecture Search

Ingredient of magic

Layer

• BatchNorm
• Convolution
• Pooling
• Fully Connected (Dense)
• Dropout
• Linear
• LSTM
• RNN

General speaking

• Input
• Hidden
• Output

Activation Function

• Sigmoid
• Hyperbolic Tangent
• Rectified Linear Unit (ReLU)
• Leaky ReLU
• Softmax

Loss Function

• Cross-Entropy
• Hinge
• Huber
• Kullback-Leibler
• MAE (L1)
• MSE (L2)

Optimizer / Optimization Algorithm

• Exponential Moving Average (Exponentially Weighted Moving Average)

• Adadelta

• Adagrad

• Adam

• Conjugate Gradients

• BFGS

• Momentum

• Nesterov Momentum

• Newton’s Method

• RMSProp

• Stochastic Gradient Descent (SGD)

Parameter

• Learning Rate: Used to limit the amount each weight is corrected each time it is updated.
• Epochs: The number of times to run through the training data while updating the weight.

Regularization

• Data Augmentation
• Dropout
• Early Stopping
• Ensembling
• Injecting Noise
• L1 Regularization
• L2 Regularization

Common Concept

Big Pucture: Machine Learning vs. Deep Learning

Terminology / Tricks

• one-hot encoding
• ground truth
• Data Parallelism
• Vanilla - means standard, usual, or unmodified version of something.
  • Vanilla gradient descent (aka. Batch gradient descent) - means the basic gradient descent algorithm without any bells or whistles.

Tricks for language model - a sort of overview

• Word Representation

  • Embedding
  • Train Embedding

• CNN for NLP

• RNN for NLP

• Capsule net with GRU

  • Kaggle kernel - Capsule net with GRU
  • Kaggle kernel - Capsule net with GRU on Preprocessed data

Network Framework

• LeNet - CNN
• AlexNet - CNN
• ZFNet
• VGG-Net - CNN
• GoogleNet - CNN
• ResNet - CNN
• DenseNet
• ResNeXt
• DPN (Dual Path Network)
• CliqueNet

Applications

CV

• Image Classification

NLP

• Basis
  • Text segmentation
  • Part-of-speech tagging (POS tagging)
• Speech Recognition
  • End-to-End Models:
    • (Traditional --> HMM)
    • CTC
    • RNN Transducer
    • Attention-based Model
  • Improved attention
    • Single head attention
    • Multi-headed attention
  • Word Pieces
  • Sequence-Training
    • Beam-Search Decoding Based EMBR
• Named Entity Recognition (NER)
• Neural Machine Translation (NMT)
  • Encoder LSTM + Decoder LSTM
  • Google NMT (GNMT)
• Speech Synthesis
  • WaveNet: A Generative Model for Raw Audio
  • Tacotron: An end-to-end speech synthesis system
• Personalized Recommendation
• Machine Translation
• Sentiment classification
• Chatbot
  • Sequential Matching Network (SMN)

Other Sequence Learning Problem

• Music generation
• DNA sequence analysis
• Video activity recognition

Books Recommendation

• Deep Learning - MIT
  • github pdf
  • chinese translation pdf
  • online reading
• Dive into Deep Learning (D2L Book) (d2l.ai) / 動手學深度學習
  • English github / Chinese github
  • PyTorch / PyTorch (Chinese)
  • Numpy
  • STAT 157 UC Berkeley
• Speech and Language Processing 2ed.
• Deep Learning with Python

Tools

Visualize Drawing Tools

• NN-SVG - FCNN, LeNet, AlexNet style
  • github
• draw.io
• Netscope
• Graphviz - Graph Visualization Software
  • Keras model visualization
  • pydot

Latex

• HarisIqbal88/PlotNeuralNet

Toy

• martisak/dotnets

Resources

Dataset/Corpus

Corpus/NLP Dataset

• SemCor
• SENSEVAL
• SemEval
  • wiki
• WMT17
  • News
  • Chinese-English

Animate Dataset

• nico-opendata
• Danbooru2018 - A Large-Scale Crowdsourced and Tagged Anime Illustration Dataset
• MyAnimeList Dataset
  • MyAnimeList

Github Repository

Example

• hadikazemi/Machine-Learning

Summary

• brightmart/text_classification - all kinds of text classification models and more with deep learning

Application

• Leon
  • leon-ai/leon

Mature Tools

NLP

• Chinese
  • jieba
• English
  • spaCy - Industrial-Strength Natural Language Processing in Python
  • explosion/spaCy
  • gensim
  • nltk
  • fairseq - Facebook AI Research Sequence-to-Sequence Toolkit

Tutorial

Course

• Tensorflow and deep learning without a PhD series by @martin_gorner
• YSDA Natural Language Processing course
• Tensorflow and deep learning without a PhD - Martin Görner
• dataflowr | deep learning courses - github
  • Hands-on tour to deep learning with PyTorch

Interactive Learning

• Kaggle Learn Deep Learning
• Intel AI Developer Program - AI Courses
  • Natural Language Processing
  • Computer Vision
  • Deep Learning

