TensorFlowで趣味の画像収集サーバーを作る 4月号

Slide 1

Slide 1 text

C-LIS CO., LTD. #dltfb https://goo.gl/MX72Bc

Slide 2

Slide 2 text

C-LIS CO., LTD. ༗ࢁܓೋʢ,FJKJ"3*:"."ʣ C-LIS CO., LTD. AndroidΞϓϦ։ൃऀ ػցֶशॳ৺ऀ ΍ͬͯ·ͤΜ Photo : Koji MORIGUCHI (AUN CREATIVE FIRM)

Slide 3

Slide 3 text

C-LIS CO., LTD. IUUQGVLVZVLJOFUQPTU ;͘Ώ͖ϒϩά AlphaGoͷαʔόʔྉۚ͸࠷௿60ԯԁ!!? #dltfb

Slide 4

Slide 4 text

C-LIS CO., LTD. IUUQGVLVZVLJOFUQPTU ;͘Ώ͖ϒϩά AlphaGoͷαʔόʔྉۚ͸࠷௿60ԯԁ!!? #dltfb

Slide 5

Slide 5 text

C-LIS CO., LTD. #dltfb

Slide 6

Slide 6 text

5FOTPS'MPXͰ  झຯͷը૾ऩूαʔόʔΛ࡞Δ ݄̐߸ #dltfb https://goo.gl/MX72Bc

Slide 7

Slide 7 text

C-LIS CO., LTD. (%(,PCFʢBUେࡕ޻ۀେֶʣ

Slide 8

Slide 8 text

C-LIS CO., LTD. (%(,ZPUPʢBUۙـେֶʣ

Slide 9

Slide 9 text

C-LIS CO., LTD. "OESPJE#B[BBSBOE$POGFSFODFʢ੨ࢁֶӃେֶʣ Google TensorFlowͱAndroid͕ܨ͕Δະདྷ IUUQTTQFBLFSEFDLDPNLFJKJHPPHMFUFOTPSqPXUPBOESPJELBYJLBSVXFJMBJ

Slide 10

Slide 10 text

C-LIS CO., LTD. ษڧձͰͷൃදςʔϚ IUUQTHEHLPCFEPPSLFFQFSKQFWFOUT

Slide 11

Slide 11 text

޷Έͷ؟ڸ່ͬը૾ΛࣗಈͰऩू͍ͨ͠

Slide 12

Slide 12 text

C-LIS CO., LTD. ࣮ݱΛ્Ή̏ͭͷน σʔλͷน ܭࢉྔͷน ػցֶशͷน

Slide 13

Slide 13 text

C-LIS CO., LTD. ࡞੒ͨ͠σʔληοτ ؟ڸ່ͬը૾ຕ ඇ؟ڸ່ͬը૾ຕ ؟ڸ່ͬ ඇ؟ڸ່ͬ ޡݕग़

Slide 14

Slide 14 text

C-LIS CO., LTD. $*'"34BNQMF ଟΫϥε෼ྨ IUUQTXXXDTUPSPOUPFEV_LSJ[DJGBSIUNM

Slide 15

Slide 15 text

C-LIS CO., LTD. ̎Ϋϥε෼ྨ 1 0

Slide 16

Slide 16 text

C-LIS CO., LTD. ΍ͬͨ͜ͱ ूΊͨσʔλΛ$*'"3ͷϑΝΠϧܗࣜʹ߹ΘͤͯQBDLͯ͠  αϯϓϧϓϩάϥϜʹ༩͑ͨ Label Red Green Blue 1 0 Label Red Green Blue 1 Label Red Green Blue 2 :

Slide 17

Slide 17 text

C-LIS CO., LTD. DJGBSQZͷϞσϧ ̎૚ͷંΓ৞ΈɾϓʔϦϯά૚ ̎૚ͷશ݁߹૚

Slide 18

Slide 18 text

C-LIS CO., LTD. ݕূ݁Ռ ܇࿅σʔλ ςετσʔλ 1SFDJTJPO ʙ

Slide 19

Slide 19 text

C-LIS CO., LTD. ՝୊ ֶशʢςετʣσʔλͷॆ࣮  ؟ڸͰ͸ͳ͘ɺΩϟϥΫλʔΛݟ͍ͯΔՄೳੑ͕͋Δ ൑ఆਫ਼౓Λ্͛Δ  ʢάϥϑͷਂ૚Խʗύϥϝʔλʔௐ੔ʣ

