Shopee-IET-ML-Competition (1st Place)

The aim of the project is to build a model to classify images to 18 categories. The training/test sets are images provided by Shopee, which are classified into 18 categories.

Metrics: Accuracy

Getting Started

These instructions will get you a copy of the project up and running on your machine for development and testing purposes.

Prerequisites

First, you have to download the data from here and store them under data folder in the folder of this repository. Furthermore, The script has been only tested using Python 3.6 and for the libraries needed to run the script, you can check requirements.txt.

Notebooks

The notebook contains the analysis done before and after the training.

• EDA.ipynb: Exploratory data analysis, the analysis done before training.
• Model Analysis.ipynb: Analysis of the error of the models and correlation between models, done after the training.

Scripts

Below are the details of the scripts:

• split_train_val.py: to split the train images to n folds and store them in a directory.
• training.py: to train the model
• generate_df.py: to generate mapTest.csv which contain a sorted filepath to the test data.
• generate_test.py: to generate validation and test predictions from the models trained.
• average.py: to average the predictions of validation or test.
• voting.py: to do a weighted majority voting

Training

First, you will have to run split_train_val.py to create a locally stored splits train data.

python split_train_val.py --train_dir ../data/train --output_dir ../data/train_val_v1 --n_splits 7

Next, you can immediately do the training by running the script training.py. All the configs can be easily configured from inside the script.

P.S. It is recommended to run the script from inside the folder to avoid RelativePathError, or equivalently do cd src.

Result

After fine-tuning several pretrained models combined with weighted majorith voting, we achieved 0.86956 in the private leaderboard.

Authors

Team: h1n4

• Hardian Lawi - hardianlawi
• Kevin Luvian - kevinluvian
• Timothy Gabriel Kurniawan - timothygk

Acknowledgments

Extending my gratitude to NTU-IET and Shopee who made this competition possible.