Shopee - Price Match Guarantee

Introduction

The aim of this competition is to predict which items are the same products. This is source code/solution to get the Silver Medal in Shopee - Price Match Guarantee.

Alt text

Disclaimer

This is our work, we DO NOT represent any organization
There’s no reproducibility guarantee for notebook which uses GPU and TPU
Dataset and generated dataset falls under Shopee Terms and Conditions which can be seen on Kaggle Datasets

Environment List

Some of the notebook are run on different environment. Use Kaggle TPU on training the model and GPU while submitting the code to get public/private score.

Environment Name	Description
Kaggle CPU	2C/4T CPU, 16GB RAM
Kaggle GPU	2C/4T CPU, 16GB RAM, Nvidia Tesla P100
Kaggle TPU	2C/4T CPU, 16GB RAM, TPU v3-8

Notebook Description

Filename	Link to Kaggle Kernel	Environment	Description
eff3-512.ipynb	https://www.kaggle.com/mfalfafa/shopee-effb3-512-all-training	Kaggle TPU	Create multimodal model with EfficientNet B3 pretrained model and image size 512x512 pixel
effb5-512.ipynb	https://www.kaggle.com/mfalfafa/shopee-effb5-512-v2-all-training	Kaggle TPU	Create multimodal model with EfficientNet B5 pretrained model using image size 512x512 pixel with MLP
nf-net-f0.ipynb	https://www.kaggle.com/mfalfafa/shopee-nf-net-all-training	Kaggle TPU	Create multimodal model with NFNet-F0 pretrained model using image size 512x512 pixel with MLP
nf-net-f1.ipynb	https://www.kaggle.com/mfalfafa/shopee-nf-net-f1-all-training	Kaggle TPU	Create multimodal model with NFNet-F1 pretrained model using image size 512x512 pixel with MLP
roberta-base-id.ipynb	https://www.kaggle.com/mfalfafa/shopee-roberta-base-id-all-training	Kaggle TPU	Create multimodal model with RoBERTa-Base-Id pretrained model with MLP
tfrecord-512-gen.ipynb	https://www.kaggle.com/mfalfafa/shopee-tfrecord-512-all	Kaggle CPU	Generate TFRecord with image size 512x512 pixel for training of image model Images
text-cleaner.ipynb	https://www.kaggle.com/mfalfafa/shopee-text-cleaner-for-roberta-base-id-all	Kaggle CPU	Create text preprocessing for BERT model
text-vectorizer.ipynb	https://www.kaggle.com/mfalfafa/text-vectorizer-for-all-training	Kaggle CPU	Create vectorized text for MLP model
final-solution.ipynb	https://www.kaggle.com/mfalfafa/shopee-final-solution	Kaggle GPU	Final solution for product predictions. This notebook used for submission

Generated Models

The models are generated using training notebooks. This models can be used to make quick submission using final-solution notebook.

Training file	Generated model	Link to Kaggle datasets
effb3-512.ipynb	`EfficientNet B3 + MLP with ArcMargin`	https://www.kaggle.com/mfalfafa/shopee-effb3-512
effb5-512.ipynb	`EfficientNet B5 + MLP with ArcMargin`	https://www.kaggle.com/mfalfafa/shopee-effb5-512-v2
nf-net-f0.ipynb	`NFNet-F0 + MLP with ArcMargin`	https://www.kaggle.com/mfalfafa/shopee-nfnet-512
nf-net-f1.ipynb	`NFNet-F1 + MLP with ArcMargin`	https://www.kaggle.com/mfalfafa/shopee-nfnet-f1-512
roberta-base-id.ipynb	`RoBERTa + MLP with ArcMargin`	https://www.kaggle.com/mfalfafa/shopee-roberta-base-id

Dependency Datasets

Dependency datasets are used for training the model and submitting the solution.

Dataset name	Description	Link to Kaggle datasets
keras_efficientnet_whl	Python libraries for Keras EfficientNet model	https://www.kaggle.com/alanchn31/keras-efficientnet-whl
nfnets_keras	Python libraries for NFNet model	https://www.kaggle.com/dsofen/nfnets-keras
tfroberta_base_indonesian	Pretrained model for RoBERTa-Base-Id	https://www.kaggle.com/mfalfafa/tfroberta-base-indonesian
NFNET_Model_Checkpoints_270421	Pretrained models for NFNet	https://www.kaggle.com/dsofen/nfnet-model-checkpoints-270421

LB/Leaderboard Score

Notebook filename	Submission filename	Public LB	Private LB
final-solution.ipynb	submission.csv	0.744	0.733

Reproducibility Guide

This guide assume you have necessary files (full dataset provided by Shopee and dependency datasets), move it to correct directory path and run it on Kaggle Notebook.

Run text-vectorizer.ipynb, text-cleaner.ipynb
Run tfrecord-512-gen.ipynb
Run effb3-512.ipynb
Run effb5-512.ipynb
Run nf-net-f0.ipynb
Run nf-net-f1.ipynb
Run final-solution.ipynb