Slide 20

Slide 20 text

άϥϑͷਂ૚Խ

Slide 21

Slide 21 text

C-LIS CO., LTD. $*'"3ʹ͍ͭͯ  ΘΕΘΕ͸΋ͬͱৄ͘͠஌Δඞཁ͕͋Δ IUUQTXXXUFOTPSqPXPSHWFSTJPOTSUVUPSJBMTEFFQ@DOOJOEFYIUNM

Slide 22

Slide 22 text

-PDBMMZ$POOFDUFE-BZFSJTԿ

Slide 23

Slide 23 text

C-LIS CO., LTD. # local3 with tf.variable_scope('local3') as scope: # Move everything into depth so we can perform a single matrix multiply. dim = 1 for d in pool2.get_shape()[1:].as_list(): dim *= d reshape = tf.reshape(pool2, [FLAGS.batch_size, dim]) weights = _variable_with_weight_decay('weights', shape=[dim, 384], stddev=0.04, wd=0.004) biases = _variable_on_cpu('biases', [384], tf.constant_initializer(0.1)) local3 = tf.nn.relu(tf.matmul(reshape, weights) + biases, name=scope.name) _activation_summary(local3) # local4 with tf.variable_scope('local4') as scope: weights = _variable_with_weight_decay('weights', shape=[384, 192], stddev=0.04, wd=0.004) biases = _variable_on_cpu('biases', [192], tf.constant_initializer(0.1)) local4 = tf.nn.relu(tf.matmul(local3, weights) + biases, name=scope.name) _activation_summary(local4) DJGBSQZ

Slide 24

Slide 24 text

C-LIS CO., LTD.

Slide 25

Slide 25 text

C-LIS CO., LTD.

Slide 26

Slide 26 text

C-LIS CO., LTD. "This kind of layer is just like a convolutional layer, but without any weight-sharing. That is to say, a different set of ﬁlters is applied at every (x, y) location in the input image. Aside from that, it behaves exactly as a convolutional layer." IUUQTDPEFHPPHMFDPNQDVEBDPOWOFUXJLJ-BZFS1BSBNT-PDBMMZ DPOOFDUFE@MBZFS@XJUI@VOTIBSFE@XFJHIUT

Slide 27

Slide 27 text

C-LIS CO., LTD. Ͱɺ݁ہ -PDBMMZ$POOFDUFE-BZFSJTԿ ਫ਼౓্͕͕Δͷʁ ଎౓্͕͕Δͷʁ ͲΜͳ͍͍͜ͱ͕͋Δͷʁ

Slide 28

Slide 28 text

C-LIS CO., LTD. ંΓ৞Έ૚Λ૿΍ͯ͠Έͨ ˍύϥϝʔλʔΛௐ੔ # conv0  with tf.variable_scope('conv0') as scope:  kernel = _variable_with_weight_decay('weights', shape=[32, 32, 3, 32],  stddev=1e-4, wd=0.0)  conv = tf.nn.conv2d(images, kernel, [1, 1, 1, 1], padding='SAME')  biases = _variable_on_cpu('biases', [32], tf.constant_initializer(0.0))  bias = tf.nn.bias_add(conv, biases)  conv0 = tf.nn.relu(bias, name=scope.name)  _activation_summary(conv0)    # pool0  pool0 = tf.nn.max_pool(conv0, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1],  padding='SAME', name='pool0')  # norm0  norm0 = tf.nn.lrn(pool0, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75,  name='norm0')    # conv1  with tf.variable_scope('conv1') as scope:  kernel = _variable_with_weight_decay('weights', shape=[16, 16, 32, 64],  stddev=1e-4, wd=0.0)  conv = tf.nn.conv2d(norm0, kernel, [1, 1, 1, 1], padding='SAME')  biases = _variable_on_cpu('biases', [64], tf.constant_initializer(0.0))  bias = tf.nn.bias_add(conv, biases)  conv1 = tf.nn.relu(bias, name=scope.name)  _activation_summary(conv1)

Slide 29

Slide 29 text

C-LIS CO., LTD.

