このレポジトリでは、Deep Learning、MXNet、Gluonを知るためのノートブックを紹介しています。ノートブックを利用することで、文章、画像、数式、そしてコードを1箇所にまとめて説明することを目的としています。ここでの取り組みがうまくいけば、その成果は書籍、教材、チュートリアルの一部、有用かつ流用可能なコードになると思います。私達の知りうる限りでは、(1) 最新の Deep Learning の幅広いコンセプトと (2) 実行可能なコードを容易に利用できるノートブック、これらの双方を提供するものはいままでありませんでした。この試みを行う特別な理由がいまのところ見つかっていないとしても、いずれ、その理由が見つかると考えています。
このコンテンツのもう一つの独特な側面として執筆プロセスがあります。このリソースは完全にオープンかつフリーで利用することができます。通常、書籍というものは、文体を統一したり、コンテンツを絞り込むために、少数の著者によって執筆されます。ここでは、コミュニティからの貢献を歓迎しており、専門家やコミュニティのメンバーと各章をともに執筆することを望んでいます。例えば、誤字・脱字といった訂正については、多くの協力をみなさんから頂いています。
ここでは、基本的なコンセプトから発展的な内容、幅広い応用に至るまで、MXNetを利用して説明をしています。MXNetは、そのスピードに関して非常に有名であり、本番環境で幅広く利用されています。また、MXNetの新しいインタフェースであるGluonによって、MXNetの利用はさらに簡単になります。
これらのノートブックを利用するにあたって、MXNetのインストールが必要になります。幸いにも、インストールは特にLinuxの場合は簡単に行うことができます。こちらのインストールガイドを参照してください。また、Jupyter notebookのインストールも必要で、実行にあたってはPython 3を利用してください。
著者やそれ以外の方も、ここのコンテンツを利用して講演することが増えています。いくつかの講演スライド (例えば、KDD 2017における6時間の講演等) は非常に大きいため、こちらのレポジトリに別途保存しています。もし、ここのコンテンツを利用して作成したチュートリアルや教材などがあれば、ぜひ共有してください。
もともとのコンテンツは英語で、こちらでは日本語に訳していますが、中国語にも翻訳されています (web versionとGitHub source)
-
Chapter 1: Crash course
-
Chapter 2: Introduction to supervised learning
- Linear regression (from scratch)
- Linear regression (with
gluon) - Binary classification with logistic regression (
gluonw bespoke loss function) - Multiclass logistic regression (from scratch)
- Multiclass logistic regression (with
gluon) - Overfitting and regularization (from scratch)
- Overfitting and regularization (with
gluon) - Perceptron and SGD primer
- Learning environments
-
Chapter 3: Deep neural networks (DNNs)
- Multilayer perceptrons (from scratch)
- Multilayer perceptrons (with
gluon) - Dropout regularization (from scratch)
- Dropout regularization (with
gluon) - Introduction to
gluon.Blockandgluon.nn.Sequential() - Writing custom layers with
gluon.Block - Serialization: saving and loading models
- Advanced Data IO
- Debugging your neural networks
-
Chapter 4: Convolutional neural networks (CNNs)
-
Chapter 5: Recurrent neural networks (RNNs)
- Simple RNNs (from scratch)
- LSTMS RNNs (from scratch)
- GRUs (from scratch)
- RNNs (with
gluon) - Roadmap Dropout for recurrent nets
- Roadmap Zoneout regularization
-
Chapter 6: Optimization
- Introduction to optimization
- Gradient descent and stochastic gradient descent from scratch
- Gradient descent and stochastic gradient descent with
gluon - Momentum from scratch
- Momentum with
gluon - Adagrad from scratch
- Adagrad with
gluon - RMSprop from scratch
- RMSprop with
gluon - Adadelta from scratch
- Adadelta with
gluon - Adam from scratch
- Adam with
gluon
-
Chapter 7: Distributed & high-performance learning
- Fast & flexible: combining imperative & symbolic nets with HybridBlocks
- Training with multiple GPUs (from scratch)
- Training with multiple GPUs (with
gluon) - Training with multiple machines
- Roadmap Asynchronous SGD
- Roadmap Elastic SGD
-
Chapter 8: Computer vision (CV)
- Roadmap Network of networks (inception & co)
- Roadmap Residual networks
- Object detection
- Roadmap Fully-convolutional networks
- Roadmap Siamese (conjoined?) networks
- Roadmap Embeddings (pairwise and triplet losses)
- Roadmap Inceptionism / visualizing feature detectors
- Roadmap Style transfer
- Visual-question-answer
- Fine-tuning
-
Chapter 9: Natural language processing (NLP)
- Roadmap Word embeddings (Word2Vec)
- Roadmap Sentence embeddings (SkipThought)
- Roadmap Sentiment analysis
- Roadmap Sequence-to-sequence learning (machine translation)
- Roadmap Sequence transduction with attention (machine translation)
- Roadmap Named entity recognition
- Roadmap Image captioning
- Tree-LSTM for semantic relatedness
-
Chapter 10: Audio processing
- Roadmap Intro to automatic speech recognition
- Roadmap Connectionist temporal classification (CSC) for unaligned sequences
- Roadmap Combining static and sequential data
-
Chapter 11: Recommender systems
