TensorFlow에 대한 분석 내용 - TensorFlow? - 배경 - DistBelief - Tutorial - Logistic regression - TensorFlow - 내부적으로는 - Tutorial - CNN, RNN - Benchmarks - 다른 오픈 소스들 - TensorFlow를 고려한다면 - 설치 - 참고 자료

1. TensorFlow
JakeS.Choi(shchoi@diotek.com)
2015.12.17

2. 차례
TensorFlow?
배경
DistBelief
Tutorial-Logisticregression
TensorFlow-내부적으로는
Tutorial-CNN,RNN
Benchmarks
다른오픈소스들
TensorFlow를고려한다면
설치
참고자료

3. TensorFlow?
Opensourcesoftwarelibraryfornumericalcomputation
usingdataflowgraphs
Deeplearningframework이아니네?
GoogleResearchBlog에2015년11월9일발표
Apache2.0라이선스

4. Tensor-Flow?
[1]Large-ScaleDeepLearningforIntelligentComputerSystems

5. 배경
2012년기준으로구글내부적으로machinelearning,deep
learning에대한수요가증가함
[1]Large-ScaleDeepLearningforIntelligentComputerSystems
GMail,Photos,Speech등어러분야에적용

6. 배경
2011년GoogleBrainproject의일환으로대용량데이터와대규모계
산을중점으로개발을시작
[1]Large-ScaleDeepLearningforIntelligentComputerSystems

7. DistBelief
2011년에개발된Google의Deeplearninginfrastructure
기본설계방향
처리할데이터가많다→데이터병렬화
모델이너무커서메모리에한번에안올라감→모델병렬화
구글은학습에천단위코어수를사용함
[2]Introductiontoapachehorn(incubating)

8. DistBelief
Modelparallelismbenchmarks
실험결과
HWR은일반적으로40K,최대10M까지테스트
Speech의경우8대에서2.2배의성능향상,8대이상은데이터교환
비용으로성능하락이확인됨
[4]LargeScaleDistributedDeepNetworks

9. DistBelief
Dataparallelismbenchmarks
실험결과
[4]LargeScaleDistributedDeepNetworks

10. TensorFlow
기존DistBelief를개선
DistBelief은규모에대한확장성은뛰어나지만유연성이떨어졌음
TensorFlow는DistBelief보다는2배정도빠름다고함
아직whitepaper에benchmarking자료는미포함됨[3]
구글내부적으로는DistBelief에서TensorFlow로이전을완료한상태

11. Tutorial-Logisticregression
0~9숫자하나를인식하는문제,문제의해결을위한모델은logistic
regression을사용
MNIST파일을로딩,input_data는구글에서미리구현해둔코드
#ImportMINSTdata
importinput_data
mnist=input_data.read_data_sets(MNIST_data/,one_hot=True)
데이터는아래와같은형태의이미지
[6]MNISTForMLBeginners

12. Tutorial-Logisticregression
mnist.train,mnist.validation,mnist.test로구성
mnist.*.xs는아래와같은형태의28x28x1(Grayscale)이미지
[6]MNISTForMLBeginners

13. Tutorial-Logisticregression
총60,000의이미지로구성
[6]MNISTForMLBeginners

14. Tutorial-Logisticregression
mnist.*.ys는0~9까지하나만1을가지는one-hotvector로목표/정
답데이터
[6]MNISTForMLBeginners

15. Tutorial-Logisticregression
importtensorflowastf
#Parameters
learning_rate=0.01
training_epochs=25
batch_size=100
display_step=1
#tfGraphInput
x=tf.placeholder(float,[None,784])#mnistdataimageofshape28*28=784
y=tf.placeholder(float,[None,10])#0-9digitsrecognition=10classes
#Createmodel
#Setmodelweights
W=tf.Variable(tf.zeros([784,10]))
b=tf.Variable(tf.zeros([10]))
#Constructmodel
activation=tf.nn.softmax(tf.matmul(x,W)+b)#Softmax
#Minimizeerrorusingcrossentropy
cost=-tf.reduce_sum(y*tf.log(activation))#Crossentropy
#GradientDescent
optimizer=tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)

