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

MLflow Tracking を用いた実験管理 / ayniy-with-mlflow

MLflow Tracking を用いた実験管理 / ayniy-with-mlflow

「atmaCup#5 振り返り会」での発表資料

- connpass: https://atma.connpass.com/event/178585/
- Competition GitHub: https://github.com/upura/atma-comp05
- Ayniy GitHub: https://github.com/upura/ayniy
- Ayniy documentation: https://upura.github.io/ayniy-docs/
- YouTube: https://youtu.be/b-YogiJA9XA

Shotaro Ishihara

June 07, 2020
Tweet

More Decks by Shotaro Ishihara

Other Decks in Programming

Transcript

  1. 使い⽅ from mlflow import log_metric, log_param, log_artifact mlflow.set_experiment(exp_name) mlflow.start_run(run_name=run_name) log_param('model_name',

    model_name) log_param('fe_name', fe_name) log_param('train_params', params) log_param('cv_strategy', cv) log_param('evaluation_metric', evaluation_metric) log_metric('cv_score', cv_score) log_param('fold_scores', fold_scores) log_param('cols_definition', cols_definition) log_param('description', description) mlflow.end_run() 11
  2. Ayniy Documentation GitHub Slide (Japanese) Sadriddin Ayni was a Tajik

    intellectual who wrote poetry, fiction, journalism, history and lexicography. He is regarded as Tajikistan's national poet and one of the most important writers in the country's history. https://uz.wikipedia.org/wiki/Sadriddin_Ayniy 15
  3. All You Need is YAML import yaml from sklearn.model_selection import

    StratifiedKFold from ayniy.preprocessing.runner import Tabular from ayniy.model.runner import Runner f = open('configs/fe000.yml', 'r+') fe_configs = yaml.load(f) g = open('configs/run000.yml', 'r+') run_configs = yaml.load(g) cv = StratifiedKFold(n_splits=5, shuffle=True, random_state=7) tabular = Tabular(fe_configs, cv) tabular.create() runner = Runner(run_configs, cv) runner.run_train_cv() runner.run_predict_cv() runner.submission() https://upura.github.io/ayniy-docs/quick_start_guide.html 16
  4. コンペ中の作業 . 現状のAyniyで不可能な場合は実装 . fe.yaml , run.yaml 書く . runner.py

    を実⾏ . mlflow uiでCVスコア⾒る . 良さげだったらsubmit 18