Softpia Japan Seminar 20190724

C-LIS CO., LTD.

ਓ޻஌ೳηϛφʔɹʙΫϥ΢υɾϞόΠϧɾΤοδʹ͓͚Δػցֶशʙ ɹιϑτϐΞδϟύϯ ػցֶशϞσϧΛར༻ͨ͠  ϞόΠϧΞϓϦ։ൃͷࣄྫ

༗ࢁܓೋʢ,FJKJ"3*:"."ʣ $-*4$0 -5% ઐ໳ɿ"OESPJEΞϓϦ։ൃ ػցֶशྺɿ໿̐೥ ஶॻɿ  ʰ"OESPJE4UVEJPͰ͸͡ΊΔ؆୯"OESPJEΞϓϦ։ൃʱʢٕज़ධ࿦ࣾʣ  ʰ5FOTPS'MPXΛ͸͡Ί·ͨ͠ʱʢΠϯϓϨε3%ʣ 
ڞஶʰ5FOTPS'MPX׆༻ΨΠυʱʢٕज़ධ࿦ࣾʣ Photo : Koji MORIGUCHI (MORIGCHOWDER)

5FOTPS'MPXൃදʢ೥݄ʣ

ػցֶशΛར༻ͨ͋͠Ε͜Ε IUUQLJWBOUJVNIBUFCMPKQFOUSZ TensorFlowͰΞχϝΏΔΏΓͷ੍࡞ձࣾΛࣝผ͢Δ IUUQCPIFNJBIBUFOBCMPHDPNFOUSZ σΟʔϓϥʔχϯάͰ͓ͦদ͞Μͷ࿡ͭࢠ͸ݟ෼͚ΒΕΔͷ͔ʁ IUUQCPIFNJBIBUFOBCMPHDPNFOUSZ IUUQDISJTUJOBIBUFOBCMPHDPNFOUSZ Deep
LearningͰϥϒϥΠϒʂΩϟϥΛࣝผ͢Δ

؟ڸ່ͬͷΠϥετΛ  Πϯλʔωοτ͔ΒࣗಈͰऩू͍ͨ͠ʂ

؟ڸ່ͬ൑ఆ

ධՁ༻αʔόʔ ܇࿅ɾֶश༻αʔόʔ܈ σʔληοτసૹ ʢTFRecordʣ ֶशࡁϞσϧऔಘ ը૾औಘ ը૾औಘ ϥϕϧ
෇͚ σʔληοτ؅ཧ αʔόʔ ը૾ऩू ϥϕϧ ෇͚ σʔληοτ  ؅ཧΞϓϦ playground.megane.ai ֶशࡁΈϞσϧ഑ஔ ը૾ૹ৴ ൑ఆ݁Ռ .FHBOF$P

ͳͥϞόΠϧΞϓϦʹ૊ΈࠐΉͷ͔ ϨΠςϯγ͕69ʹ༩͑ΔӨڹΛ࡟ݮ͢ΔͨΊ αʔόʔʹૹ৴͢Δʢ৺ཧతɾ๏తʣϋʔυϧ͕ߴ͍σʔλΛऔΓѻ͏ͨΊ

ϞόΠϧΞϓϦ΁ͷػցֶशϞσϧͷ૊ΈࠐΈ 5FOTPS'MPXGPS.PCJMF 5FOTPS'MPX-JUF .-,JUʢ'JSFCBTF.-,JUʣ

5FOTPS'MPX-JUF σόΠε্Ͱͷਪ࿦ΛՄೳʹ͢ΔϑϨʔϜϫʔΫɻ 5FOTPS'MPXͷαϒηοτϥΠϯλΠϜͰɺར༻Ͱ ͖ͳ͍ΦϖϨʔγϣϯ͕͋Δɻ "OESPJEJ04ΤοδσόΠεʹରԠɻ IUUQTXXXUFOTPSqPXPSHMJUF

.-,JUCFUBʢ'JSFCBTF.-,JUʣ (PPHMF͕ఏڙ͢ΔαʔϏεͰɺ(PPHMF͕࡞੒ͨ͠ػցֶश ϞσϧΛ"OESPJEJ04ΞϓϦʹ૊ΈࠐΜͩΓɺΦϦδφϧͷ ΞϓϦΛ഑৴Ͱ͖Δɻ w7JTJPO w-BOHVBHF w$VTUPN IUUQTEFWFMPQFSTHPPHMFDPNNMLJU

