Vol.58 No (Apr. 2017) API 1,a) , recurrent neural network 10 38% 1 An API Suggestion Using Recurrent Neural Networks Tetsu

Size: px

Start display at page:

Download "Vol.58 No (Apr. 2017) API 1,a) , recurrent neural network 10 38% 1 An API Suggestion Using Recurrent Neural Networks Tetsu"

たけなりみのしま
3 years ago
Views:

1 API 1,a) , recurrent neural network 10 38% 1 An API Suggestion Using Recurrent Neural Networks Tetsuo Yamamoto 1,a) Received: August 8, 2016, Accepted: January 10, 2017 Abstract: Developers reuse existing source code or use libraries to develop effectively. In this study, we focus on the order of method invocation statements in existing source code and propose to suggest method invocation statements. This paper shows an approach to suggest method invocation statements using recurrent neural network. We have implemented the approach and conducted experiments to measure an accuracy with 10 open source software projects. We have investigated various prameters of recurrent neural network. Our evaluation has shown that our approach is 38% accuracy in API code suggestion, it can correctly suggest the API with top 1 candidate. Keywords: code suggestion, recurrent neural netowork, deep learing 1. Android Android API 42% [17] 1 College of Engineering, Nihon University, Koriyama, Fukushima , Japan a) tetsuo@cs.ce.nihon-u.ac.jp 1 Andorid View TextView Button onclicklistener API c 2017 Information Processing Society of Japan 769

2 j w j 1 j 1 w j 1 1 = w 1,w 2,w 3,...,w j 1 w j 1 1 w j P (w j w j 1 1 ) s T P (w T 1 )= T j=1 P (w j w j 1 1 ) 1 Java Fig. 1 A Java code snippet. API [1], [4], [13], [14], [20] API [13] API [20] 2 API recurrent neural netowork 10 38% Bengi [3] NNLM NNLM w j 1 0 N N 1-of-N I have a pen and a eraser. N 6 I 1 have 2 a 3 pen 4 and 5 eraser 6 I 1-of-N (1, 0, 0, 0, 0, 0) T have 1-of-N (0, 1, 0, 0, 0, 0) T w j 1 j n+1 w j n n 3 I have a have a pen a pen and 2.2 Bengi n Mikolov RNNLM [11], [12] RNN T c 2017 Information Processing Society of Japan 770

3 w T 1 = w 1,w 2,w 3,...,w T t =1, 2, 3,... t 2 RNNLM RNN 2 w s y t w(t) y(t) s(t) w(t) t 1-of-N N M U N M s(t) W t 1 s(t) =f(uw(t)+ws(t 1)) y(t) V M N y(t) =g(vs(t)) f(z) sigmoid f(z) = 1 1+e z g(z m ) softmax g(z m )= ezm k ez k RNN Backpropagation Through Time BPTT [19] BPTT RNN Truncated BPTT Long Short-Term Memory LSTM RNNLM RNNLM Java Fig. 2 RNN An architecture of RNN. Fig. 3 3 Flow of proposed method. c 2017 Information Processing Society of Japan 771

4 RNN 2.2 RNNLM 4 Fig. 4 List of method call statements. 3.2 API <init> # java.util.arraylist<string> add java.util.arraylist#add new java.util.arraylist<string>() java.util.arraylist#<init> 3.3 API 3.2 JDK java <eom> android 5 RNNLM Fig. 5 RNNLM unfolded in time. 1 <eom> RNNLM 3.4 RNN 2.2 RNNLM of-N RNNLM truncated BPTT <eom> BPTT BPTT LSTM Java <eom> 4 RNNLM 5 1 android.view.layoutinflater#inflate c 2017 Information Processing Society of Japan 772