16. Tutorial-Logisticregression
Nonex784(28x28)으로x라는matrix를정의,placeholder로정의된
내용은그래프계산전전달되어야함,이값은외부에서그래프로입력을
다룰때사용y는정답데이터가입력됨,여기서None으로정의된값은
mini-batchsize가됨
x=tf.placeholder(float,[None,784])
y=tf.placeholder(float,[None,10])#0-9digitsrecognition=10classes
W는inputxoutput크기의matrix,b는bias로output크기와동일한
변수를정의
W=tf.Variable(tf.zeros([784,10]))
b=tf.Variable(tf.zeros([10]))

17. Tutorial-Logisticregression
logisticregression모델을정의
tf.matmul은matrixmultiply함수
activation=tf.nn.softmax(tf.matmul(x,W)+b)#Softmax

18. Tutorial-Logisticregression
crossentropy함수
cost함수를위의수식과같이crossentropy함수와동일하게정의
optimizier는아래와같이gradientdecent를사용하고cost가최소화
되도록최적화를실행
cost=-tf.reduce_sum(y*tf.log(activation))#Crossentropy
#GradientDescent
optimizer=tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)

19. Tutorial-Logisticregression
#Initializingthevariables
init=tf.initialize_all_variables()
#Launchthegraph
withtf.Session()assess:
sess.run(init)
#Trainingcycle
forepochinrange(training_epochs):
avg_cost=0.
total_batch=int(mnist.train.num_examples/batch_size)
#Loopoverallbatches
foriinrange(total_batch):
batch_xs,batch_ys=mnist.train.next_batch(batch_size)
#Fittrainingusingbatchdata
sess.run(optimizer,feed_dict={x:batch_xs,y:batch_ys})
#Computeaverageloss
avg_cost+=sess.run(cost,feed_dict={x:batch_xs,y:batch_ys})/total_batch
#Displaylogsperepochstep
ifepoch%display_step==0:
printEpoch:,'%04d'%(epoch+1),cost=,{:.9f}.format(avg_cost)

20. Tutorial-Logisticregression
#Initializingthevariables
init=tf.initialize_all_variables()
#Launchthegraph
withtf.Session()assess:
...
printOptimizationFinished!
#Testmodel
correct_prediction=tf.equal(tf.argmax(activation,1),tf.argmax(y,1))
#Calculateaccuracy
accuracy=tf.reduce_mean(tf.cast(correct_prediction,float))
printAccuracy:,accuracy.eval({x:mnist.test.images,y:mnist.test.labels})

21. Tutorial-Logisticregression
코드를실행하면아래와같이dataflowgraph를생성함
[1]Large-ScaleDeepLearningforIntelligentComputerSystems

22. Tutorial-Logisticregression
이렇게만들어진그래프는분산처리가가능함
[1]Large-ScaleDeepLearningforIntelligentComputerSystems

23. Tutorial-Logisticregression
코드상에는forwardpass만있고backwardpass에대한내용이없
음(Optimizer정의를제외하고)
TensorFlow에서는x,y,W,b등을각각의심볼로다루고심볼계산
(Symboliccomputation)을지원함
내부적으로자동적으로미분함수를구하고gradient를계산함

24. Tutorial-Logisticregression
GraphVisualization
아래와같이summary에관한코드를추가
#Initializingthevariables
init=tf.initialize_all_variables()
tf.scalar_summary(loss,cost)
merged_summary_op=tf.merge_all_summaries()
#Launchthegraph
withtf.Session()assess:
sess.run(init)
summary_writer=tf.train.SummaryWriter('/tmp/tf_l_regression.log',graph_def=
sess.graph_def)
...

