Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Sansanデータ分析コンテスト参加者 marugari様

sansan_rd
November 01, 2016

Sansanデータ分析コンテスト参加者 marugari様

『人工知能は名刺をどこまで解読できるのか』
2016年8月8日~9月30日まで行われた、データ分析コンテスト
一番最初に手法をアップロードして頂いたmarugari様のアプローチのライトニングトーク

http://jp.corp-sansan.com/lp/data-sientist-c.html

sansan_rd

November 01, 2016
Tweet

More Decks by sansan_rd

Other Decks in Programming

Transcript

  1. Ξϓϩʔν • Private Score: 0.00234 • Deep LearningʢCVͳ͠ͷ1ຊʣ • ResNetϕʔεͷωοτϫʔΫ


    ࠷ޙͷίϯϘϦϡʔγϣϯͷϑΟϧλ਺͸128 • ը૾ͷલॲཧ
 ը૾αΠζ͸ (32, 160)
 पԑ෦ͷ৭ͷೱ͞ͰωΨϙδ൓స • Softmaxग़ྗΛ9ͭฒ΂ͨϚϧνλεΫֶश • ༧ଌͷௐ੔͸ΫϥεؒɺϑΝΠϧ಺ͲͪΒ΋͍ͯ͠ͳ͍ 3
  2. ڙཆҰ Max-Pooling • ෳ਺ͷϥϕϧ͕෇͍͍ͯΔ΋ͷ͸Max-Poolingͷํ͕ྑ͞ ͦ͏ ➡ Avg-PoolingͱMax-PoolingΛConcat ➡ ਫ਼౓վળͤͣ ➡

    Avg-PoolingΛγϣʔτΧοτͱͯ͠ɺม׵Λ͔͚ͨMax- Poolingͱ࿨ΛͱΔ ➡ ਫ਼౓վળͤͣ • ւ֎ͷ಄ͷྑ͍ਓ͕ߟ͑ͨ͜ͱͳͷͰɺ࣌ؒΛ࢖ͬͯݕ౼ ͢Δ͜ͱ͸ͳ͔ͬͨ 4
  3. Out[3]: data Pooling avg, 8x8/1 Flatten Pooling max, 8x8/1 Flatten

    Activation relu _Plus SoftmaxOutput In [1]: import mxnet as mx In [2]: data = mx.sym.Variable('data') net = data pavg = mx.sym.Pooling(net, pool_type='avg', kernel=(8,8)) pmax = mx.sym.Pooling(net, pool_type='max', kernel=(8,8)) pavg = mx.sym.Flatten(pavg) pmax = mx.sym.Flatten(pmax) pmax = mx.sym.Activation(pmax, act_type='relu') net = pavg + pmax out = mx.sym.SoftmaxOutput(net)
  4. ڙཆೋ Ϋϥεؒͷ֬཰ௐ੔ • ImageNetͱൺ΂ͯϥϕϧͷछྨ͕গͳ͍ͷͰɺ࠷௿1ͭͷ ϥϕϧ͕෇༩͞ΕΔΑ͏ʹ૊ΜͰΈΑ͏ ➡ i ൪໨ͷSoftmaxग़ྗʹ i-1 ൪໨·Ͱͷର਺Φοζൺͷ૯࿨

    Λ࢖͏ ➡ ਫ਼౓վળͤͣ • Ϋϥεؒͷ֬཰ௐ੔͘Β͍͸Deep Learningʹ·͔ͤͯྑ ͍ • ޻෉Λ͢ΔͳΒɺಉҰϑΝΠϧ಺Ͱͷௐ੔ͷํͩͬͨ 6
  5. Out[5]: data FullyConnected 1 FullyConnected 1 Concat SoftmaxOutput Concat FullyConnected

    1 FullyConnected 1 Concat SoftmaxOutput _Plus Concat FullyConnected 1 FullyConnected 1 Concat SoftmaxOutput In [4]: data = mx.sym.Variable('data') net = data fco = [] smo = [] cum = None for ii in range(3): if ii > 0: if cum is None: cum = pos else: cum = cum + pos net = mx.sym.Concat(*[net, cum]) neg = mx.sym.FullyConnected(net, num_hidden=1) pos = mx.sym.FullyConnected(net, num_hidden=1) fco.append(mx.sym.Concat(*[neg, pos])) smo.append(mx.sym.SoftmaxOutput(fco[ii])) out = mx.sym.Group(smo)
  6. ڙཆࡾ ϨΠΞ΢τ৘ใ • ྖҬͷେ͖͞ΛݟΕ͹෼͔Δ΋ͷΛޡೝ͍ࣝͯ͠Δ ➡ ྖҬͷେ͖͞ɺॎԣൺΛωοτϫʔΫʹྲྀ͠ࠐΉ ➡ ਫ਼౓վળͤͣ • ϨΠΞ΢τ৘ใͷΈͷDeep

    LearningͰ͸MAE: 0.08ఔ౓
 XGBoostͷ0.06ʹൺ΂ͯྼΔ • Deep LearningʹΑΔ༧ଌ஋ͱϨΠΞ΢τ৘ใͰStacking ͨ͠΋ͷ͸աֶश͕ࠅ͔ͬͨ • ཧ۶͔Β͢Ε͹্ख͍͘͘͸ͣͳͷͰௐࠪܧଓ 8