Slide 30

Slide 30 text

C-LIS CO., LTD. લճͱಉ͡σʔλͰݕূ ૚Λ૿΍͢ʢύϥϝʔλʔΛௐ੔ʣͱ࠷ॳ͔Β  ϩε཰͕෼ͷ̍ʹ௿Լͨ͠ɻ ୅ΘΓʹεςοϓʹ͔͔Δ͕࣌ؒ̏ഒʹͳͬͨɻ

Slide 31

Slide 31 text

σʔλͷॆ࣮

Slide 32

Slide 32 text

C-LIS CO., LTD. ࡞੒σʔλ಺༁ ؟ڸ່ͬը૾ຕ   ඇ؟ڸ່ͬը૾ຕ ؟ڸ່ͬ ඇ؟ڸ່ͬ ؟ڸ່ͬ ඇ؟ڸ່ͬ

Slide 33

Slide 33 text

C-LIS CO., LTD. ݕূσʔλͷ෼ྨΛखͰߦ͏ ҰຕֆͷΩϟϥΫλʔʹݶఆ͢Δ  ʢෳ਺ຕଘࡏ͠ʹ͍͘ֆΛબఆʣ D ࠜઇΕ͍

Slide 34

Slide 34 text

C-LIS CO., LTD. ࡞੒σʔλ಺༁ ܇࿅σʔλ ςετσʔλ

Slide 35

Slide 35 text

C-LIS CO., LTD. ܇࿅σʔλը૾ʹϊΠζΛՃ͑ͯ૿ྔ ΅͔͠ Ψ΢γΞϯϊΠζ ΠϯύϧεϊΠζ

Slide 36

Slide 36 text

Slide 37

Slide 37 text

C-LIS CO., LTD. ֶशͱݕূ ܇࿅σʔλ ςετσʔλ 1SFDJTJPO

Slide 38

Slide 38 text

C-LIS CO., LTD. ໰୊ൃੜ ΦϑΟεͷαʔόʔϚγϯͷ(16͕࢖͑ͳ͘ͳΔ $ python3 megane_co/cifar10_train.py I tensorflow/stream_executor/dso_loader.cc:105] successfully opened CUDA library libcublas.so locally I tensorflow/stream_executor/dso_loader.cc:105] successfully opened CUDA library libcudnn.so locally I tensorflow/stream_executor/dso_loader.cc:105] successfully opened CUDA library libcufft.so locally I tensorflow/stream_executor/dso_loader.cc:105] successfully opened CUDA library libcuda.so.1 locally I tensorflow/stream_executor/dso_loader.cc:105] successfully opened CUDA library libcurand.so locally Filling queue with 400 CIFAR images before starting to train. This will take a few minutes. E tensorflow/stream_executor/cuda/cuda_driver.cc:481] failed call to cuInit: CUDA_ERROR_NO_DEVICE

Slide 39

Slide 39 text

C-LIS CO., LTD. ଎౓ͷൺֱ $16 2016-04-13 10:40:51.188290: step 0, loss = 11.15 (65.8 examples/sec; 1.946 sec/batch) 2016-04-13 10:41:10.546296: step 10, loss = 10.67 (74.0 examples/sec; 1.731 sec/batch) 2016-04-13 10:41:29.646204: step 20, loss = 10.29 (74.0 examples/sec; 1.729 sec/batch) (16 2016-03-06 00:42:54.454566: step 90, loss = 10.39 (286.9 examples/sec; 0.446 sec/batch) 2016-03-06 00:42:59.533350: step 100, loss = 10.32 (290.1 examples/sec; 0.441 sec/batch) 2016-03-06 00:43:04.834940: step 110, loss = 10.27 (292.9 examples/sec; 0.437 sec/batch)

