Caffeでお手軽本格ディープラーニングiOSアプリ

@potatotips #22
#DeepLearning #MachineLearning

Transcript

1. $B⒎FͰ͓खܰຊ֨
   σΟʔϓϥʔχϯά
   J04ΞϓϦ
   5",6:" !OPSBEBJLP
   
   QPUBUPUJQT
   

   View Slide

2. দࢁ୓໼
   w !OPSBEBJLP
   w ϑϦʔϥϯε೥໨ ݩ:BIPP
   
   w J04ΞϓϦ ΢ΣϒΞϓϦͳͲΛ੍࡞
   w ػցֶशʹڵຯ͋Γ
   w ֆ΋ඳ͖·͢
   2
   

   View Slide

3. ΊΜ͖͖
   ໙ར͖
   3
   

   View Slide

4. 4
   ໙ར͖
   ࣸਅʹج͍ͮͯϥʔϝϯ԰Λਪન͢ΔΞϓϦ
   ೖྗ
   

   View Slide

5. 5
   σΟʔϓϥʔχϯά
   ͷٕज़Λ࢖༻
   ʴ ʹ
   

   View Slide

6. ը૾ೝࣝʹڧ͍
   ػցֶशΞϧΰϦζϜ
   6
   σΟʔϓϥʔχϯάͱ͸
   

   View Slide

7. w ೴ͷਆܦߏ଄Λ໛฿ͨ͠χϡʔϥϧωοτϫʔΫͷҰछ
   w େྔͷσʔλ͔Β෺ମͷಛ௃Λࣗಈతʹֶश
   ‣ ͜Ε·Ͱ͸ಛ௃ͷநग़ํ๏͸ਓ͕͕ؒΜ͹ͬͯ༻ҙ͍ͯͨ͠
   7
   
   

   View Slide

8. ࡞Ζ͏
   σΟʔϓϥʔχϯάΞϓϦ
   ୭Ͱ΋؆୯ʹ࡞ΕΔํ๏Λ͝঺հ͠·͢
   8
   

   View Slide

9. $B⒎F
   σΟʔϓϥʔχϯά༻
   ϑϨʔϜϫʔΫ
   w IUUQDB⒎FCFSLFMFZWJTJPOPSH
   w (16ԋࢉ $6%"
   Ͱߴ଎ʹֶशͰ͖Δ
   w ͙͢ʹࢼͤΔֶशࡁΈϞσϧ΋͋Δ
   w .BD049ରԠ
   9
   

   View Slide

10. Caffe for
    J04্Ͱಈ͘$B⒎F
    w IUUQTHJUIVCDPNBMFQIDB⒎F
    w $B⒎FͷGPSL
    w J04্Ͱࣝผॲཧ͕࣮༻ʹ଱͑Δ଎౓Ͱಈ͔ͤΔ
    ‣ J1IPOFTͰʙඵ
    w αʔό͍ΒͣͰ࢖͑Δ
    w ͨͩ͠9$PEF͸·ͩඇରԠ
    10
    

    View Slide

11. $B⒎FGPSJ04
    αϯϓϧ࡞Γ·ͨ͠
    w IUUQTHJUIVCDPNOPSBEBJLP
    DB⒎FJPTTBNQMF
    w ୯७ͳ෺ମೝࣝ
    w #-7$$B⒎F/FU.PEFMΛ࢖༻
    11
    

    View Slide

12. demo
    

    View Slide

13. ࢖༻͢Δσʔλ
    w MBCFMTUYUࣝผ݁ՌΛ໊લʹม׵͢ΔͨΊͷҰཡ
    w EFQMPZQSPUPUYUωοτϫʔΫఆٛ
    w NFBOCJOBSZQSPUPฏۉը૾
    w CWMD@SFGFSFODF@DB⒎FOFUDB⒎FNPEFMֶशࡁΈσʔλ
    13
    

    View Slide

14. ॲཧͷྲྀΕ
    ࣝผର৅ͷը૾ͷಡΈࠐΈ
    w ૾ͷը૾
    $MBTTJpFSΫϥεͷॳظԽ
    w ͭͷϞσϧσʔλ΁ͷϑΝΠϧύεΛࢦఆ
    $MBTTJpFSͷ࣮ߦ
    w ը૾Λࢦఆͯ݁͠ՌΛऔಘ
    ࣝผ݁Ռͷग़ྗ
    14
    

    View Slide

15. UIImage* image = [UIImage imageNamed:@"sample.jpg"];
    cv::Mat src_img, img;
    UIImageToMat(image, src_img);
    cv::cvtColor(src_img, img, CV_RGBA2BGRA);
    ը૾ͷಡΈࠐΈ
    w 6**NBHFΛಡΈࠐΈ
    w DW.BUܗࣜʹม׵
    w Χϥʔ഑ྻΛ3(#"͔Β#(3"ʹม׵
    

    View Slide

16. // ϑΝΠϧύεΛstringܕʹม׵
    string model_file_str = std::string([model_file UTF8String]);
    string label_file_str = std::string([label_file UTF8String]);
    string trained_file_str = std::string([trained_file UTF8String]);
    string mean_file_str = std::string([mean_file UTF8String]);
    Classifier classifier = Classifier(model_file_str, trained_file_str,
    mean_file_str, label_file_str);
    $MBTTJpFSͷॳظԽ
    w ϞσϧఆٛɺϥϕϧɺֶशࡁΈϞσϧɺฏۉը૾ͷύεΛऔಘ
    w ֤ϑΝΠϧύεΛTUETUSJOHʹม׵
    w $MBTTJpFSͷΠϯελϯεΛ࡞੒
    

    View Slide

17. // ࣝผͷ࣮ߦ
    std::vector result = classifier.Classify(img);
    $MBTTJpFSͷ࣮ߦ
    w ը૾Λࢦఆ͢Δ͚ͩʂ
    

    View Slide

18. for (std::vector::iterator it = result.begin();
    it != result.end();
    ++it) {
    NSString* label = [NSString stringWithUTF8String:it->first.c_str()];
    NSNumber* probability = [NSNumber numberWithFloat:it->second];
    NSLog(@"label: %@, prob: %@", label, probability);
    }
    ࣝผ݁Ռͷग़ྗ
    w TUEWFDUPSܗࣜͰෳ਺ͷࣝผީิ͕ಘΒΕΔ
    w JUFSBUPSͰճ֤ͯ͠ީิΛऔಘ
    w JUpSTUϥϕϧɺJUTFDPOE֬཰
    

    View Slide

19. ·ͱΊ
    w $B⒎FΛ࢖͑͹ΦϦδφϧͷֶशϞσϧ͕࡞ΕΔ
    w $B⒎FGPSJ04ͳΒαʔό͍ΒͣͰࣝผॲཧ͕ग़དྷΔ
    w αϯϓϧϓϩδΣΫτͷ͝঺հ
    w ΦϦδφϧͷֶशϞσϧͰΞϓϦΛ࡞Ζ͏ʂ
    19
    

    View Slide

20. ͝ਗ਼ௌ͋Γ͕ͱ͏͍͟͝·ͨ͠
    20
    

    View Slide