5 y(1) android.view.view#findviewbyid y(2) 1 s(1) 2 s(2) <eom> y(1) y(6) 6 y(1) 2 android.view.view#findviewbyid y(1) android.view.view#findviewbyid 1-of-N y(2) 3 android.widget.textview#settext y(1) y(6) M epoch <eom> 1 RNN Android 1 android.view.layoutinflater#inflate 2 android.view.view#findviewbyid 3 android.widget.textview#settext 3 y(3) softmax softmax 1 1-of-N y(3) android.view.view#findviewbyid 1-of-N android.view.view#findviewbyid Top-k Android 10 Android API android 1 android-23 *1 Android SDK API 23 Android-Universal-Image-Loader *2 MPAn- *1 html#downloads *2 Loader c 2017 Information Processing Society of Japan 773

6 1 Table 1 Target projects. LOC android-23 1, ,670 3,058 Android-Universal -Image-Loader 88 13, MPAndroidChart , SlidingMenu 33 4, ViewPagerIndicator 42 4, butterknife 91 10, fresco , glide , iosched , picasso 76 13, droidchart *3 SlidingMenu *4 ViewPagerIndicator *5 butterknife *6 fresco *7 glide *8 iosched *9 picasso * 10 GitHub Android Java 1 Java LOC RNN android ,491 4,784 29,463 3,019 RNN Chainer * 11 NVIDIA GeFroce GTX GB memory GPU MB 2 M 512 *3 *4 *5 *6 *7 *8 *9 *10 *11 epoch a 1,a 2,...,a k, <eom> a 1 a 2 a 2 a 3 a k 1 k k 1 a k 1 a 1 a k 1 a k android.view.layoutinflater#inflate android.view.view 2 android.view.layoutinflater#inflate android.view.view#findviewbyid 2 android.widget.textview Top-k android-23 9 c 2017 Information Processing Society of Japan 774

7 2 android.view.layoutinflater#inflate Table 2 Method calls following android.view. LayoutInflater#inflate. 1 android.view.view#findviewbyid android.view.view#getbottom android.view.view#getviewtreeobserver android.view.view#getheight android.view.view#gettop android.view.view#settag android.view.view#setvisibility android.view.view#setfocusable android.view.view#setonclicklistener android.view.view#getleft android.view.view#getmeasuredheight android.view.view#requestlayout android.view.view#setlayoutparams android.view.view#getvisibility android.view.view#gettag android.view.view#offsetleftandright android.view.view#<init> android.view.view#getlayoutparams android.view.view#setclickable android.view.view#getpaddingbottom android.view.view#findviewbyid Table 3 Method calls following android.view. View#findViewById. 1 android.widget.textview#settext android.widget.textview#<init> android.widget.textview#setvisibility android.widget.textview#settypeface android.widget.textview#settextsize android.widget.textview# setmovementmethod 7 android.widget.textview# setbackgroundresource 8 android.widget.textview#setpadding android.widget.textview#settag android.widget.textview# setonclicklistener 11 android.widget.textview#append android.widget.textview#settextcolor android.widget.textview#setallcaps android.widget.textview#gettext android.widget.textview#setgravity android.widget.textview# setcontentdescription 17 android.widget.textview#setlayoutparams android.widget.textview#setid android.widget.textview# announceforaccessibility M 512 epoch epoch Top-k k Top-1 epoch 2 38% android-23 8,482 3, epoch 4 6 epoch epoch API epoch epoch M , android-23 glide android-23 API glide c 2017 Information Processing Society of Japan 775

