SunPro会誌 2016 技術書典

Size: px

Start display at page:

Download "SunPro会誌 2016 技術書典"

ゆあつちた
5 years ago
Views:

1 1 hiromu 1.1 SunPro Deep Convolutional Generative Adversarial Network (DCGAN) DCGAN 3 DCGAN non-root TensorFlow 1.2 DCGAN DCGAN Generative Adversarial Network(GAN, ) GAN 2014 Goodfellow [1] 2

2 1 2 GAN GAN Generator Discriminator 2 Generator z Discriminator ( ) Generator Discriminator Generator Generator Discriminator Generator Discriminator 1.1 GAN 2 Discriminator Generator Discriminator

3 3 1 DCGAN GAN Discriminator Generator Generator 2015 Radford DCGAN[2] Radford GAN Generator Discriminator GAN Springenberg [3] Generator ReLU( Tanh) Discriminator LeakyReLU non-root Tensorflow DCGAN Tensorflow PC 3 3 CPU 100% Amazon EC2 Tensorflow OS( ) CentOS 6.5 Tensorflow Python 2.7 Python 2.6 root quota 3GB miniconda Python 2.7 Anaconda Python (numpy scipy ) Anaconda 3GB miniconda

4 1 4 miniconda *1 Python 2.7 Linux 64-bit # $ https :// repo. continuum. io / miniconda / Miniconda2 - latest - Linux - x86_64. sh # $ bash Miniconda2 - latest - Linux - x86_64. sh # P A T H $ export PATH ="/ home / hiromu / miniconda2 / bin : $PATH " /home/hiromu/miniconda2 miniconda Tensorflow Python Tensorflow *2 miniconda Tensorflow pip # T e n s o r f l o w o p e n s s l z l i b $ conda create -n tensorflow python =2.7 # T e n s o r f l o w $ source activate tensorflow # Linux 64 - bit, CPU only, Python 2.7 ( U R L ) ( tensorflow )$ export TF_BINARY_URL =https :// storage. googleapis. com / tensorflow / linux / cpu / tensorflow rc0 - cp27 - none - linux_x86_64. whl # T e n s o r f l o w ( tensorflow )$ pip install -- upgrade $TF_BINARY_URL $ python Python Continuum Analytics, Inc. ( default, Dec , 18:08:32) < > >>> import tensorflow Traceback ( most recent call last ): File "< stdin >", line 1, in < module > File "/ home / hiromu / miniconda2 / envs / tensorflow / lib / python2.7/ site - packages / tensorflow / init. py ", line 23, in < module > from tensorflow. python import * File "/ home / hiromu / miniconda2 / envs / tensorflow / lib / python2.7/ site - packages / tensorflow / python / init. py ", line 48, in <module > from tensorflow. python import pywrap_tensorflow File "/ home / hiromu / miniconda2 / envs / tensorflow / lib / python2.7/ site - packages / tensorflow / python / pywrap_tensorflow. py ", line 28, in <module > _pywrap_tensorflow = swig_import_helper () File "/ home / hiromu / miniconda2 / envs / tensorflow / lib / python2.7/ site - packages / tensorflow / python / pywrap_tensorflow. py ", line 24, in swig_import_helper _mod = imp. load_module (' _pywrap_tensorflow ', fp, pathname, description ) ImportError : / lib64 / libc. so.6: version `GLIBC_2.17 ' not found ( required by / home / hiromu / miniconda2 / envs / tensorflow / lib / python2.7/ site - packages / tensorflow / python / _pywrap_tensorflow. so ) * 1 * 2

5 5 1 Tensorflow CentOS glibc root glibc glibc glibc-2.17 rpmfind.net CentOS 7.2 glibc-2.17 *3 rpm miniconda Tensorflow # ( ) $ cd / home / hiromu / miniconda2 / envs / tensorflow # g l i b c $ wget ftp :// fr2. rpmfind. net / linux / centos / / updates / x86_64 / Packages / glibc el7_2.6. x86_64. rpm # $ rpm2cpio glibc el7_2.6. x86_64. rpm cpio -idv Tensorflow glibc CentOS 7.2 libstdc *4 # libstdc ++ $ wget ftp :// fr2. rpmfind. net / linux / centos / / os / x86_64 / Packages / libstdc el7. x86_64. rpm # $ rpm2cpio libstdc el7. x86_64. rpm cpio -idv Python $ / home / hiromu / miniconda2 / envs / tensorflow / lib64 /ld - linux -x so.2 -- library - path / home / hiromu / miniconda2 / envs / tensorflow / lib64 :/ home / hiromu / miniconda2 / envs / tensorflow / usr / lib64 / home / hiromu / miniconda2 / envs / tensorflow / bin / python Python Continuum Analytics, Inc. ( default, Dec , 18:08:32) < > >>> import tensorflow >>> alias $ alias tfpy ="/ home / hiromu / miniconda2 / envs / tensorflow / lib64 /ld - linux -x so.2 -- library - path / home / hiromu / miniconda2 / envs / tensorflow / lib64 :/ home / hiromu / miniconda2 / envs / tensorflow / usr / lib64 / home / hiromu / miniconda2 / envs / tensorflow / bin / python " Python scipy Pillow * x86_64.html * html