25. Tutorial-Logisticregression
GraphVisualization
...
#Trainingcycle
forepochinrange(training_epochs):
avg_cost=0.
total_batch=int(mnist.train.num_examples/batch_size)
#Loopoverallbatches
foriinrange(total_batch):
batch_xs,batch_ys=mnist.train.next_batch(batch_size)
#Fittrainingusingbatchdata
sess.run(optimizer,feed_dict={x:batch_xs,y:batch_ys})
#Computeaverageloss
avg_cost+=sess.run(cost,feed_dict={x:batch_xs,y:batch_ys})/total_batch
summary_str=sess.run(merged_summary_op,feed_dict={x:batch_xs,y:bat
ch_ys})
summary_writer.add_summary(summary_str,epoch*total_batch+i)
...

26. Tutorial-Logisticregression
GraphVisualization
Tensorflow가동작중로그를지정한경로에남김
그런후아래와같이실행하면http://localhost:6006에접속이가능함
tensorboard--logdir=/tmp/tf_l_regression.log

27. Tutorial-Logisticregression
GraphVisualization
아래와같이남겨진scalar값은그림과같이값의변화를확인할수있음
tf.scalar_summary(loss,cost)

28. Tutorial-Logisticregression
GraphVisualization

29. TensorFlow-돌아와서
DNN뿐아니라다른machinelearing등대규모matrix연산이필요
한곳에적용이가능함
하드웨어에대한이해가필요없음
Deeplearning을위한다양한최적화된함수를지원
자동미분을계산하여backwardpass에대한코드작성이불필요함
다양한모델과결합이쉬움
C++,Pythonfrontend를지원
[1]Large-ScaleDeepLearningforIntelligentComputerSystems

30. TensorFlow-내부적으로는
분산,이종/다중디바이스를위한고려,현재는singlemachine코드만
공개된상태
구글에서도해당기능에대해서중요하게생각하고있으며향후공개가
능성이높을것으로판단됨[7]
Multi-device(GPU/CPU)에대한동작은지원하고있음[8]
[3]TensorFlow:Large-scalemachinelearningonheterogeneoussystems

31. TensorFlow-내부적으로는
Scheduling과리소스(CPU,GPU,메모리등)별node배치가핵심일
것으로판단됨
노드배치는costmodel을사용함
Cost는input,output의크기와operation으로통계적으로추정
혹은이미실행된배치에서측정된비용으로계산됨
이비용에는계산,데이터교환비용등이포함됨
사용자에의해특정작업을특정디바이스에할당할수도있음
디바이스간간선은동적으로send/receivenode가추가되어디바이
스로부터추상화시킴
[3]TensorFlow:Large-scalemachinelearningonheterogeneoussystems

32. TensorFlow-내부적으로는
특별히Gradient기반의그래프계산에는몇가지힌트를사용할수있
음
forwardpass를통해계산된결과는backwardpass에서다시사
용됨
따라서노드할당에이를고려하여할당
단,메모리한계가있기때문에long-livedtensor에대한
swapping도같이고려됨
[3]TensorFlow:Large-scalemachinelearningonheterogeneoussystems

33. TensorFlow-내부적으로는
커널구현은cuBLAS,cuDNN,Eigen등을사용
데이터병렬화에대한내용에대한고려가되어있음
그래프생성후디바이스별노드배치는모델병렬화의내용
아직이부분도API상에서확인되지않음
[3]TensorFlow:Large-scalemachinelearningonheterogeneoussystems

34. Tutorial-CNN
importtensorflowastf
importinput_data
defweight_variable(shape):
#mean=0.0
initial=tf.truncated_normal(shape,stddev=0.1)
returntf.Variable(initial)
defbias_variable(shape):
initial=tf.constant(0.1,shape=shape)
returntf.Variable(initial)
defconv2d(x,W):
returntf.nn.conv2d(x,W,strides=[1,1,1,1],padding='SAME')
defmax_pool_2x2(x):
returntf.nn.max_pool(x,ksize=[1,2,2,1],strides=[1,2,2,1],padding='SAME')