8 4 Table 4 android-23 A summary of the accuracy of andorid Top-1 38% 38% 38% 35% 33% 34% 33% 33% 34% 32% 33% 31% 33% 31% 32% Top-5 75% 74% 73% 72% 72% 71% 71% 71% 72% 70% 72% 70% 71% 70% 70% Top-10 88% 89% 88% 88% 88% 86% 87% 86% 87% 88% 88% 86% 87% 86% 85% Top-20 97% 97% 96% 96% 96% 96% 96% 96% 96% 96% 96% 96% 95% 95% 94% Top-1 31% 31% 31% 31% 32% 31% 32% 33% 31% 30% 31% 31% 31% 30% 30% Top-5 69% 69% 70% 70% 69% 69% 68% 69% 68% 67% 70% 69% 68% 68% 68% Top-10 85% 85% 85% 85% 85% 84% 85% 86% 84% 85% 85% 85% 84% 85% 85% Top-20 95% 95% 95% 95% 95% 94% 95% 96% 95% 95% 95% 95% 94% 95% 95% 7 3 Table 7 Resultsof3hiddenlayers. 8,482 Top-1 3,284 38% Top-5 6,528 76% Top-10 7,573 89% Top-20 8,232 97% 6 android-23 Fig. 6 A line graph of the accuracy of andorid Table 5 Results of 1 hidden layer. 8,482 Top-1 3,012 35% Top-5 6,062 71% Top-10 7,360 86% Top-20 8,129 95% 6 2 Table 6 Resultsof2hiddenlayers. 8,482 Top-1 3,242 38% Top-5 6,398 75% Top-10 7,494 88% Top-20 8,255 97% API API 8 2 1,024 Table 8 Results of 2 hidden layers and 1,024 Units. 8,482 Top-1 3,219 37% Top-5 6,386 75% Top-10 7,506 88% Top-20 8,257 97% Robbes [15] n Top-k 2 Top-1 60% Top-10 88% View- PageIndicator Top-1 50% Top-10 90%Robbes Top-1 c 2017 Information Processing Society of Japan 776

9 9 Table 9 A comparison among projects. Total Top-1 Top-5 Top-10 Top-20 Top-1 Top-5 Top-10 Top-20 accuracy accuracy accuracy accuracy Android-Universal-Image-Loader % 80% 87% 96% MPAndroidChart 1, ,011 1,112 33% 68% 86% 94% SlidingMenu % 67% 85% 92% ViewPagerIndicator % 48% 70% 97% android-23 8,482 3,242 6,398 7,494 8,255 38% 75% 88% 97% butterknife % 45% 60% 62% fresco % 59% 80% 91% glide % 76% 89% 95% iosched 2, ,883 2,282 2,568 32% 70% 85% 95% picasso % 61% 81% 94% 10 2 Table 10 The accuracy of the first 2 letters of method name. Top-1 Top-10 accuracy accuracy Android-Universal-Image-Loader 60% 94% MPAndroidChart 63% 94% SlidingMenu 52% 95% ViewPagerIndicator 40% 90% android-23 68% 97% butterknife 89% 100% fresco 61% 97% glide 61% 97% iosched 56% 96% picasso 58% 96% Top-10 butterknife Top-1 89% 4.4 Mikolov RNNLM RNNLM Top-1 38% Top-1 100% API open - read - close open - read - read - close 2 open - read API close read 2 Top-1 100% Top-5 Top-10 API 4 Top-5 70% Top-10 80% API API 4.5 RNNLM RNNLM RNNLM Andorid API API c 2017 Information Processing Society of Japan 777

10 API API API 5. API Prospector [8] Xsnippet [16] PARSEWeb [18] Michail API [9], [10] Strathcona [6], [7] API Mishne [13] API API Bruch [4] HMM Han [5] HMM API Nguen [14] HMM Andorid API ASTLan [1] AST AST CSCC [2] API [20] 2 API API 6. API 10 38% 1 API JSPS 15K00108 [1] Nguyen, A.T. and Nguyen, T.N.: Graph-Based Statistical Language Model for Code, Proc IEEE/ACM 37th IEEE International Conference on Software Engineering, pp (2015). [2] Asaduzzaman, M., Roy, C.K., Schneider, K.A. and Hou, D.: CSCC: Simple, Efficient, Context Sensitive Code Completion, Proc IEEE International Conference on Software Maintenance and Evolution, pp (2014). [3] Bengio, Y., Ducharme, R., Vincent, P. and Janvin, C.: A Neural Probabilistic Language Model, The Journal of Machine Learning Research, Vol.3, pp (2003). [4] Bruch, M., Monperrus, M. and Mezini, M.: Learning from examples to improve code completion systems, Proc. 7th Joint Meeting of the European Software Engineering Conference (ESEC ) and the ACM SIGSOFT Symposium on the Foundations of Software Engineering, pp (2009). [5] Han, S.H.S., Wallace, D. and Miller, R.: Code Completion from Abbreviated Input, Proc. 24th IEEE/ACM c 2017 Information Processing Society of Japan 778

International Conference on Automated Software Engineering, pp.332 343 (2009). [6] Holmes, R. and Murphy, G.C.: Using structural context to recommend source code examples, Proc.