- Introduction to recommender systems
- Roadmap Latent factor models
- Roadmap Deep latent factor models
- Roadmap Bilinear models
- Roadmap Learning from implicit feedback
-
Chapter 12: Time series
- Introduction to Forecasting (with
gluon) - Generalized Linear Models/MLP for Forecasting (with
gluon) - Roadmap Factor Models for Forecasting
- Roadmap Recurrent Neural Network for Forecasting
- Linear Dynamical System (from scratch)
- Exponential Smoothing and Innovative State-space modeling (from scratch)
- Roadmap Gaussian processes for Forecasting
- Roadmap Bayesian Time Series Models
- Roadmap Modeling missing data
- Roadmap Combining static and sequential data
- Introduction to Forecasting (with
-
Chapter 13: Unsupervised learning
- Roadmap Introduction to autoencoders
- Roadmap Convolutional autoencoders (introduce upconvolution)
- Roadmap Denoising autoencoders
- Variational autoencoders
- Roadmap Clustering
-
Chapter 14: Generative adversarial networks (GANs)
- Introduction to GANs
- Deep convolutional GANs (DCGANs)
- Roadmap Wasserstein-GANs
- Roadmap Energy-based GANS
- Roadmap Conditional GANs
- Image transduction GANs (Pix2Pix)
- Roadmap Learning from Synthetic and Unsupervised Images
-
Chapter 15: Adversarial learning
- Roadmap Two Sample Tests
- Roadmap Finding adversarial examples
- Roadmap Adversarial training
-
Chapter 16: Tensor Methods
- Introduction to tensor methods
- Roadmap Tensor decomposition
- Roadmap Tensorized neural networks
-
Chapter 17: Deep reinforcement learning (DRL)
- Roadmap Introduction to reinforcement learning
- Roadmap Deep contextual bandits
- Deep Q-networks (DQN)
- Double-DQN
- Roadmap Policy gradient
- Roadmap Actor-critic gradient
-
Chapter 18: Variational methods and uncertainty
- Roadmap Dropout-based uncertainty estimation (BALD)
- Weight uncertainty (Bayes by Backprop) from scratch
- Weight uncertainty (Bayes by Backprop) with
gluon - Weight uncertainty (Bayes by Backprop) for Recurrent Neural Networks
- Roadmap Variational autoencoders
-
Chapter 19: Graph Neural Networks
- Appendix 1: Cheatsheets
- Roadmap
gluon - Roadmap PyTorch to MXNet (work in progress)
- Roadmap Tensorflow to MXNet
- Roadmap Keras to MXNet
- Roadmap Math to MXNet
- Roadmap
We've designed these tutorials so that you can traverse the curriculum in more than one way.
- Anarchist - Choose whatever you want to read, whenever you want to read it.
- Imperialist - Proceed through all tutorials in order. In this fashion you will be exposed to each model first from scratch, writing all the code ourselves but for the basic linear algebra primitives and automatic differentiation.
- Capitalist - If you don't care how things work (or already know) and just want to see working code in
gluon, you can skip (from scratch!) tutorials and go straight to the production-like code using the high-levelgluonfront end.
This evolving creature is a collaborative effort (see contributors tab). The lead writers, assimilators, and coders include:
- Zachary C. Lipton (@zackchase)
- Mu Li (@mli)
- Alex Smola (@smolix)
- Sheng Zha (@szha)
- Aston Zhang (@astonzhang)
- Joshua Z. Zhang (@zhreshold)
- Eric Junyuan Xie (@piiswrong)
- Kamyar Azizzadenesheli (@kazizzad)
- Jean Kossaifi (@JeanKossaifi)
- Stephan Rabanser (@steverab)
In creating these tutorials, we've have drawn inspiration from some the resources that allowed us to learn deep / machine learning with other libraries in the past. These include:
- Soumith Chintala's Deep Learning with PyTorch: A 60 Minute Blitz
- Alec Radford's Bare-bones intro to Theano
- Video of Alec's intro to deep learning with Theano
- Chris Bishop's Pattern Recognition and Machine Learning
- Already, in the short time this project has been off the ground, we've gotten some helpful PRs from the community with pedagogical suggestions, typo corrections, and other useful fixes. If you're inclined, please contribute!