35. Tutorial-CNN
mnist=input_data.read_data_sets('MNIST_data/',one_hot=True)
x=tf.placeholder('float',[None,784])
#-1==batchsize
x_image=tf.reshape(x,[-1,28,28,1])
#5x5patchsizeof1channel,32features
W_conv1=weight_variable([5,5,1,32])
b_conv1=bias_variable([32])
h_conv1=tf.nn.relu(conv2d(x_image,W_conv1)+b_conv1)
h_pool1=max_pool_2x2(h_conv1)
#5x5patchsizeof32channel,64features
W_conv2=weight_variable([5,5,32,64])
b_conv2=bias_variable([64])
h_conv2=tf.nn.relu(conv2d(h_pool1,W_conv2)+b_conv2)
h_pool2=max_pool_2x2(h_conv2)

36. Tutorial-CNN
#imagesizereducedto7x7,fullconnected
W_fc1=weight_variable([7*7*64,1024])
b_fc1=bias_variable([1024])
h_pool2_flat=tf.reshape(h_pool2,[-1,7*7*64])
h_fc1=tf.nn.relu(tf.matmul(h_pool2_flat,W_fc1)+b_fc1)
keep_prob=tf.placeholder('float')
h_fc1_drop=tf.nn.dropout(h_fc1,keep_prob)
W_fc2=weight_variable([1024,10])
b_fc2=bias_variable([10])
y_conv=tf.nn.softmax(tf.matmul(h_fc1_drop,W_fc2)+b_fc2)
y_=tf.placeholder('float',[None,10])
cross_entropy=-tf.reduce_sum(y_*tf.log(y_conv))

37. Tutorial-CNN
train_step=tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction=tf.equal(tf.argmax(y_conv,1),tf.argmax(y_,1))
accuracy=tf.reduce_mean(tf.cast(correct_prediction,'float'))

38. Tutorial-RNN
JürgenSchmidhuber[10]가제안한문제
timestep 0 1 2 3 4 ... 20 21 ... 53 54
x[0] 0.5 0.7 0.3 0.1 0.2 ... 0.5 0.9 ... 0.8 0.2
x[1] 0 0 1 0 0 0 1 0 0
이입력은0.3+.9=1.2가정답으로출력되어야함
[9]rnn_example.ipynb

39. Tutorial-RNN
아래코드는이문제를생성하는코드
defgen_data(min_length=51,max_length=55,n_batch=5):
X=np.concatenate([np.random.uniform(size=(n_batch,max_length,1)),
np.zeros((n_batch,max_length,1))],
axis=-1)
y=np.zeros((n_batch,))
#Computemasksandcorrectvalues
forninrange(n_batch):
#Randomlychoosethesequencelength
length=np.random.randint(min_length,max_length)
#ichangedthistoaconstant
#length=55
#ZerooutXaftertheendofthesequence
X[n,length:,0]=0
#Settheseconddimensionto1attheindicestoadd
X[n,np.random.randint(length/2-1),1]=1
X[n,np.random.randint(length/2,length),1]=1
#MultiplyandsumthedimensionsofXtogetthetargetvalue
y[n]=np.sum(X[n,:,0]*X[n,:,1])
return(X,y)

40. Tutorial-RNN
importtensorflowastf
importnumpyasnp
fromtensorflow.models.rnnimportrnn_cell
fromtensorflow.models.rnnimportrnn
#Definingsomehyper-params
num_units=2#thisistheparameterforinput_sizeinthebasicLSTMcell
input_size=2#num_unitsandinput_sizewillbethesame
batch_size=50
seq_len=55
num_epochs=100

41. Tutorial-RNN
###ModelConstruction
#weusethebasicLSTMcellprovidedinTensorFlow
cell=rnn_cell.BasicLSTMCell(num_units)
#numunitsistheinput-sizeforthiscell
#createplaceholdersforXandy
inputs=[tf.placeholder(tf.float32,shape=[batch_size,input_size])for_inrange(se
q_len)]
result=tf.placeholder(tf.float32,shape=[batch_size])
#notethatoutputsisalistofseq_len
outputs,states=rnn.rnn(cell,inputs,dtype=tf.float32)
#eachelementisatensorofsize[batch_size,num_units]
#weactuallyonlyneedthelastoutputfromthemodel,ie:lastelementofoutputs