Slide 40

Slide 40 text

໰୊ൃੜ

Slide 41

Slide 41 text

(PPHMF$MPVE.BDIJOF-FBSOJOH

Slide 42

Slide 42 text

C-LIS CO., LTD. (PPHMF$MPVE.BDIJOF-FBSOJOH ($1/&95Ͱൃද͞Εͨ  Ϋϥ΢υϕʔεͷػցֶशϓϥοτϑΥʔϜ

Slide 43

Slide 43 text

C-LIS CO., LTD. (PPHMF$MPVE.BDIJOF-FBSOJOH 1SPKFDU/BNFNFHBOF@DP

Slide 44

Slide 44 text

C-LIS CO., LTD. (PPHMF$MPVE.BDIJOF-FBSOJOH ฦࣄɺདྷͯ·ͤΜͰͨ͠ʂ

Slide 45

Slide 45 text

͔͜͜Β͸$16ͷΈͰ͓ૹΓ͠·͢

Slide 46

Slide 46 text

C-LIS CO., LTD. ݕূʢ৽Ϟσϧʴ৽σʔληοτʣ ܇࿅σʔλ ςετσʔλ 1SFDJTJPO ʙ

Slide 47

Slide 47 text

C-LIS CO., LTD. 5SBJOJOHʢ৽Ϟσϧʴ৽σʔληοτʣ CBUDI 4UFQ

Slide 48

Slide 48 text

C-LIS CO., LTD. 5SBJOJOHʢچϞσϧʴچσʔληοτʣ CBUDI 4UFQ

Slide 49

Slide 49 text

C-LIS CO., LTD. ݕূʢچϞσϧʴ৽σʔληοτʣ ܇࿅σʔλ ςετσʔλ 1SFDJTJPO ʙ

Slide 50

Slide 50 text

C-LIS CO., LTD. 5SBJOJOHʢچϞσϧʴ৽σʔληοτʣ CBUDI 4UFQ

Slide 51

Slide 51 text

C-LIS CO., LTD. ՝୊ ֶशʢςετʣσʔλʹదԠ͢ΔϑΟϧλʔͷ  छྨͱڧ౓Λݕ౼ ൑ఆਫ਼౓Λ্͛Δɹˠɹ$*'"3ͷσʔλͰ֬ೝ  ʢάϥϑͷਂ૚Խʗύϥϝʔλʔௐ੔ʗޡ൑ఆύλʔϯΛׂΓग़͢ʣ

Slide 52

Slide 52 text

C-LIS CO., LTD. ࣍ճ༧ࠂ   $MPVE7JTJPO"QJ5FOTPS'MPXษڧձ (%(,PCF IUUQTHEHLPCFEPPSLFFQFSKQFWFOUT

Slide 53

Slide 53 text

C-LIS CO., LTD. 5FOTPS'MPXΛ࢖ͬͨػցֶश͜ͱ͸͡ΊCZ5PSV6&/0:"."  IUUQXXXTMJEFTIBSFOFU5PSV6FOPZBNBUFOTPSqPXHEH σΟʔϓϥʔχϯάͰ͓ͦদ͞Μͷ࿡ͭࢠ͸ݟ෼͚ΒΕΔͷ͔ʁɹʙ४උฤʙ  IUUQCPIFNJBIBUFOBCMPHDPNFOUSZ ػցֶशͷσʔληοτը૾ຕ਺Λ૿΍͢ํ๏  IUUQRJJUBDPNCPIFNJBOJUFNTDEGD  5FOTPS'MPX  IUUQTHJUIVCDPNUFOTPSqPXUFOTPSqPX ࢀߟ

Slide 54

Slide 54 text

C-LIS CO., LTD. C-LIS CO., LTD.

Slide 55

Slide 55 text

ຊࢿྉ͸ɺ༗ݶձࣾγʔϦεͷஶ࡞෺Ͱ͢ɻܝࡌ͞Ε͍ͯΔΠϥετ͸ɺಛʹهࡌ͕ͳ͍৔߹͸ࠜઇΕ͍ͷஶ࡞෺Ͱ͢ɻ ຊࢿྉͷશ෦ɺ·ͨ͸Ұ෦ʹ͍ͭͯɺஶ࡞ऀ͔ΒจॻʹΑΔڐ୚Λಘͣʹෳ੡͢Δ͜ͱ͸ې͡ΒΕ͍ͯ·͢ɻ 5IF"OESPJE4UVEJPJDPOJTSFQSPEVDFEPSNPEJpFEGSPNXPSLDSFBUFEBOETIBSFECZ(PPHMFBOEVTFE BDDPSEJOHUPUFSNTEFTDSJCFEJOUIF$SFBUJWF$PNNPOT"UUSJCVUJPO-JDFOTF ֤੡඼໊ɾϒϥϯυ໊ɺձ໊ࣾͳͲ͸ɺҰൠʹ֤ࣾͷ঎ඪ·ͨ͸ొ࿥঎ඪͰ͢ɻຊࢿྉதͰ͸ɺɺɺäΛׂѪͯ͠ ͍·͢ɻ 5IF"OESPJESPCPUJTSFQSPEVDFEPSNPEJpFEGSPNXPSLDSFBUFEBOETIBSFECZ(PPHMFBOEVTFEBDDPSEJOH UPUFSNTEFTDSJCFEJOUIF$SFBUJWF$PNNPOT"UUSJCVUJPO-JDFOTF