YOLOで遊ぶ / pyml-niigata-20181013

Python機械学習勉強会 in 新潟 Restart #2 で発表した資料

Transcript

1. YOLOͰ༡Ϳ Pythonػցֶशษڧձ in ৽ׁ 2018/10/13 @kasacchiful
2. ࣗݾ঺հ • ּݪ ޺ (@kasacchiful) • ৽ׁࢢࡏॅ • Ruby͕͓ؾʹೖΓ • JaSST৽ׁ࣮ߦҕһ • JAWS-UG৽ׁࢧ෦ • ୈ1ճษڧձΛ11݄24೔(౔)ʹ։࠵ܾఆʂ
3. એ఻
4. ઌ݄NDSͰ͜Μͳ࿩͠·ͨ͠ • ࢿྉ: https://speakerdeck.com/kasacchiful/nds57 • σϞιʔε: https://github.com/kasacchiful/nds57-sample/blob/master/ nds57_vgg16_sample.ipynb
5. OSC 2018 Niigata Ͱ஻Γ·͢ IUUQTXXXPTQOKQPTDOJJHBUBNPEVMFTFHVJEFFWFOUQIQ FJE
6. JAWS-UG ৽ׁࢧ෦ઃཱ • 11݄24೔(౔)ʹୈ1ճษڧձΛ։࠵ʂ IUUQTKBXTVHOJJHBUBDPOOQBTTDPN
7. ຊ୊
8. ໨࣍ • YOLOͬͯԿʁ෺ମݕग़ͬͯԿʁ • ·ͣ͸ಈ͔ͯ͠ΈΑ͏ • ΧελϜֶशͤͯ͞ΈΑ͏
9. ໨࣍ • YOLOͬͯԿʁ෺ମݕग़ͬͯԿʁ • ·ͣ͸ಈ͔ͯ͠ΈΑ͏ • ΧελϜֶशͤͯ͞ΈΑ͏
10. YOLOͬͯԿʁ • ෺ମݕग़(Object Detection)ͷ࣮૷ͷ1ͭ • ଞʹ΋ɺFaster R-CNN΍SSDͳͲ͕͋Δ • DarknetϑϨʔϜϫʔΫ্Ͱߏங (C࣮૷) → Keras࣮૷ͳͲ΋͋Δ • ࠓ೥य़ʹv3ެ։ ( https://arxiv.org/abs/1804.02767 ) • v2͕19૚ → v3͕53૚ • ݕग़଎౓͕एׯ஗͘ͳ͕ͬͨɺݕग़ਫ਼౓͕େ෯Ξοϓ • “You Only Live Once" Λ΋ͬͯ͡ɺ”You Only Look Once”
11. ෺ମݕ஌ͬͯԿʁ • ͋Δը૾ͷத͔ΒఆΊΒΕͨ෺ମͷΧςΰϦ(Ϋϥ ε)ͱͦͷҐஔΛݕग़ • ڭࢣ͋Γֶशͷ෼ྨʹ૬౰ • ಈըʹ΋ద༻Մೳ • ෺ମͷҐஔ͸ɺ෺ମͷྖҬΛۣܗͰғͬͨΓɺྖ ҬΛॴఆͷ৭ͰృΓͭͿͨ͠Γ
12. ྫ͑͹͜Μͳײ͡ ./darknet detect cfg/yolov3.cfg yolov3.weights data/dog.jpg
13. ެࣜͷσϞΛݟͯΈΑ͏ IUUQTQKSFEEJFDPNEBSLOFUZPMP
14. ෺ମݕ஌ͷྺ࢙ ྖҬީิͷબఆ ಛ௃ྔͷநग़ ػցֶशͰ෼ྨ ैདྷ ྖҬީิͷબఆ σΟʔϓϥʔχϯά R-CNN Fast R-CNN Faster R-CNN YOLO SSD σΟʔϓϥʔχϯά
15. ෺ମݕ஌ͷࣄྫ • LIXIL (ABEJA SIX 2018Ͱͷߨԋ) • γεςϜΩονϯΛ࢖ͬͨਓؒͷߦಈ෼ੳʹ෺ମݕ஌Λ׆༻ • ෺ମݕ஌ͷϩά͔Βɺ࢖͍΍͍͢γεςϜΩονϯͷ։ൃʹͭͳ͍͛ͯΔ • ෺ମݕ஌ʹ͸SSD͕࢖ΘΕ͍ͯΔ • ࢀߟ: • https://six2018.abejainc.com/reports/ • https://logmi.jp/279849
16. ໨࣍ • YOLOͬͯԿʁ෺ମݕग़ͬͯԿʁ • ·ͣ͸ಈ͔ͯ͠ΈΑ͏ • ΧελϜֶशͤͯ͞ΈΑ͏
17. darknetΠϯετʔϧ $ git clone https://github.com/pjreddie/darknet $ cd darknet $ make
18. ඞཁʹԠͯ͡Makefile΋ฤू GPU=0 #=> NVIDIAͷGPU͕࢖͑ͯɺCUDA͕ΠϯετʔϧࡁͳΒ1ʹ CUDNN=0 OPENCV=0 #=> OPENCV͕ΠϯετʔϧࡁͳΒ1ʹ OPENMP=0 DEBUG=0
19. Makefileमਖ਼ͨ͠Β࠶౓make $ make
