Meli Data Challenge 2021 - First Place Solution

Overview

Meli Data Challenge 2021 - First Place Solution

My solution for the Meli Data Challenge 2021, first place in both public and private leaderboards.

The Model

My final model is an ensemble combining recurrent neural networks and XGBoost regressors. Neural networks are trained to predict the stock days probability distribution using the RPS as loss function. XGBoost regressors are trained to predict stock days using different objectives, here the intuition behind this:

  • MSE loss: the regressor trained with this loss will output values close to the expected mean.
  • Pseudo-Huber loss: an alternative for the MAE loss, this regressor outputs values close to the expected median.
  • Quantile loss: 11 regressors are trained using a quantile loss with alpha 0, 0.1, 0.2, ..., 1. This helps to build the final probability distribution.

The outputs of all these level-0 models are concatenated to train a feedforward neural network with the RPS as loss function.

diagram

The last 30 days of the train dataset are used to generate the labels and the target stock input. The remaining 29 days are used to generate the time series input.

The train/validation split is done at a sku level:

  • For level-0 models: 450000 sku's are used for training and the rest for validation.
  • For the level-1 model: the sku's used for training level-0 models are removed from the dataset and the remaining sku's are split again into train/validation.

Once all models are trained, the last 29 days of the train dataset and the provided target stock values are used as input to generate the submission.

Disclaimer: the entire solution lacks some fine tuning since I came up with this little ensemble monster towards the end of the competition. I didn't have the time to fine-tune each model (there are technically 16 models to tune if we consider each quantile regressor as an independent model).

How to run the solution

Requirements

  • TensorFlow v2.
  • Pandas.
  • Numpy.
  • Scikit-learn.

CUDA drivers and a CUDA-compatible GPU is required (I didn't have the time to test this on a CPU).

Some scripts require up to 30GB of RAM (again, I didn't have the time to implement a more memory-efficient solution).

The solution was tested on Ubuntu 20.04 with Python 3.8.10.

Downloading the dataset

Download the dataset files from https://ml-challenge.mercadolibre.com/downloads and put them into the dataset/ directory.

On linux, you can do that by running:

cd dataset && wget \
https://meli-data-challenge.s3.amazonaws.com/2021/test_data.csv \
https://meli-data-challenge.s3.amazonaws.com/2021/train_data.parquet \
https://meli-data-challenge.s3.amazonaws.com/2021/items_static_metadata_full.jl

Running the scripts

All-in-one script

A convenient script to run the entire solution is provided:

cd src
./run-solution.sh

Note: the entire process may take more than 3 hours to run.

Step by step

If you find trouble running the al-in-one script, you can run the solution step by step following the instructions bellow:

cd into the src directory:

cd src

Extract time series from the dataset:

python3 ./preprocessing/extract-time-series.py

Generate a supervised learning dataset:

python3 ./preprocessing/generate-sl-dataset.py

Train all level-0 models:

python3 ./train-all.py

Train the level-1 ensemble:

python3 ./train-ensemble.py

Generate the submission file and gzip it:

python3 ./generate-submission.py && gzip ./submission.csv

Utility scripts

The training_scripts directory contains some scripts to train each model separately, example usage:

python3 ./training_scripts/train-lstm.py
Owner
Matias Moreyra
Electronics Engineer, Software Developer.
Matias Moreyra
Contour-guided image completion with perceptual grouping (BMVC 2021 publication)

Contour-guided Image Completion with Perceptual Grouping Authors Morteza Rezanejad*, Sidharth Gupta*, Chandra Gummaluru, Ryan Marten, John Wilder, Mic

Sid Gupta 6 Dec 27, 2022
ObjectDetNet is an easy, flexible, open-source object detection framework

Getting started with the ObjectDetNet ObjectDetNet is an easy, flexible, open-source object detection framework which allows you to easily train, resu

5 Aug 25, 2020
AirCode: A Robust Object Encoding Method

AirCode This repo contains source codes for the arXiv preprint "AirCode: A Robust Object Encoding Method" Demo Object matching comparison when the obj

Chen Wang 30 Dec 09, 2022
Over-the-Air Ensemble Inference with Model Privacy

Over-the-Air Ensemble Inference with Model Privacy This repository contains simulations for our private ensemble inference method. Installation Instal

Selim Firat Yilmaz 1 Jun 29, 2022
Deep Reinforcement Learning for Keras.

