Kochi University of Technology Aca Title 環境分野への深層学習応用研究の立ち上げについて Author(s) 中根, 英昭, 若槻, 祐貴 Citation 高知工科大学紀要, 15(1): Date of issue U

Size: px

Start display at page:

Download "Kochi University of Technology Aca Title 環境分野への深層学習応用研究の立ち上げについて Author(s) 中根, 英昭, 若槻, 祐貴 Citation 高知工科大学紀要, 15(1): Date of issue U"

たかよしのしろ
5 years ago
Views:

Kochi University of Technology Aca Title 環境分野への深層学習応用研究の立ち上げについて Author(s) 中根, 英昭, 若槻, 祐貴 Citation 高知工科大学紀要, 15(1): 111-120 Date of

1 Kochi University of Technology Aca Title 環境分野への深層学習応用研究の立ち上げについて Author(s) 中根, 英昭, 若槻, 祐貴 Citation 高知工科大学紀要, 15(1): Date of issue URL Rights Text version publisher Kochi, JAPAN

2 1 2 ( ) nakane.hideaki@kochi-tech.ac.jp MLP CNN ) PC GPU Graphics Processing Unit ) (a) DNA (b) 3, 4) 111

3 ) GPU PC OS Linux C++ GPU Caffe Chainer Tensorflow H 2 O Tensorflow Tensorflow GPU Linux-PC Tensorflow Windows/Linux OS PC GPU GeForce GTX1070 GPU Tensorflow Google Windows Tensorflow ) GPU Linux Keras Tensorflow Theano Qiita Tensorflow Windows GPU ) Window-PC Tensorflow 1. CPU Core i GPU GeForce GTX1070 or GTX1080Ti OS Windows 10 Keras / TensorFlow Python 3.6 (Anaconda 4.4.0) 0 9 MNIST CPU 1 GPU 1 R Python Tensorflow Tensorflow Keras 8) Python Keras Tensorflow Theano CNTK 8) Tensorflow Python Anaconda Anaconda Spyder R Rstudio 1 Windows PC GPU GTX GTX1080Ti Keras / Tensorflow Windows Keras Francois Chollet Deep Learning with Python Software / hardware requirement 9) You ll need access to a UNIX machine; it s possible to use Windows too, but I don t recommend it 112

4 1. CNN 1 Linux 3.2 Keras / Tensorflow ) CNN 7 Keras / Tensorflow , Adam ReLU softmax RGB Keras Python 1 CNN ) , RGB , ch 2 17, CNN ) 113

5 CNN ) Keras Deep Learning with Keras regression network 14) Keras 4. model.add(activation("softmax")) categorical crossentropy mean squared error ,550 MLP CNN MLP RMSE m 3 /s 114

6 m 3 /s 15) MLP CNN 2 MLP 1 km 16) MLP 17) ,

7 ) 31 km m m m

Xception 20) fine-tuning 100 21) 2017 4 11.

8 ) 4.3 CNN Deep Learning with Python 9) ) 1000 ImageNet Xception 20) fine-tuning ) ) Fine-tuning Xception Keras Documentation 8) m 2.4 m

9 CNN MLP CNN MLP ) Keras / Tensorflow Chainer 118

10 SI-CAT 1), 1, AI 2017, pp , ),, URL = kochi-tech.ac.jp/frontier.html 3) J. Miyake, Y. Kaneshita, S. Asatani, S. Tagawa, Shimoda, H. Niioka and T. Hirano, Graphical classification of DNA sequences of HLA alleles by Deep learning, Human Cell, published online, URL = /s , ) H. Niioka, S. Asatani, A. Yoshimura, H. Ohigashi, S. Tagawa, and J. Miyake, Classification of C2C12 cells at differentiation by convolutional neural network of deep learning using phase contrast images, Human Cell, published online, URL = %2Fs , ),,,,,, 2016, ),,,, ) Qiita, TensorFlow Windows GPU, URL = 24e9491eb8a4ce42eea6, ) Keras Documentation, URL = 9) F. Chollet, Software/hardware requirement, Deep Learning with Python, Manning, ), Deep Learning km, 28, ),, URL = 12),,,,,,,, , pp. 1 6, ),,, URL = SrchDamData.exe?ID= ) A. Gulli and S. Pal, regression network, Deep Learning with Keras, pp , Packt, ),, 29, ), URL = know/kurashi/kaiseki.html. 17),, 29, ),, 29, ),,,, B1, Vol. 72, No. 4, pp. I 187 I 192, ) F. Chollet, Xception: Deep Learning with Depthwise Separable Convolutions, URL = arxiv.org/abs/ v3, ), / 3, 29, ),, 29, ), CNN, 29,

11 Startup of Deep Learning Application to Environmental Research Hideaki Nakane 1 Yuki Wakatsuki 2 (Received: May 7th, 2018) 1 School of Environmental Science and Engineering, Kochi University of Technology 185 Miyanokuchi, Tosayamada, Kami City, Kochi , JAPAN 2 Graduate School of Engineering, Kochi University of Technology 185 Miyanokuchi, Tosayamada, Kami City, Kochi , JAPAN nakane.hideaki@kochi-tech.ac.jp Abstract: We implemented a computing environment for deep learning, a form of artificial intelligence technology, at the Nakane Laboratory in the second half of FY In FY 2017, deep learning technologies were successfully applied to various environmental research themes. We developed regression and classification models for the following targets. For regression, we used the amount of inflow into the Sameura Dam, the water level of the Yasuda River, the water level of the Kagami River, and the total amount of solar insolation in Kochi. The targets of classification were the presence / absence of precipitation in Kochi, three species of flowers / lilies, and forest tree species. Although our objective is diverse, we could create estimation models using two kinds of deep learning methods; multilayer perceptron (MLP) and convolutional neural network (CNN). In regression cases, it was shown that the water levels and the inflow amounts to the dam during entire periods from flooding to droughts can be well predicted using the MLP models. The CNN models with transfer study demonstrated performance equivalent to or better than humans for classification of tree species. 120

( ) % 50 % )

( ) % 50 % ) Kochi University of Technology Aca 佐岡地区本村における歴史景観の調査屋敷地の変 Title 遷から読み解く集落景観の特質 Author(s) 池内, 克徳, 藤原, 駿, 渡辺, 菊眞, 楠瀬, 慶太 Citation 高知工科大学紀要, 15(1): 57-68 Date of 2018-07-31 issue URL http://hdl.handle.net/10173/1944