chainer.md

##The DIY Guide to Chainer

Please make Pull Requests for good resources, or create Issues for any feedback! Thanks!

---

Chainer is a deep learning framework (released June 8, 2015) that is not as well known as Tensorflow or Torch. However, it is extremely intuitive and flexible. This is because Chainer adopts a Define-BY-Run model as opposed to a Define-AND-Run model. This paradigm gives you immense control of your network, and makes deep nets easy to debug and experiment with.

Fun fact: Chainer is very popular in the Japanese deep learning community

Installation

pip install chainer

Install from source or with CUDA

---

###Hello World #####Introduction

• Define-BY-Run vs Define-AND-Run - Short intro & Slides (slides 9 - 15) & My explanation
• Basics concepts - Overview (slides 20 - 23)
  • Variable (like Placeholders in TF, Tensors in Torch) - Example & Documentation
  • Links (like Layers in Keras and Torch) - Example & Documentation
  • Neural network - Example & MNIST MLP code & AlexNet code
  • Optimization - Example & Documentation
  • Save / Load - Example
• Recurrent Neural Network
  • Chainer actually let you break down and play around with RNN
  • Basic RNN guide
  • Karpathy's Char-RNN example & RNNLM example
  • LSTM Variants (how to define new LSTMs like StatefulPeepHoleLSTM)
• Computation graph - Documentation, Example visualization
• Model Zoo - Documentation, Code
• GPU support
  • Done through CuPy, as easy as W.to_gpu()
  • Explanation & Slides (slide 29 - 33)
  • EC2 setup guide (5 commands)

#####Instructional Examples

• MNIST - Walkthrough & Code
• CIFAR-10 - Different ConvNet examples
• RNNLM - Char-RNN & Word-RNN
• Word2Vec - Vanilla implmentation
• SVM - Vanilla implmentation

###Advanced Nets #####Autoencoders

• Regular and Stacked Denoising autoencoder
• Convolutional autoencoder
• Variational (alternative) and Adverserial autoencoder
• Fauxtograph blog (Generating fashion from VAE), Code
• Binary net

#####Convolutional NN

• ImageNet, Karpathy blog
• Convolutional autoencoder
• DCGAN, Demo (Anime face generation), Sample
• Deep Dream (Generating Van Gogh paintings), Sample
• R-CNN, Faster R-CNN, Sample
• Siamese network (contrastive loss)
• Filter visualization, Sample
• Illustration2Vec (Converts image to vec, useful in tag generation), Demo

#####RNN

• Seq2Seq dialogue system, machine translation system
• Attention mechanism (from Graves 2013)
• Bidirectional RNN

#####Reinforcement Learning

• DQN, Pong demo
• Another DQN example, Platformer demo

###Super Short Feedback Survey (Pretty please!)