Deep Reinforcement Learning for Keras What is it? keras-rl implements some state-of-the art deep reinforcement learning algorithms in Python and seaml

Keras-RL 0 Dec 15, 2022
68 keypoint annotations for COFW test data

68 keypoint annotations for COFW test data This repository contains manually annotated 68 keypoints for COFW test data (original annotation of CFOW da

31 Dec 06, 2022
Use tensorflow to implement a Deep Neural Network for real time lane detection

LaneNet-Lane-Detection Use tensorflow to implement a Deep Neural Network for real time lane detection mainly based on the IEEE IV conference paper "To

MaybeShewill-CV 1.9k Jan 08, 2023
Revisiting Oxford and Paris: Large-Scale Image Retrieval Benchmarking

Revisiting Oxford and Paris: Large-Scale Image Retrieval Benchmarking We revisit and address issues with Oxford 5k and Paris 6k image retrieval benchm

Filip Radenovic 188 Dec 17, 2022
Deep learning with TensorFlow and earth observation data.

Deep Learning with TensorFlow and EO Data Complete file set for Jupyter Book Autor: Development Seed Date: 04 October 2021 ISBN: (to come) Notebook tu

Development Seed 20 Nov 16, 2022
Python implementation of Lightning-rod Agent, the Stack4Things board-side probe

Iotronic Lightning-rod Agent Python implementation of Lightning-rod Agent, the Stack4Things board-side probe. Free software: Apache 2.0 license Websit

2 May 19, 2022
Use CLIP to represent video for Retrieval Task

A Straightforward Framework For Video Retrieval Using CLIP This repository contains the basic code for feature extraction and replication of results.

Jesus Andres Portillo Quintero 54 Dec 22, 2022
BasicNeuralNetwork - This project looks over the basic structure of a neural network and how machine learning training algorithms work

BasicNeuralNetwork - This project looks over the basic structure of a neural network and how machine learning training algorithms work. For this project, I used the sigmoid function as an activation

Manas Bommakanti 1 Jan 22, 2022
Improving Object Detection by Estimating Bounding Box Quality Accurately

Improving Object Detection by Estimating Bounding Box Quality Accurately Abstrac

2 Apr 14, 2022
This is the dataset and code release of the OpenRooms Dataset.

This is the dataset and code release of the OpenRooms Dataset.

Visual Intelligence Lab of UCSD 95 Jan 08, 2023
Remote sensing change detection tool based on PaddlePaddle

PdRSCD PdRSCD(PaddlePaddle Remote Sensing Change Detection)是一个基于飞桨PaddlePaddle的遥感变化检测的项目,pypi包名为ppcd。目前0.2版本,最新支持图像列表输入的训练和预测,如多期影像、多源影像甚至多期多源影像。可以快速完

38 Aug 31, 2022
A PyTorch implementation of Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks

SVHNClassifier-PyTorch A PyTorch implementation of Multi-digit Number Recognition from Street View Imagery using Deep Convolutional Neural Networks If

Potter Hsu 182 Jan 03, 2023
Pytorch modules for paralel models with same architecture. Ideal for multi agent-based systems

WideLinears Pytorch parallel Neural Networks A package of pytorch modules for fast paralellization of separate deep neural networks. Ideal for agent-b

1 Dec 17, 2021
This is the repository for the AAAI 21 paper [Contrastive and Generative Graph Convolutional Networks for Graph-based Semi-Supervised Learning].

CG3 This is the repository for the AAAI 21 paper [Contrastive and Generative Graph Convolutional Networks for Graph-based Semi-Supervised Learning]. R

12 Oct 28, 2022
A Kitti Road Segmentation model implemented in tensorflow.

KittiSeg KittiSeg performs segmentation of roads by utilizing an FCN based model. The model achieved first place on the Kitti Road Detection Benchmark

Marvin Teichmann 890 Jan 04, 2023
A PyTorch Implementation of SphereFace.

SphereFace A PyTorch Implementation of SphereFace. The code can be trained on CASIA-Webface and the best accuracy on LFW is 99.22%. SphereFace: Deep H

carwin 685 Dec 09, 2022