0x4l3x/herbarium-2022-fgvc9-flora-of-north-america

The project elaborated for the kaggle competition Herbarium 2022 - FGVC9

Herbarium 2022 - FGVC9 Competition Solution

This repository contains the code, paper, and poster for our solution to the Herbarium 2022 - FGVC9 competition. The competition involves identifying plant species through herbarium sheets. Our paper provides a detailed explanation of the solution adopted, along with a study of the machine learning field that comprises it.

Abstract

Plants are of fundamental importance to life on Earth. With the threat of climate change advancing, new ideas are needed to combat it. This is where this paper comes in, a solution to the Herbarium 2022 - FGVC9 competition. It consists of identifying plant species through herbarium sheets. This paper details the solution adopted, as well as a study of the machine learning field that comprises it.

Files Overview

• bit_train/bit_pytorch: The bit_train folder corresponds to the bit_pytorch folder available on the BigTransfer GitHub repository.

• bit_hyperrule.py: This file corresponds to the bit_hyperrule.py file in the BigTransfer repository. It is related to the hyperparameter rules for BigTransfer adaptation.

• HerbariumTrainingScript.py: The HerbariumTrainingScript.py file contains all the training methods for the ResNets trained over the Herbarium dataset.

• statVisualization.ipynb: The statVisualization.ipynb file is a notebook for making graphs with the metrics of the models.

• badSamples.ipynb: The badSamples.ipynb notebook corresponds to code for analyzing the bad samples of the dataset.

• DataVisualization.ipynb: The DataVisualization.ipynb notebook contains methods for visualizing the data of the Herbarium 2022 competition.

Additional Resources

• Paper - Herbarium.pdf: The Paper - Herbarium.pdf file contains the detailed write-up of our solution. You can access the paper in PDF format.

• The Herbarium 2022 - poster.pdf: The The Herbarium 2022 - poster.pdf file is the visual representation of our solution, as presented in the competition.