.-,JU7JTJPO #BSDPEF4DBO 'BDF%FUFDUJPO *NBHF-BCFMJOH -BOENBSL%FUFDUJPO 0CKFDU%FUFDUJPO5SBDLJOH 5FYU3FDPHOJUJPO IUUQTEFWFMPQFSTHPPHMFDPNNMLJU

.-,JU-BOHVBHF -BOHVBHF*EFOUJpDBUJPO 0O%FWJDF5SBOTMBUJPO 4NBSU3FQMZ IUUQTEFWFMPQFSTHPPHMFDPNNMLJU

.-,JU$VTUPN ඞཁʹԠͨ͡Ϟσϧͷμ΢ϯϩʔυ ϞσϧͷࣗಈΞοϓσʔτ Ϟσϧͷ"#ςετ Ϟσϧͷಈతʢ%ZOBNJDʣͳબ୒ IUUQTEFWFMPQFSTHPPHMFDPNNMLJU

ػցֶशϫʔΫϑϩʔ σʔληοτͷ੔උ ϞσϧͷֶशͱධՁ ϞσϧͷσϓϩΠ ਪ࿦

'PPE(BMMFSZ ୺຤ʹอଘ͞Ε͍ͯΔࣸਅͷ৯΂෺͚ͩΛදࣔʢ৯΂෺ Ҏ֎Λಁաͯ͠දࣔʣͯ͠Ұཡ͢ΔΞϓϦɻ IUUQTHJUIVCDPNLFJKJGPPE@HBMMFSZ@XJUI@UFOTPSqPX

def _export_graph(sess, input_tensors, output_tensors, output_dir): output_path = os.path.join(output_dir, 'model.tflite') converter
= TFLiteConverter.from_session(sess, input_tensors, output_tensors) # converter.post_training_quantize = True tflite_model = converter.convert() open(output_path, "wb").write(tflite_model) Ϟσϧͷग़ྗʢ5FOTPS'MPX-JUFʣ

val tfInference = Interpreter( model, options) val resizedImageBuffer = ByteBuffer
.allocateDirect(IMAGE_BYTES_LENGTH) .order(ByteOrder.nativeOrder()) val inputBuffer = ByteBuffer .allocateDirect(IMAGE_BYTES_LENGTH * 4) .order(ByteOrder.nativeOrder()) val resultBuffer = ByteBuffer .allocateDirect(4) .order(ByteOrder.nativeOrder()) Ϟσϧͷ࣮ߦ JPLFJKJGPPEHBMMFSZ*NBHF3FDPHOJ[FSLU

val scaledBitmap = Bitmap.createScaledBitmap(bitmap, IMAGE_WIDTH, IMAGE_HEIGHT, false) resizedImageBuffer.rewind() scaledBitmap.copyPixelsToBuffer(resizedImageBuffer) inputBuffer.rewind()
for (index in (0..IMAGE_BYTES_LENGTH - 1)) { inputBuffer.putFloat(resizedImageBuffer[index].toInt().and(0xFF).toFloat()) } inputBuffer.rewind() resultBuffer.rewind() tfInference.run(inputBuffer, resultBuffer) resultBuffer.rewind() val confidence = resultBuffer.getFloat() Ϟσϧͷ࣮ߦ JPLFJKJGPPEHBMMFSZ*NBHF3FDPHOJ[FSLU

