# AI News, Introduction to Chainer&#58; Neural Networks in Python

## Introduction to Chainer&#58; Neural Networks in Python

Neural networks provide a vast array of functionality in the realm of statistical modeling, from data transformation to classification and regression.

Unfortunately, due to the computational complexity and generally large magnitude of data involved, the training of so called deep learning models has been historically relegated to only those with considerable computing resources.

This history is extremely useful when trying to train neural networks because by calling the backward() method on a variable we can perform backpropagation or (reverse-mode) auto-differentiation, which provides our chosen optimizer with all the information needed to successfully update the weights of our neural networks.

Of course the solution to the least squares optimization involved here can be calculated analytically via the normal equations much more efficiently, but this process will demonstrate the basic components of each network you’ll train going forward.

In general, the structure you’ll want to keep common to all neural networks that you make in Chainer involves making a forward function, which takes in your different parametric link functions and runs the data through them all in sequence.

Then writing a train function, which runs that forward pass over your batches of data for a number of full passes through the data called epochs, and then after each forward pass calculates a specified loss/objective function and updates the weights of the network using an optimizer and the gradients calculated through the backward method.

Finally, they’ll tell the optimizer to keep track of and update the parameters of the specified model layers by calling the setup method on the optimizer instance with the layer to be tracked as an argument.

To make it simple here, let’s start by including only 3 link layers in our network (you should feel free to mess around with this at your leisure later though to see what changes make for better classifier performance).

We’ll need a link taking the input images which are $$28 \times 28=784$$ down to some other (probably lower) dimension, then a link stepping down the dimension even further, and finally we want to end up stepping down to the 10 dimensions at the end (with the constraint that they sum to 1).

Additionally, since compositions of linear functions are linear and the benefit of deep learning models are their ability to approximate arbitrary nonlinear functions, it wouldn’t do us much good to stack repeated linear layers together without adding some nonlinear function to send them through&lt;/a&gt;2.

Thus the final forward pass structure will look like: And when it comes time to train our model, with this size of data, we’ll want to batch process a number of samples and aggregate their loss collectively before updating the weights.

We can now train the network (I’d recommend a low number of epochs and a high batch size to start with in order to reduce training time, then these can be altered later to increase validation performance).

Should you wish to save your trained model for future use, Chainer’s most recent release provides the ability to serialize link elements and optimizer states into an hdf5 format via: Then we can quickly restore the state of our previous model and resume training from that point by loading the serialized files.

It’s worth noting that the entire routine we developed for structuring and training a high-performance MNIST classifier with a 3.4% error rate was accomplished in just over 25 lines of code and around 10 seconds of CPU training time, in which it processed a total of 300,000 images.

## Complex neural networks made easy by Chainer

Chainer is an open source framework designed for efficient research into and development of deep learning algorithms.

This gives much greater flexibility in the implementation of complex neural networks, which leads in turn to faster iteration, and greater ability to quickly realize cutting-edge deep learning algorithms.

This mechanism makes backword computation possible by tracking back the entire path from the final loss function to the input, which is memorized through the execution of forward computation—without defining the computational graph in advance.

To train a neural network, three steps are needed: (1) build a computational graph from network definition, (2) input training data and compute the loss function, and (3) update the parameters using an optimizer and repeat until convergence.

Therefore, when implementing recurrent neural networks, for examples, users are forced to exploit special tricks (such as the scan() function in Theano) which make it harder to debug and maintain the code.

The following code shows the implementation of two-layer perceptron in Chainer: In the constructer (__init__), we define two linear transformations from the input to hidden units, and hidden to output units, respectively.

Once forward computation is finished for a minibatch on the MNIST training data set (784 dimensions), the following computational graph can be obtained on-the-fly by backtracking from the final node (the output of the loss function) to the input (note that SoftmaxCrossEntropy is also introduced as the loss function): The point is that the network definition is simply represented in Python rather than a domain-specific language, so users can make changes to the network in each iteration (forward computation).

This imperative declaration of neural networks allows users to use standard Python syntax for branching, without studying any domain specific language (DSL), which can be beneficial as compared to the symbolic approaches that TensorFlow and Theano utilize and also the text DSL that Caffe and CNTK rely on.