MOOC

• Stanford - CS231n: Convolutional Neural Networks for Visual Recognition
• Stanford - CS244n: Natural Language Processing with Deep Learning
  • Winter 2019 - first time using PyTorch
  • Videos
  • Winter 2017 - using TensorFlow
  • Videos
  • Hank's blog (github)
  • CS224n Chinese camp
• MIT Deep Learning
  • Github - Tutorials, assignments, and competitions for MIT Deep Learning related courses
• PKU - 人工智慧實踐：Tensorflow筆記

Document

• DeepNotes
  • deepnet - Implementations of CNNs, RNNs and cool new techniques in deep learning from scratch
• UFLDL Tutorial
  • Starter Code - github
• Machine Learning Cheatsheet
• 深度學習500問
• Machine Learning Notebook

Github

• Machine Learning cheatsheets for Stanford's CS 229
  • Webpage - CS 229 ― Machine Learning
• Deep Learning cheatsheets for Stanford's CS 230
  • Webpage - CS 230 ― Deep Learning kmario23/deep-learning-drizzle: Drench yourself in Deep Learning, Reinforcement Learning, Machine Learning, Computer Vision, and NLP by learning from these exciting lectures!!
• graykode/nlp-tutorial - Natural Language Processing Tutorial for Deep Learning Researchers
• Microsoft Natural Language Processing Best Practices & Examples
• Microsoft AI education materials for Chinese students, teachers and IT professionals
• Lambda Deep Learning Demos
  • github
• Azure/MachineLearningNotebooks
• smilelight/lightNLP
• RasaHQ/rasa - Open source machine learning framework to automate text- and voice-based conversations
• xinjli/pytensor: A numpy deep learning framework
• boat-group/fancy-nlp: NLP for human. A fast and esay-to-use natural language processing (NLP) toolkit, satisfying your imagination about NLP.
• Hironsan/anago: Bidirectional LSTM-CRF and ELMo for Named-Entity Recognition, Part-of-Speech Tagging and so on.
• BrikerMan/Kashgari: Kashgari is a Production-ready NLP Transfer learning framework for text-labeling and text-classification, includes Word2Vec, BERT, and GPT2 Language Embedding.
• jiesutd/NCRFpp: NCRF++, a Neural Sequence Labeling Toolkit. Easy use to any sequence labeling tasks (e.g. NER, POS, Segmentation). It includes character LSTM/CNN, word LSTM/CNN and softmax/CRF components.
• jiqizhixin/ML-Tutorial-Experiment: Coding the Machine Learning Tutorial for Learning to Learn

Slides

• Supervised Deep Learning

Conclusion

NLP

• crownpku/Awesome-Chinese-NLP: A curated list of resources for Chinese NLP

Summaries

• StateOfTheArt.ai
• Didi Chinese NLP
  • github
• Papers With Code: Browse state-of-the-art

NLP

• graykode/nlp-roadmap: ROADMAP(Mind Map) and KEYWORD for students those who have interest in learning NLP
• Tracking Progress in Natural Language Processing

CV

• Awesome Computer Vision - A curated list of awesome computer vision resources

Article

NLP

• 初入NLP領域的一些小建議

Lexical Database

• WordNet
  • Open Multilingual Wordnet
• BabelNet

Other

Manipulate Github Large File (>100MB)

.gitattributes

• Git large file storage
• Bitbucket tutorial - Git LFS
• Configuring Git Large File Storage
• Moving a file in your repository to Git Large File Storage
  • BFG Repo-Cleaner - brew install bfg
    • github
  • Removing sensitive data from a repository - git filter-branch

Time measure

• Python decorator to measure the execution time of methods

Export Markdown

• Pandoc
  • User Manual
• toDOCX
  • pandoc -o output.docx -f markdown -t docx filename.md
  • PDFtoDOCX
• toPPT
  • Smallpdf PDF to PPT Converter

Machine Learning/Deep Learning Platform

• Anaconda
• PaddlePaddle
  • github

Deprecated notes

• h5py - HDF5 for Python: To store model in HDF5 binary data format
• pyyaml - PyYAML: YAML framework

Programming Framework

Framework	Organization	Support Language	Remark
TensorFlow	Google	Python, C++, Go, JavaScript, ...
Keras	fchollet	Python	on top of TensorFlow, CNTK, or Theano
PyTorch	Facebook	Python
CNTK	Microsoft	C++
OpenNN	C++
Caffe	BVLC	C++, Python
MXNet	DMLC	Python, C++, R, ...
Torch7	Facebook	Lua
Theano	U. Montreal	Python
Deeplearning4J	DeepLearning4J	Java, Scala
Leaf	AutumnAI	Rust
Lasagne	Lasagne	Python
Neon	NervanaSystems	Python

Pending Project

Subject	Technique	Framework	Complexity	Remark
Online ImageNet Classifier	CNN	Keras	○○●●●	(TODO) Using Keras Applications combine with RESTful API
First TF.js	(TODO) Using TensorFlow.js to load pre-trained model and make prediction on the browser
YOLO	CNN	Tensorflow	(TODO) Real-time Object Detection
Word Similarity	(TODO) Word Similarity Based on Dictionary and Based on Corpus