Ϟσϧͷ࣮ߦ Input Output Interpreter tfInference inputBuffer resultBuffer લॲཧ
σίʔυ

ϞσϧΛ૊ΈࠐΉ্Ͱͷ՝୊ Ϟσϧͷେ͖͞ ࣮ߦ଎౓ ফඅϝϞϦ Ϟσϧͷอޢ ϢʔβʔϓϥΠόγʔ

.#Λ௒͑ΔϞσϧ IUUQTHJUIVCDPNLFJKJGPPE@HBMMFSZ@XJUI@UFOTPSqPXSFMFBTFTUBH

ϞσϧͷΞʔΩςΫνϟΛมߋ

ϞσϧʢύϥϝʔλʔʣͷྔࢠԽ CJUුಈখ਺఺਺ΛCJU੔਺ʹม׵ʢྔࢠԽʣ

ྔࢠԽͷબ୒ࢶ 1PTUUSBJOJOH2VBOUJ[BUJPO 2VBOUJ[BUJPOBXBSF5SBJOJOH

1PTUUSBJOJOHRVBOUJ[BUJPO def _export_graph(sess, input_tensors, output_tensors, output_dir): output_path =
os.path.join(output_dir, 'model.tflite') converter = TFLiteConverter.from_session(sess, input_tensors, output_tensors) converter.post_training_quantize = True tflite_model = converter.convert() open(output_path, "wb").write(tflite_model)

if model.QUANTIZATION: g = tf.get_default_graph() tf.contrib.quantize.create_training_graph(input_graph=g, quant_delay=2000000) IUUQTHJUIVCDPNUFOTPSqPXUFOTPSqPXUSFFSUFOTPSqPXDPOUSJCRVBOUJ[F
2VBOUJ[BUJPOBXBSFUSBJOJOH

ϞσϧΛྔࢠԽ

"OESPJE/FVSBM/FUXPSLT"1* ϞόΠϧ୺຤্ͰػցֶशͷܭࢉॲཧΛ࣮ߦ͢ΔͨΊʹઃܭ͞Εͨɻ IUUQTEFWFMPQFSBOESPJEDPNOELHVJEFTOFVSBMOFUXPSLT "OESPJEʢ"1*ϨϕϧʣҎ߱ͰରԠɻ IUUQTCMPHLFJKJJPUFOTPSqPXBEWFOU@DBMFOEBSIUNM ͍·//"1*ʢ5FOTPS'MPX-JUFʣ͸࢖͑Δͷ͔

ਪ࿦଎౓ͷൺֱ ػछ໊ NN API͋Γ NN APIͳ͠ Essential PH-1
556,323ns 185,372,624ns Pixel 2 450,807ns 187,395,464ns Pixel 3 477,489ns 129,994,563ns IUUQTHJUIVCDPNLFJKJGPPE@HBMMFSZ@XJUI@UFOTPSqPXSFMFBTFTUBHUqJUF@OOBQJ

͞·͟·ͳ੍໿ εΧϥʔͰͷԋࢉ͕Ͱ͖ͳ͍ɻʷOPSNBMJ[FE@JNBHFJNBHF@QI εϥΠε͕࢖͑ͳ͍ɻʷJNBHFJNBHF< > )ZQFSCPMJD5BOHFOU͕࢖͑ͳ͍ɻʷPVUQVUUGUBOI PVU@P⒎TFU
IUUQTXXXUFOTPSqPXPSHMJUFHVJEFPQT@DPNQBUJCJMJUZVOTVQQPSUFE@PQFSBUJPOT

ਫ਼౓͕ۃ୺ʹ௿Լ

max: 14.586626 min: -3.8083103 ϞσϧΛߏ੒͍ͯ͠Δύϥϝʔλʔͷ࠷େɾ࠷খ஋Λ֬ೝ

Ϟσϧઃܭͷ࣌఺Ͱʮଌఆͱௐ੔ʯ σʔληοτͷ੔උ ϞσϧͷֶशͱධՁ ϞσϧͷσϓϩΠ ਪ࿦ Ϟσϧͷઃܭ

ଌఆʹج͍ͮͨϞσϧͷௐ੔ Ϟσϧͷઃܭ Ϟσϧͷௐ੔ TensorFlow Lite΁ม׵ ʢྔࢠԽʣ αΠζɾ࣮ߦ଎౓  ଌఆ

ΞϓϦଆͰͷ଎౓վળ /BWJFSגࣜձࣾ φϏΤʣ ౦ژ౎จژ۠ຊڷຊڷ࢛ஸ໨Ϗϧ' σΟʔϓϥʔχϯάΛ༻͍ͨը૾ม׵ٕज़ͷ։ൃ͓Αͼιϑτ΢ΣΞͷఏڙ IUUQTXXXOBWJFSDP Ϟσϧ͸มߋͤͣɺ࣮ߦ଎౓Λ̏ʙ̐ഒʹߴ଎Խ

