Skip to content

ajdajd/apes_together_strong_

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

38 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

apes_together_strong

MLHackFest 2019 Repo

How to use

  1. Create your own conda or virtual environment.
  2. Run pip install -r requirements.txt.
  3. Add Kaggle API credentials. See instructions here.
  4. get wild

Structure

test folder contains all experiment notebooks from Kaggle competitions with the pipeline created in utils/models.

src folder is reserved for competition purposes, need to copy tests/utils in the folder to use the modules.

data folder is used for downloading dataset in Kaggle using KAGGLE API.

src and test folders must have their own submission folder when generating predictions in competition/challenges for easier tracking of the submission files.

Notes:

  • You must join the competition first before you are able to download data thru Kaggle API.

  • You must join the Kaggle competition first before you are able to download its dataset thru the Kaggle API.

  • When using CatBoostCV specify in init the obj i.e regression or binary to use the correct algorithm to use. LGBMCV works for both, no need to specify the obj.

  • When dealing with Regression Problem transform the target to np.log for easier training then transform again back to original state using np.exp. if negative values are encountered in the prediction values, just use pd.Series.clip function to clip the values to its min, and max.

  • When dealing with Classification Problem with large dataset 500K ~ 1M+ Instances, consider to downsample the majority class for easier feedback loop iteration, don't use SMOTE or other stuff, that doesn't work!

  • When everything doesn't work, use target encoding under utils/cat_encoding.py that will automagically make the model better, but ofcourse make sure you have solid CV and DO NOT OVERFIT

  • RandomForest is the only model that Rafael Trusts in sklearn that can be used in competitions, unless you ensemble/stack predictions, use LogisticRegression.

  • Submission files makes it easier for us to check if our Cross Validation correlates with the Public Leaderboard in Kaggle by formatting the name of the submission file using {model_used}_{challenge}_(my_cv_score}.csv

About

Mirror for Team ATS's repo for ML Hackfest 2019

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published