On the other hand, although Torch and MXNet also allow users to employ imperative modeling of neural networks, they still use the define-and-run approach for building a computational graph object, so debugging requires special care.

recurrent neural network is a type of neural network that takes sequence as input, so it is frequently used for tasks in natural language processing such as sequence-to-sequence translation and question answering systems.

Since the computational graph of a recurrent neural network contains directed edges between previous and current time steps, its construction and backpropagation are different from those for fixed neural networks, such as convolutional neural networks.

In current practice, such cyclic computational graphs are unfolded into a directed acyclic graph each time for model update by a method called truncated backpropagation through time.

The following example shows a simple recurrent neural network with one recurrent hidden unit: Only the types and size of layers are defined in the constructor as well as on the multi-layer perceptron.

Which means “a” is one of the most probable words, and a noun or adjective tend to follow after “a.” To humans, the results look almost the same, being syntactically wrong and meaningless, even when using different inputs.

Chainer-based deep reinforcement learning library, ChainerRL has been released. https://github.com/pfnet/chainerrl (This

user must provide an appropriate definition of the problem (called “environment”) that is to be solved using reinforcement learning.

The format of defining the environment in ChainerRL follows that of OpenAI’s Gym (https://github.com/openai/gym), a benchmark toolkit for reinforcement learning.

step() sends an action to the environemnt, then returns 4-tuple (next observation, reward, whether it reachs the terminal of episode, and additional information). obs,

DRL, neural networks correspond to policy that determines an action given a state, or value functions (V-function or Q-function), that estimate the value of a state or action.

In ChainerRL, policies and value functions are represented as a Link object in Chainer that implements __call__() method. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 class

ChainerRL is currently a beta version, feedbacks are highly appreciated if you are interested in reinforcement learning.

We are planning to keep improving ChainerRL, by making it easier to use and by adding new algorithms.

(This post is translated from the original post written by Yasuhiro Fujita.) ChainerRL contains a set of Chainer implementations of deep reinforcement learning (DRL) algorithms.

First, user must provide an appropriate definition of the problem (called “environment”) that is to be solved using reinforcement learning.

The format of defining the environment in ChainerRL follows that of OpenAI’s Gym (https://github.com/openai/gym), a benchmark toolkit for reinforcement learning.

In DRL, neural networks correspond to policy that determines an action given a state, or value functions (V-function or Q-function), that estimate the value of a state or action.

In ChainerRL, policies and value functions are represented as a Link object in Chainer that implements __call__() method.

After creating the agent, training can be done either by user’s own training loop, or a pre-defined training function as follows.

As ChainerRL is currently a beta version, feedbacks are highly appreciated if you are interested in reinforcement learning.

Alex Rubinsteyn: Python Libraries for Deep Learning with Sequences

PyData NYC 2015 Recurrent Neural Networks extend the applicability of deep learning into many different problems involving sequential data, such as ...

In this episode of AI Adventures, Yufeng takes us on a tour of TensorBoard, the visualizer built into TensorFlow, to visualize and help debug models. Associated ...

Deep Learning With Python & Tensorflow - PyConSG 2016

Speaker: Ian Lewis Description Python has lots of scientific, data analysis, and machine learning libraries. But there are many problems when starting out on a ...

Chris Moody introduces lda2vec

Chris speaks at data.bythebay.io! Q&A with Chris and Alexy: Standard natural language processing (NLP) is a messy and difficult ..

Keras TensorFlow Model in 7 Minutes

In this tutorial you will learn how to stand up a docker environment to do Keras TensorFlow development. Full Blog Post ...

Understanding Black-box Predictions via Influence Functions

How can we explain the predictions of a black-box model? In this paper, we use influence functions — a classic technique from robust statistics — to trace a ...

Introduction to Cognitive Toolkit | AI8

This session will feature the Cognitive Toolkit, digging into how developers can create their own deep and machine learning models.

Convergence Analysis of Two-layer Neural Networks with ReLU Activation

NIPS 2017 short video.

Hyperbolic Tangent Tanh as a Neural Networks Activation Function

We'll find the derivative of tanh - hyperbolic tangent function This video is a lecture of Neural Network Fundamentals in Python online course on Udemy.

Unsupervised Deep Learning in Python

More courses at the home page: New #Udemy course on Unsupervised #DeepLearning in #Python Early bird 50% off coupon: ..