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

Mercari ML Meetup

Mercari ML Meetup

How Mercari achieves machine learning engineering and operation.

shibuiwilliam

March 15, 2022
Tweet

More Decks by shibuiwilliam

Other Decks in Technology

Transcript

  1. Agenda ❖ Topic: 違反検知 ➢ Mercari ML Secret No. 1

    ~Modelling~ ➢ Mercari ML Secret No. 2 ~SysML~
  2. $whoami • 澁井 雄介 : @shibuiwilliam or @cvusk ◦ 株式会社メルカリ

    機械学習エンジニア • 経歴 ◦ 文学修士(イギリス旅行史) →富士通→IBM→SAS→パナソニック ◦ 株式会社メルカリ ▪ 機械学習によるサービス改善 ▪ 機械学習プラットフォーム開発 cat : 0.55 dog: 0.45 文系 : 0.60 理系 : 0.40 Object Detection
  3. 特定カテゴリ二値分類による違反出品検出 • 特定のカテゴリに特化した違反出品検出器を二値分類で実装 安全に! Binary classifier for a certain category

    Binary classifier for a certain category Binary classifier for a certain category Binary classifier for a certain category 0 0 1 0 Different model per objective Different training process per model ・・・ !!
  4. • Machine learning in cyclic workflow 目指すは機械学習をE2Eでカバーするプラットフォーム 素早く! Deploy Monitoring

    Evaluation Hyper parameter optimization Re-Training Training Hyper parameter optimization import tensorflow as tf def neural_net(x_dict): x = x_dict['images'] layer_1 = tf.layers.dense(x, n_hidden_1) layer_2 = tf.layers.dense(layer_1, n_hidden_2) out_layer = tf.layers.dense(layer_2, num_classes) return out_layer ...
  5. 日中はコーディング→学習→デプロイを効率化 素早く! Deploy Training Hyper parameter optimization import tensorflow as

    tf def neural_net(x_dict): x = x_dict['images'] layer_1 = tf.layers.dense(x, n_hidden_1) layer_2 = tf.layers.dense(layer_1, n_hidden_2) out_layer = tf.layers.dense(layer_2, num_classes) return out_layer ...
  6. Search Space Katib Grid Search Random Search Bayesian Optimization Hyperband

    Optimization Genetic Algorithm(coming soon) • Optimization distributed on Kubernetes cluster to search for the best ML parameters • Automatically distributed over cluster • Efficiently searches for the optimized parameters • Early stopping once reached target https://github.com/kubeflow/katib