42. Tutorial-RNN
#Weactuallywanttheoutputtobesize[batch_size,1]
#Sowewillimplementalinearlayertodothis
W_o=tf.Variable(tf.random_normal([2,1],stddev=0.01))
b_o=tf.Variable(tf.random_normal([1],stddev=0.01))
outputs2=outputs[-1]
outputs3=tf.matmul(outputs2,W_o)+b_o
#computethecostforthisbatchofdata
cost=tf.reduce_mean(tf.pow(outputs3-result,2))
#computeupdatestoparametersinordertominimizecost
train_op=tf.train.RMSPropOptimizer(0.005,0.2).minimize(cost)

43. Tutorial-RNN
controlflow중loopcontrol에대한지원이없어Pythonloop로동
작하여성능이슈가있을것으로판단됨[11]
BLSTM,CTC은미구현,가변길이의RNN입력이불가

44. Benchmarks
convnet-benchmarks의자료[12]
AlexNet-Input128x3x224x224
Library Class Time(ms) forward(ms) backward(ms)
Nervana-fp16 ConvLayer 92 29 62
CuDNN[R3]-fp16 cudnn.SpatialConvolution 96 30 66
CuDNN[R3]-fp32 cudnn.SpatialConvolution 96 32 64
Nervana-fp32 ConvLayer 101 32 69
fbfft fbnn.SpatialConvolution 104 31 72
cudaconvnet2* ConvLayer 177 42 135
CuDNN[R2]* cudnn.SpatialConvolution 231 70 161
Caffe(native) ConvolutionLayer 324 121 203
TensorFlow conv2d 326 96 230
Torch-7(native) SpatialConvolutionMM 342 132 210
CL-nn(Torch) SpatialConvolutionMM 963 388 574
Caffe-CLGreenTea ConvolutionLayer 1442 210 123

45. 다른오픈소스들
Theano
2009년부터시작
Symbolicgraph처리를TensorFlow보다먼저구현한라이브러리
scan이라불리는loopingcontrol을지원하여RNN구현에용이함
Block,Keras와같은Theano를기반으로한상위레벨의라이브러리
가존재함
Keras는backend로TensorFlow를지원하고있음
Multi-GPU에대한지원이없음
Python기반
Touch
2002년부터시작
RNN에대한지원이없음
비공식적브랜치에서는지원
YannLeCun이contributor로있는프로젝트
Lua기반
[13]EvaluationofDeepLearningToolkits

46. 다른오픈소스들
기타
Caffe:2013년UCBerkeley,Yangqingjia개발
Deeplearning4J:2014년AdamGibson이개발
DeepDist:2014년Facebook,Dirkneumann이개발
ApacheHorn:2012년분산ANN을기반한Hama에서파생됨,
DistBelief오픈소스화를위한국내개발자프로젝트
[13]EvaluationofDeepLearningToolkits

47. 다른오픈소스들
Keras
지원하는기능이라면구조명세수준
fromkeras.modelsimportSequential
fromkeras.layers.coreimportDense,Dropout,Activation
fromkeras.layers.embeddingsimportEmbedding
fromkeras.layers.recurrentimportLSTM
model=Sequential()
model.add(Embedding(max_features,256,input_length=maxlen))
model.add(LSTM(output_dim=128,activation='sigmoid',inner_activation='hard_si
gmoid'))
model.add(Dropout(0.5))
model.add(Dense(1))
model.add(Activation('sigmoid'))
model.compile(loss='binary_crossentropy',optimizer='rmsprop')
model.fit(X_train,Y_train,batch_size=16,nb_epoch=10)
score=model.evaluate(X_test,Y_test,batch_size=16)

48. TensorFlowv0.5
분산처리관련구현미공개
C++frontend를지원하나Python쪽에집중되어있음
C++에서는그래프생성,자동미분계산등이누락되어있음[5]
C++예제코드는Python에서그래프생성하여파일(protocal
buffer포맷)로저장후로딩하도록되어있음
그래프생성코드는Python으로작성된것으로추정됨
EEG라불리는성능프로파일링툴은미공개상태
[3]TensorFlow:Large-scalemachinelearningonheterogeneoussystems

