Last update: 28 July, 2017

Chainer – a deep learning frameworkChainer provides a set of features required for research and development using deep learning such as designing neural networks, training, and evaluation.

Designing a network Training, evaluation

Dataset

Features and Characteristics of Chainer

Powerful

☑ CUDA

☑ cuDNN

☑ NCCL

Versatile

☑ Convolutional Network

☑ Recurrent Network

☑ Many Other Components

☑ Various Optimizers

Intuitive

☑ Define-by-Run

☑ High debuggability

Supports GPU calculation using CUDA

High-speed training/inference by cuDNN

Supports a fast, multi-GPU learning using NCCL

N-dimensional Convolution, Deconvolution, Pooling, BN, etc.

RNN components such as LSTM, Bi-directional LSTM, GRU and Bi-directional GRU

Many layer definitions and various loss functions used in neural networks

Various optimizers, e.g., SGD, MomentumSGD, AdaGrad, RMSProp, Adam, etc.

Easy to write a complicated network

User-friendly error messages. Easy to debug using pure Python debuggers.

Well-abstracted common tools for various NN learning, easy to write a set of learning flows☑ Simple APIs

Popularity Growth of Chainer

Neural network = Computational graph

NN can be interpreted as a computational graph that applies many linear and nonlinear functions to input vectors

How to handle a computational graph

A definition of computational graph exists apart from code that performs computation according to the definition

StaticThe actual code that performs computation is treated as a definition of computational graph

Dynamic

Chainer is the first deep-learning framework to adopt “Define-by-Run”*

How about Chainer? → Dynamic

● Define-and-Run（static graph）Consists of two steps: first to build a computational graph, then feed data to the computational graph (Caffe, theano, TensorFlow, etc.)

● Define-by-Run（dynamic graph）Describing a forward-pass computation means to construct a computational graph for the backward computation (Chainer, DyNet, PyTorch, etc.）

* autograd adopted Define-by-Run but it was not a framework for deep learning.

Define-and-Run and Define-by-Run

# Building

x = Variable(‘x’)

y = Variable(‘y’)

z = x + 2 * y

# Evaluation

for xi, yi in data:

eval(z, (xi, yi))

# Build, evaluate at the same time

for xi, yi in data:

x = Variable(xi)

y = Variable(yi)

z = x + 2 * y

You can make a branch to change the forward computation depending on the data

Define-and-Run Define-by-Run

How to write a Convolutional Networkimport chainerimport chainer.links as Limport chainer.functions as F

class LeNet5(chainer.Chain): def __init__(self): super(LeNet5, self).__init__() with self.init_scope(): self.conv1 = L.Convolution2D(1, 6, 5, 1) self.conv2 = L.Convolution2D(6, 16, 5, 1) self.conv3 = L.Convolution2D(16, 120, 4, 1) self.fc4 = L.Linear(None, 84) self.fc5 = L.Linear(84, 10)

• Start writing a model by inheriting Chain class• Register parametric layers inside the init_scope

• Write forward computation in __call__ method (no need to write backward computation)

def __call__(self, x): h = F.sigmoid(self.conv1(x)) h = F.max_pooling_2d(h, 2, 2) h = F.sigmoid(self.conv2(h)) h = F.max_pooling_2d(h, 2, 2) h = F.sigmoid(self.conv3(h)) h = F.sigmoid(self.fc4(h)) return self.fc5(h)

Training modelsmodel = LeNet5()model = L.Classifier(model)

# Dataset is a list! ([] to access, having __len__)dataset = [(x1, t1), (x2, t2), ...]

# iterator to return a mini-batch retrieved from datasetit = iterators.SerialIterator(dataset, batchsize=32)

# Optimization methods (you can easily try various methods by changing SGD to # MomentumSGD, Adam, RMSprop, AdaGrad, etc.)opt = optimizers.SGD(lr=0.01)opt.setup(model)

updater = training.StandardUpdater(it, opt, device=0) # device=-1 if you use CPUtrainer = training.Trainer(updater, stop_trigger=(100, 'epoch'))trainer.run()

For more details, refer to official examples: https://github.com/pfnet/chainer/tree/master/examples

Define-by-Run brings flexibility and intuitiveness

“Forward computation” becomes a definition of network• Depending on data, it is easy to change a network structure

• You can define a network itself by Python code＝The network structure can be treated as a program instead of data.

For Chainer, the “forward computation” can be written in Python• Enables you to write a network structure freely using the syntax of Python

• Define-by-Run makes it easy to insert any process like putting a print statement between network computations (In case of define-and-run which compiles a network, this kind of debugging is difficult)

• Easy to reuse code of the same network for other purposes with few changes (e.g. by just adding a conditional branch partially)

• Easy to check intermediate values and the design of the network itself using external debugging tools etc.