11 International Conference on Automated Software Engineering, pp (2009). [6] Holmes, R. and Murphy, G.C.: Using structural context to recommend source code examples, Proc. 27th International Conference on Software Engineering, pp (2005). [7] Holmes, R., Walker, R. and Murphy, G.: Approximate Structural Context Matching: An Approach to Recommend Relevant Examples, IEEE Trans. Software Engineering, Vol.32, No.12, pp (2006). [8] Mandelin, D., Xu, L., Bodik, R. and Kimelman, D.: Jungloid mining: Helping to navigate the API jungle, Proc ACM SIGPLAN Conference on Programming Language Design and Implementation, pp (2005). [9] Michail, A.: Data mining library reuse patterns using generalized association rules, Proc. 22nd International Conference on Software Engineering, pp (2000). [10] Michail, A.: Browsing and searching source code of applications written using a GUI framework, Proc. 24th International Conference on Software Engineering, pp (2002). [11] Mikolov, T., Karafiat, M., Burget, L., Cernocky, J. and Khudanpur, S.: Recurrent Neural Network based Language Model, Interspeech 2010, pp (2010). [12] Mikolov, T. and Kombrink, S.: Extensions of recurrent neural network language model, ICASSP, pp (2011). [13] Mishne, A., Shoham, S. and Yahav, E.: Typestate-based semantic code search over partial programs, ACM SIG- PLAN Notices, Vol.47, No.10, pp (2012). [14] Nguyen, T.T., Pham, H.V., Vu, P.M. and Nguyen, T.T.: Recommending API Usages for Mobile Apps with Hidden Markov Model, Proc IEEE/ACM International Conference on Automated Software Engineering, pp (2015). [15] Robbes, R. and Lanza, M.: Improving code completion with program history, Automated Software Engineering, Vol.17, No.2, pp (2010). [16] Sahavechaphan, N. and Claypool, K.: XSnippet: Mining For sample code, Proc. 21st Annual ACM SIGPLAN Conference on Object-Oriented Programming Systems, Languages, and Applications, pp (2006). [17] Syer, M.D., Nagappan, M., Hassan, A.E. and Adams, B.: Revisiting Prior Empirical Findings For Mobile Apps: An Empirical Case Study on the 15 Most Popular Open- Source Android Apps, CASCON: Conference of the Center for Advanced Studies on Collaborative Research, pp (2013). [18] Thummalapenta, S. and Xie, T.: Parseweb: A programmer assistant for reusing open source code on the web, Proc. 22nd IEEE/ACM International Conference on Automated Software Engineering, pp (2007). [19] Werbos, P.J.: Backpropagation Through Time: What It Does and How to Do It, Proc. IEEE, Vol.78, No.10, pp (1990). [20] Vol.56, No.2, pp (2015) IEEE-CS c 2017 Information Processing Society of Japan 779

& Vol.5 No (Oct. 2015) TV 1,2,a) , Augmented TV TV AR Augmented Reality 3DCG TV Estimation of TV Screen Position and Ro

& Vol.5 No (Oct. 2015) TV 1,2,a) , Augmented TV TV AR Augmented Reality 3DCG TV Estimation of TV Screen Position and Ro TV 1,2,a) 1 2 2015 1 26, 2015 5 21 Augmented TV TV AR Augmented Reality 3DCG TV Estimation of TV Screen Position and Rotation Using Mobile Device Hiroyuki Kawakita 1,2,a) Toshio Nakagawa 1 Makoto Sato