6 1 6 # s c i p y p i p $ pip install scipy # P i l l o w c o n d a $ conda install pillow 1.4 DCGAN UEC FOOD-256 UEC FOOD-256 [4] Web *5 FOOD *6 DCGAN PFID Pittsburgh Fast-Food Image Dataset Intel Labs Pittsburgh [6] Web * * 5 * 6 * 7

7 7 1 Food 101 Food 101 [5] Web * (tar.gz 5GB ) 1 Food 101 DCGAN 64x64px ( ) import sys from PIL import Image convert.py for path in sys. argv [1:]: image = Image. open ( path ) if image. mode!= 'RGB ': print ' Error : %s ' % path continue ratio = 64.0 / min ( image. size ) image. thumbnail ( map ( lambda x: int (x * ratio ), image. size )) x, y = map ( lambda x: (x - 64) / 2, image. size ) image. crop ((x, y, x + 64, y + 64)). save ( path ) # f o o d i m a g e s $ python convert. py */*. jpg Error : bread_pudding / jpg Error : lasagna / jpg Error : steak / jpg $ ln -s */*. jpg./ $ rm jpg jpg jpg 3 ( ) * 8

1 8 1.2 1.5 DCGAN DCGAN Github Tensorflow *9 data food101 Tensorflow Web quota tmpfs Gist *10 # / t m p $ mktemp -d / tmp / tf. XXXXX / tmp / tf. gq5vz $ python main.

8 DCGAN DCGAN Github Tensorflow *9 data food101 Tensorflow Web quota tmpfs Gist *10 # / t m p $ mktemp -d / tmp / tf. XXXXX / tmp / tf. gq5vz $ python main. py -- dataset food is_train True -- log_dir / tmp / tf. gq5vz Epoch: 01 [0899/7950] time: , d_loss: f, g_loss: Epoch %(899/7950) * 9 * 10

9 9 1 d_loss Discriminator ( ) g_loss Generator Tensorboard Web 6006 Tensorboard d_loss g_loss $ tfpy / home / hiromu / miniconda2 / envs / tensorflow / bin / tensorboard -- logdir =/ tmp / tf. gq5vz *11 Epoch * 11 ( 28d69cce110aafcab4152ef9e807a030)

10 第 1 章ディープラーニングでご飯を作ってみる 10 図 1.3 DCGAN によるご飯画像の生成結果 1.6 おわりにここまで長文をお読みいただきありがとうございましたもともとは画像に対応した z ベクトルを求め料理画像の足し算や引き算モーフィングなどで遊んでみたかったのですがここまで書き進めた時点で締め切りを 5 日も過ぎているという状況で進捗ダメでしたこの z ベクトルを使った演算は画像生成の種となるデータに制約を設けずランダムなベクトルから画像を対応づける写像 (Generator) を学習するという DCGAN

11 11 1 ( IME ) 1 DCGAN hiromu1996[at]gmail.com 1.7 [1] Goodfellow, Ian, et al. Generative adversarial nets. Advances in Neural Information Processing Systems, [2] Radford, Alec, Metz, Luke, and Chintala, Soumith. Unsupervised representation learning with deep convolutional generative adversarial networks. International Conf. on Learning Representations, [3] Springenberg, Jost Tobias, et al. Striving for simplicity: The all convolutional net. International Conf. on Learning Representations, [4] Kawano, Yoshiyuki, and Yanai, Keiji. Automatic Expansion of a Food Image Dataset Leveraging Existing Categories with Domain Adaptation. Proc. of ECCV Workshop on TASK-CV, [5] Bossard, Lukas, Guillaumin, Matthieu and Van Gool, Luc. Food Mining Discriminative Components with Random Forests. European Conf. on Computer Vision, [6] Mei, Chen, et al. PFID: Pittsburgh fast-food image dataset. IEEE Intl. Conf. on Image Processing, 2009.

Anaconda (2019/7/3)

Anaconda (2019/7/3) Published on Research Center for Computational Science (https://ccportal.ims.ac.jp) Home > Anaconda3-2019.03 (2019/7/3) Anaconda3-2019.03 (2019/7/3) 1 利用方法 conda, anaconda に関する情報はウェブ上にたくさんありますのでそれらも参考にしてください