49. TensorFlow를고려한다면
작은규모의파라미터에서도나쁘지않은성능
아직분산처리는공개되지않았지만단일GPU에서모델이다올라가지
않는다면대안을가진라이브러리는많지않음
RNN쪽지원은아직더기다려야할것같음
v0.5에비해나쁘지않는완성도
구느님의V8엔진같은신적화를기대해볼수있을지도
TensorFlow뿐아니라다른라이브러리도구현코드가줄어들어모델
자체에집중할수있을것으로보임
Keras같은high-level라이브러리를선택하여backend를변경해
도될것같음
이제WYSIWYG툴만나오면됨

50. 설치
GPU요구사항
CUDA7.0,cuDNN6.5
CUDA(Computecapability)버전3.5이상,비공식적으로는
CUDA3.0GPU도지원
Titan,Tesla(K20,K40)권장
Multiprocessor최소8개이상
OS:Linux,OSX
Python2.7
Windows는아직지원하지않음
DockerToolbox를통해시도는가능
버전이다른라이브러리는시도하지않는것이정신건강에좋음

51. 설치
TensorFlow를동작하는방법은3가지방법이있음
Docker이미지
가장편안한방법
Docker사용법을익혀야함
Python패키지
로컬에서작업가능
소스컴파일이필요한작업은확인불가능,몇몇예제는bazel빌드로
동작
소스빌드
가장어려운방법
약간은bazel을익혀야함

52. 설치-Docker
tensorflow,tensorflow-full두가지이미지가있으며tensorflow-full
은소스가포함되어있음
dockerrun-p127.0.0.1:8888:8888-itb.gcr.io/tensorflow/tensorflow-full
-p127.0.0.1:8888:8888는jupyter(iPython)포트를포워딩하기위함
아래와같이jupyter실행이가능함
root@bacf85f6aea3:~#/run_jupyter.sh

53. 설치-Docker(GPU)
Dockertensorflow,tensorflow-full이미지는GPU를지원하지않음
Docker내부에서GPU를동작하게하려면소스를받아야함
gitclone--recurse-submoduleshttps://github.com/tensorflow/tensorflow
GPU사용시아래와같은환경변수가등록되어야함
exportLD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64
exportCUDA_HOME=/usr/local/cuda
tensorflow/tensorflow/tools/docker에docker_run_gpu.sh를실행
하면local의환경변수를읽어docker이미지를생성함

54. 설치-Python
Python2.7을사용함아직3.0은지원하지않음,만약여러버전이공존
한다면virtualenv사용을권장
python-pip,python-dev,python-virtualenv는설치되어있어야함,아
래예시는GPU지원pip패키지를설치
$virtualenv--system-site-packages-p/usr/bin/python2.7tensorflow
$source~/tensorflow/bin/activate
(tensorflow)$pipinstall--upgradehttps://storage.googleapis.com/tensorflow/linux
/gpu/tensorflow-0.5.0-cp27-none-linux_x86_64.whl
GPU사용시아래와같은환경변수가등록되어야함
exportLD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64
exportCUDA_HOME=/usr/local/cuda

55. 설치-Python
정상설치되었다면아래명령이오류를발생하지않음
(tensorflow)$python-c'importtensorflow'
만약아래와같은오류가발생한다면pipinstall
'protobuf=3.0.0a3'로protobuf버전을일치시켜야함
File/home/shchoi/pyenv/tensorflow/local/lib/python2.7/site-packages/tensorflow
/core/framework/tensor_shape_pb2.py,line22,inmodule
serialized_pb=_b('n,tensorflow/core/framework/tensor_shape.protox12nte
nsorflowdnx10TensorShapeProtox12-nx03x64imx18x02x03(
x0bx32.tensorflow.TensorShapeProto.Dimx1a!nx03x44imx12x0c
nx04sizex18x01x01(x03x12x0cnx04namex18x02x01(
tbx06proto3')
TypeError:__init__()gotanunexpectedkeywordargument'syntax'