20. ࢼ͠ʹ࣮ߦ $ ./darknet detect cfg/yolov3.cfg yolov3.weights data/dog.jpg
21. ͜Μͳײ͡ʹग़ྗ͞ΕΔ ݁Ռ͸predictions.jpgʹग़ྗ
22. σϞ
23. ໨࣍ • YOLOͬͯԿʁ෺ମݕग़ͬͯԿʁ • ·ͣ͸ಈ͔ͯ͠ΈΑ͏ • ΧελϜֶशͤͯ͞ΈΑ͏
24. GPU͕ແ͍ͱπϥΠͷͰ • ֶश؀ڥ: • Amazon EC2 p2.xlarge Πϯελϯε • GPU: NVIDIA K80 (Google Colabͱಉ͡) • Deep Learning Base AMI (Ubuntu) Ver.10.0 • CUDAΠϯετʔϧࡁ
25. darknetΠϯετʔϧ $ git clone https://github.com/pjreddie/darknet $ cd darknet
26. Makefileฤू GPU=1 CUDNN=0 OPENCV=0 OPENMP=0 DEBUG=0
27. make $ make
28. ֶशͷྲྀΕ 1. ը૾ΛूΊΔ 2. Ξϊςʔγϣϯ͢Δ 3. ը૾Λਫ૿͢͠Δʢࠓճ͸লུʣ 4. ֶश͢Δ
29. ࠓճͷֶश಺༰ • ΠλϦΞϯͷը૾ͷத͔ΒɺΠλϦΞϯͷྖҬΛ ݕग़ • class਺: 1ͭ (ΠλϦΞϯ͚ͩ) • ը૾͸72ຕ • ຊ౰͸΋ͬͱूΊΔඞཁ͕͋Δ͚Ͳɺࠓճ͸গ ྔͷσʔλͰ΍ͬͯΈΔ
30. 1. ը૾ΛूΊΔ • ΠλϦΞϯͷը૾ΛूΊΔ • Bing Search API v7Λ࢖ͬͯɺը૾Λ700ຕ͘ Β͍ूΊͯɺͦͷத͔Β࢖͑ͦ͏ͳը૾Λ72 ຕબͿ
31. 2. Ξϊςʔγϣϯ͢Δ • ը૾ʹҐஔ৘ใΛՃ͑ͯɺڭࢣσʔλΛ࡞੒͢Δ • ࠓճ͸ɺVoTTΛ࢖༻ • https://github.com/Microsoft/VoTT • VoTT͸YOLOͷܗࣜͰग़ྗͯ͘͠ΕΔͷͰศར • Ξϊςʔγϣϯ࡞ۀ͸ɺσʔλ͕ଟ͚Ε͹ଟ͍΄Ͳɺख ؒͱ͕͔͔࣌ؒΔͷͰେม͕ͩɺػցֶशͷΩϞʹͳΔ
32. 3. ը૾Λਫ૿͢͠Δ • ࠓճ͸໘౗ͩͬͨͷͰ΍͍ͬͯͳ͍ • ݩͷը૾͔ΒϥϯμϜʹͣΒͨ͠Γɺճస͞ ͤͨΓͱɺมܗͤͨ͞ը૾Λ࡞੒ͯ͠ɺը૾ Λ૿΍͢
33. 4. ֶश͢Δ • Ξϊςʔγϣϯͨ͠ڭࢣσʔλΛdarknetͷ ϑΥϧμʹίϐʔ͠ɺֶशΛ࣮ߦ • ࠓճ͸VoTT͕ग़ྗͨ͠YOLOͷcfgΛ࢖༻
34. ڭࢣσʔλΛల։ $ unzip italian_output.zip $ cd darknet/cfg $ mv ~/italian_output ./italian $ cd italian $ ls data yolo-obj.cfg $ ls data obj obj.data obj.names test.txt train.txt $ ls data/obj 000001.jpg 000001.txt 000002.jpg 000002.txt ...
35. ֶश $ cd ~/darknet $ ./darknet detector train ./cfg/italian/data/obj.data \ ./cfg/italian/yolo-obj.cfg
36. ਪ࿦ $ cd ~/darknet $ ./darknet detector test ./cfg/italian/data/obj.data \ ./cfg/italian/yolo-obj.cfg \ ./backup/yolo-obj_500.weights \ ./italian1.jpg
37. ͜Μͳײ͡ʹग़ྗ͞ΕΔ ݁Ռ͸predictions.jpgʹग़ྗ
38. σϞ
39. ·ͱΊ • ෺ମݕ஌΋ɺ਺ʑͷ࣮૷͕ొ৔͍ͯ͠ΔͷͰɺൺ ֱత؆୯ʹࢼͤ·͢ • ֶश༻σʔλͷऩूͱΞϊςʔγϣϯ͕͠ΜͲ͍ • YOLO΍SSDΛࢼ͢ྫ͸ൺֱతଟ͍ • Keras౳Ͱͷ࣮૷΋͋ΔͷͰɺࢼͯ͠Έ͍ͯͩ͘͞