KTUNE

KTUNE is machine learning wrapper for ManyCore CPU which motivated by Keras

How to Optimize

OpenMP

We use OpenMP for using Multithreading.

MKL library

We also use Math Kernel Library given by Intel

Find an optimized number of threads.

We developed an algorithm for find optimized number of threads.

Compile

$ mv ./src
$ export LD_PRELOAD=libmkl_core.so:libmkl_sequential.so
$ make
$ mv Ktune.so ./../

How to use?

$ python3
>>>> import Ktune
>>>> t = Ktune.Sequential()
>>>> t.help()
  
You can use .help() to refer to it.
  
here is the function of Ktune
  
1. network
This is a function that determines what kind of network to use
 .networkd(kind of network)
example : .network(fully connected)
  
2. add
This is a function that add layer
 .add(number of layer , 1st layer number(input layer) , 2nd layer number ...., output layer number)
example : .add(4, 784 , 120 , 40 , 10)
  
3. action
This is a function that setting action function
 .action(action function name)
example : .action(sigmoid)
  
4. loss
This is a function that setting loss function
 .loss(loss function name)
example : .loss(mean_squared_error)
  
5. optimizer
This is a function that setting optimizer function
 .optimizer(optimizer function name)
example : .optimizerz(sgd)
  
6. data
This is a function that setting determines what kind of data to use
 .data(data name)
example : .data(mnist)
  
7. fit
This is a function that training data
 .fit(Batch_size,Epoch,Learning_rate)
example : .fit(5,100,0.8)

Requirement

Intel C++ Compile ( icc )

MKL library

Our Execution environment is intel¢ç Xeon Phi¢â Processor 7210 (16GB, 1.30 GHz, 64 core)

Paper

We published Paper whith Ktune you can find our Paper in this site (korean institute of information scientists and engineers) http://www.dbpia.co.kr/Article/NODE07322653

ksy9164 / Ktune