56. 설치-소스
소스클론및관련라이브러리설치,아울러구글이만든빌드툴인Bazel
을설치해야함
$gitclone--recurse-submoduleshttps://github.com/tensorflow/tensorflow
$sudoapt-getinstallpython-numpyswigpython-dev

57. 설치-소스
configure실행
$cdtensorflow
$./configure
DoyouwishtobuildTensorFlowwithGPUsupport?[y/n]y
GPUsupportwillbeenabledforTensorFlow
PleasespecifythelocationwhereCUDA7.0toolkitisinstalled.Referto
README.mdformoredetails.[defaultis:/usr/local/cuda]:/usr/local/cuda
PleasespecifythelocationwhereCUDNN6.5V2libraryisinstalled.Referto
README.mdformoredetails.[defaultis:/usr/local/cuda]:/usr/local/cuda
SettingupCudainclude
SettingupCudalib64
SettingupCudabin
SettingupCudanvvm
Configurationfinished

58. 설치-소스
만약CUDA3.0GPU라면아래와같이실행하여설정이가능함
TF_UNOFFICIAL_SETTING=1./configure

59. 설치-소스
예제프로그램빌드및실행
$bazelbuild-copt--config=cuda//tensorflow/cc:tutorials_example_trainer
$bazel-bin/tensorflow/cc/tutorials_example_trainer
...
000008/000007lambda=2.000000x=[0.894427-0.447214]y=[1.788854-0.8
94427]
000002/000000lambda=2.000000x=[0.894427-0.447214]y=[1.788854-0.8
94427]
000004/000000lambda=2.000000x=[0.894427-0.447214]y=[1.788854-0.8
94427]
000008/000007lambda=2.000000x=[0.894427-0.447214]y=[1.788854-0.8
94427]
000002/000000lambda=2.000000x=[0.894427-0.447214]y=[1.788854-0.8
94427]
000008/000007lambda=2.000000x=[0.894427-0.447214]y=[1.788854-0.8
94427]
000004/000000lambda=2.000000x=[0.894427-0.447214]y=[1.788854-0.8
94427]
$

60. 설치-소스
아래와같은오류가발생한다면구글링필요,대부분환경설정관련오류
bazelbuild-copt//tensorflow/cc:tutorials_example_trainer
......
WARNING:Sandboxedexecutionisnotsupportedonyoursystemandthushermeti
cityofactionscannotbeguaranteed.Seehttp://bazel.io/docs/bazel-user-manual.ht
ml#sandboxingformoreinformation.Youcanturnoffthiswarningvia--ignore_un
supported_sandboxing.
ERROR:Loadingoftarget'//tools/jdk:ijar'failed;buildaborted:nosuchpackage'to
ols/jdk':BUILDfilenotfoundonpackagepath.
ERROR:Loadingfailed;buildaborted.
INFO:Elapsedtime:0.565s
이경우는.bazelrc를설정하여해결
build--package_path%workspace%:/home/shchoi/share/src/bazel-0.1.1/base_wo
rkspace
fetch--package_path%workspace%:/home/shchoi/share/src/bazel-0.1.1/base_wo
rkspace
query--package_path%workspace%:/home/shchoi/share/src/bazel-0.1.1/base_wo
rkspace

61. 설치-소스
Python을위한pippackage는아래와같이빌드가가능함
bazelbuild-copt//tensorflow/tools/pip_package:build_pip_package

62. 참고자료
[1]Large-ScaleDeepLearningforIntelligentComputer
Systems
[2]Introductiontoapachehorn(incubating)
[3]TensorFlow:Large-scalemachinelearningon
heterogeneoussystems
[4]LargeScaleDistributedDeepNetworks
[5]LoadingaTensorFlowgraphwiththeC++API
[6]MNISTForMLBeginners
[7]Github-DistributedVersion
[8]UsingGPUs
[9]rnn_example.ipynb
[10]Hochreiter,Sepp,andJürgenSchmidhuber.Longshort-
termmemory.Neuralcomputation9.8(1997):1735-1780.
[11]Symbolicloops(likescaninTheano)
[12]convnet-benchmarks
[13]EvaluationofDeepLearningToolkits