Ϟσϧͷอޢ ʰΞϓϦʹ૊ΈࠐΉʹ౰ͨͬͯɺϞσϧʢΞʔΩςΫνϟɾ ύϥϝʔλʔʣ͕֎෦ʹྲྀग़͠ͳ͍Α͏ʹอޢ͍ͨ͠ʱ

Ϟσϧͷܗࣜ 5FOTPS'MPX-JUFͷϞσϧ͸'MBU#V⒎FSTɻ 'MBU#V⒎FST͸ɺσʔλͷల։΍ύʔεΛͤͣετϨʔδ ্ͷσʔλʹϝϞϦΞυϨεΛϚοϓͯ͠ΞΫηε͢Δ ͜ͱͰɺϝϞϦͷϑοτϓϦϯτΛ࡟ݮ͢Δ  ʢʹ҉߸Խͨ͠'MBU#V⒎FSTͷϑΝΠϧ͸ɺಁաతʹಡΈ ࠐΉ͜ͱ͕Ͱ͖ͳ͍ʣɻ

෮߸Խͯ͠Ұ౓ϩʔΧϧϑΝΠϧʹอଘ assets app local ϑΝΠϧಡΈࠐΈ Interpreter ෮߸Խॲཧ

ϢʔβʔϓϥΠόγʔ ػցֶशͷϞσϧΛܧଓతʹվળ͢ΔͨΊʹ͸ɺσʔλ͕ඞཁɻ Ϣʔβʔͷ͍ΔΞϓϦͷ৔߹ɺϢʔβʔ͔ΒσʔλΛఏڙʢૹ৴ɾอଘʣ͢Δ ߹ҙΛಘΔඞཁ͕͋Δɻ ϓϥΠόγʔσʔλͷऔΓѻ͏ମ੍Λ੔͑Δඞཁ͕͋Δɻσʔλͷ಺༰ʹΑͬ ͯ͸๏཯ʹΑΔن੍΋͋Δɻ

'FEFSBUFE-FBSOJOH ͢΂ͯͷσʔληοτΛαʔόʔ্Ͱॲཧ͢ΔʮूதτϨʔχϯάʯͱ͸ҟͳ ΓɺҰͭҰͭͷ୺຤্ͰϞσϧͷֶशͷલஈॲཧΛߦ্ͬͨͰɺαʔόʔʹૹ ৴ͯ͠ฏۉԽ͢Δɻ IUUQTEFWFMPQFSTKQHPPHMFCMPHDPNGFEFSBUFEMFBSOJOHDPMMBCPSBUJWFIUNM

C-LIS CO., LTD. ຊࢿྉ͸ɺ༗ݶձࣾγʔϦεͷஶ࡞෺Ͱ͢ɻຊࢿྉͷશ෦ɺ·ͨ͸Ұ෦ʹ͍ͭͯɺஶ࡞ऀ͔ΒจॻʹΑΔڐ୚Λಘͣʹෳ੡͢Δ͜ͱ͸ې͡ΒΕ͍ͯ·͢ɻ 5IF"OESPJE4UVEJPJDPOJTSFQSPEVDFEPSNPEJpFEGSPNXPSLDSFBUFEBOETIBSFECZ(PPHMFBOEVTFEBDDPSEJOHUPUFSNTEFTDSJCFEJOUIF$SFBUJWF$PNNPOT"UUSJCVUJPO-JDFOTF ֤੡඼໊ɾϒϥϯυ໊ɺձ໊ࣾͳͲ͸ɺҰൠʹ֤ࣾͷ঎ඪ·ͨ͸ొ࿥঎ඪͰ͢ɻຊࢿྉதͰ͸ɺɺɺäΛׂѪ͍ͯ͠·͢ɻ 5IF"OESPJESPCPUJTSFQSPEVDFEPSNPEJpFEGSPNXPSLDSFBUFEBOETIBSFECZ(PPHMFBOEVTFEBDDPSEJOHUPUFSNTEFTDSJCFEJOUIF$SFBUJWF$PNNPOT"UUSJCVUJPO-JDFOTF

Softpia Japan Seminar 20190724

Softpia Japan Seminar 20190724

More Decks by ARIYAMA Keiji

Other Decks in Technology

Featured

Transcript