Chainer v2.0.1Significantly reduced memory consumption, organized API in response to the users feedback

Aggressive Buffer Release to reduce the memory consumption during training→

CuPy has been released as an independent library. This allows for array operations using GPU via an interface highly compatible with NumPy.

https://cupy.chainer.org

https://chainer.org

CuPy Independent library to handle all GPU calculations in Chainer

Lower cost to migrate CPU code to GPU with NumPy-compatible API

GPU-execute linear algebra algorithms such as a singular value decomposition

Rich in examples such as KMeans, Gaussian Mixture Model

import numpy as np

x = np.random.rand(10)

W = np.random.rand(10, 5)

y = np.dot(x, W)

import cupy as cp

x = cp.random.rand(10)

W = cp.random.rand(10, 5)

y = cp.dot(x, W)

GPU

https://github.com/cupy/cupy

Add-on packages for ChainerDistribute deep learning, deep reinforcement learning, computer vision

ChainerMN (Multi-Node): additional package for distributed deep learning　　High scalability (100 times faster with 128GPU)

ChainerRL: deep reinforcement learning library　　DQN, DDPG, A3C, ACER, NSQ, PCL, etc. OpenAI Gym support

ChainerCV: provides image recognition algorithms, dataset wrappers　　Faster R-CNN, Single Shot Multibox Detector (SSD), SegNet, etc.

ChainerMNChainer + Multi-Node

ChainerMN: Multi-nodeKeeping the easy-to-use characteristics of Chainer as is, ChainerMN enables to use multiple nodes which have multiple GPUs easily to make training faster

GPU

GPU

InfiniBand

GPU

GPU

InfiniBand

MPI

NVIDIA NCCL

Destributed deep learning with ChainerMN100x speed up with 128 GPUs

Comparison with other frameworksChainerMN is the fastest at the comparison of elapsed time to train ResNet-50 on ImageNet dataset for 100 epochs (May 2017)

We confirmed that if we increase the number of nodes, the almost same accuracy can be achieved

Speedup without dropping the accuracy

Scale-out test on Microsoft Azure

Easy-to-use API of ChainerMN

You can start using ChainerMN just by wrapping one line!

optimizer = chainer.optimizers.MomentumSGD()

optimizer = chainermn.DistributedOptimizer( chainer.optimizers.MomentumSGD())

ARM template will be announced soon

https://github.com/mitmul/ARMTeamplate4ChainerMN

↑ Click this to make a master node ↑ Click this to make worker nodes

Scaling via web interfaceYou can launch a scale-set of Azure instances super easily!

ChainerRLChainer + Reinforcement Learning

Reinforcement Learning:

ChainerRL: Deep Reinforcement Learning Library

Train an agent which interacts with the environment to maximize the rewards

Action

Env

Observation, Reward

Reinforcement Learning with ChainerRL1. Create an environment

Action

Env

Observation, Reward

Distribution: Softmax, Mellowmax, Gaussian,…

Policy: Observation → Distribution of actions

2. Define an agent model

Reinforcement Learning with ChainerRL

2. Define an agent model (contd.)Q-Function: Observation → Value of each action (expectation of the sum of future rewards)

ActionValue: Discrete, Quadratic

Reinforcement Learning with ChainerRL

Action

Env

Observation, Reward

3. Create an agent

Reinforcement Learning with ChainerRL

4. Interact with the environment!

Reinforcement Learning with ChainerRL

Algorithms provided by ChainerRL• Deep Q-Network (Mnih et al., 2015)• Double DQN (Hasselt et al., 2016)• Normalized Advantage Function (Gu et al., 2016)• (Persistent) Advantage Learning (Bellemare et al., 2016)• Deep Deterministic Policy Gradient (Lillicrap et al., 2016)• SVG(0) (Heese et al., 2015)• Asynchronous Advantage Actor-Critic (Mnih et al., 2016)• Asynchronous N-step Q-learning (Mnih et al., 2016)• Actor-Critic with Experience Replay (Wang et al., 2017) <- NEW!• Path Consistency Learning (Nachum et al., 2017) <- NEW!• etc.

ChainerRL Quickstart Guide• Define a Q-function in a Jupyter notebook and learn the Cart Pole

Balancing problem with DQNhttps://github.com/pfnet/chainerrl/blob/master/examples/quickstart/quickstart.ipynb

ChainerCVChainer + Computer Vision

Evaluate your model on popular datasets

Running and training deep-learning models easier for Computer Vision tasks

ChainerCV https://github.com/pfnet/chainercv

Datasets

Pascal VOC, Caltech-UCSD

Birds-200-2011, Stanford Online

Products, CamVid, etc.

Models

Faster R-CNN, SSD, SegNet (will add more

models!)

Trainingtools

Evaluationtools

DatasetAbstraction

Train popular models with your data

Start computer vision research using deep learning much easier

ChainerCV

Latest algorithms with your dataProvide complete model code, training code, inference code for segmentation algorithms (SegNet, etc.) and object detection algorithms (Faster R-CNN, SSD, etc.), and so on

All code is confirmed to reproduce the resultsAll training code and model code reproduced the experimental results shown in the original paper

https://github.com/pfnet/chainercv

• If you want to see some examples of ChainerCV and the reproducing code for some papers, please check the official Github repository (chainer/chainercv)

• The right figure shows the result of the inference code of Faster RCNN example

• The pre-trained weights are automatically downloaded!

https://github.com/pfnet/chainercv

$ pip install chainercv

•

•

→

←

•••

Intel Chainer

Intel Chainer with MKL-DNN Backend

CPU

CuPy

NVIDIA GPU

CUDA

cuDNN

BLAS

NumPy

Chainer

MKL-DNN

Intel Xeon/Xeon Phi

MKL

Intel Chainer with MKL-DNN BackendMKL-DNN• Neural Network library optimized for Intel architectures• Supported CPUs:✓ Intel Atom(R) processor with Intel(R) SSE4.1 support

✓ 4th, 5th, 6th and 7th generation Intel(R) Core processor

✓ Intel(R) Xeon(R) processor E5 v3 family (code named Haswell)

✓ Intel(R) Xeon(R) processor E5 v4 family (code named Broadwell)

✓ Intel(R) Xeon(R) Platinum processor family (code name Skylake)

✓ Intel(R) Xeon Phi(TM) product family x200 (code named Knights Landing)

✓ Future Intel(R) Xeon Phi(TM) processor (code named Knights Mill)

• MKL-DNN accelerates the computation of NN on the above CPUs

Intel Chainer with MKL-DNN Backendconvnet-benchmarks* result:

Intel Chainer Chainer with NumPy (MKL-Build)

Alexnet Forward 429.16 ms 5041.91 ms

Alexnet Backward 841.73 ms 5569.49 ms

Alexnet Total 1270.89 ms 10611.40 ms

~8.35x faster than NumPy backend！

Intel Chainer with MKL-DNN BackendIntel is developing Intel Chainer as a fork of Chainer v2

https://github.com/intel/chainer

Applications using Chainer

Object Detection

https://www.youtube.com/watch?v=yNc5N1MOOt4

Semantic Segmentation

https://www.youtube.com/watch?v=lGOjchGdVQs

Ponanza Chainer● Won the 2nd place at The 27th World Computer Shogi Championship

● Based on Ponanza which was the champion for two years in a row (2015, 2016)

● “Ponanza Chainer” applied Deep Learning for ordering the possible next moves for which “Ponanza” should think ahead deeply

● “Ponanza Chainer” wins “Ponanza” with a probability of 80%

47

Team PFN

Issei Yamamoto

Akira Shimoyama

Team Ponanza

Paints Chainer● Auto Sketch Colorization

● Train a neural network with a large dataset of paintings

● It takes a line drawings as input, and output a colorized image!

● You can also give color hits which indicates preferable colors

48

https://paintschainer.preferred.tech

Installation of Chainer

1. Install CUDA Toolkit 8.0https://developer.nvidia.com/cuda-downloads

2. Install cuDNN v6.0 Libraryhttps://developer.nvidia.com/rdp/cudnn-download

3. Install NCCL for Multi-GPUshttps://github.com/NVIDIA/nccl

4. Install CuPy and Chainer% pip install cupy% pip install chainer

Chainer on Ubuntu

For more details, see the official installation guide: http://docs.chainer.org/en/stable/install.html

Chainer on Windows with NVIDIA GPU1. Install Visual C++ 2015 Build Tools

http://landinghub.visualstudio.com/visual-cpp-build-tools

2. Install CUDA Toolkit 8.0https://developer.nvidia.com/cuda-downloads

3. Install cuDNN v6.0 Library for Windows 10https://developer.nvidia.com/rdp/cudnn-downloadPut all files under C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0

4. Install Anaconda 4.3.1 Python 3.6 or 2.7https://www.continuum.io/downloads

5. Add environmental variables- Add “C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\bin” to PATH variable- Add “C:\Program Files (x86)\Windows Kits\10\Include\10.0.10240.0\ucrt” to INCLUDE variable

6. Install Chainer on Anaconda Prompt> pip install chainer

Chainer on AzureUse Data Science Virtual Machine for Linux (Ubuntu)• Ready for CUDA 8.0 & cuDNN 5.1• After ssh, ”pip install --user chainer”

1

2

3

Chainer Model Exporttfchain: TensorFlow export (experimental)

Caffe-export: Caffe export (experimental)

• https://github.com/mitmul/tfchain• Supports Linear, Convolution2D, MaxPooling2D, ReLU• Just add @totf decorator right before the forward method of the model

• Currently closed project• Supports Conv2D, Deconv2D, BatchNorm, ReLU, Concat, Softmax,

Reshape

External Projects for Model Portability

DLPack

• https://mil-tokyo.github.io/webdnn/• The model conversion to run it on a web browser supports Chainer

WebDNN

• https://github.com/dmlc/dlpack

• MXNet, Torch, Caffe2 have joined to discuss the guideline of memory layout of tensor and the common operator interfaces

Companies supporting Chainer

Companies supporting Chainer

Contributing to Chainer

Chainer is an open-source project.• You can send a PR from here: https://github.com/chainer/chainer• The development speed of Deep Learning research is super fast, therefore,

to provide the state-of-the-art technologies through Chainer, we continuously update the development plans:

• Chainer v3.0.0 will be released on 26th September!• Will support gradient of gradient (higher order differentiation)• Will add the official Windows support ensured by Microsoft

The release schedule after v2.0.1 (4th July)→