Snake Species Identification

Overview

This repository contains supplementary material to the working note paper A Deep Learning Method for Visual Recognition of Snake Species submitted to the SnakeCLEF 2021 challenge - a part of LifeCLEF 2021 workshop.

The goal of the challenge was creating a method for image-based snake species identification. The proposed method is based on deep residual neural networks, namely ResNet, ResNeXt and ResNeSt, fine-tuned from ImageNet-1k pre-trained checkpoints.

Content

The repository contains Jupyter Notebooks for training and testing CNNs:

• 01a_data_exploration.ipynb - The notebook contains a high-level exploration of the training dataset including visualization of snake species distribution and geographical distribution of images.
• 01b_image_exploration.ipynb - The notebook contains exploration of images in the training dataset including examples of noisy data from Flickr and estimation of the confidence interval of noisy images in Flickr using Student's t-distribution.
• 02_training.ipynb - The notebook contains a training pipeline in fastai framework.
• 03a_testing.ipynb - The notebook contains a pipeline for creating predictions on the validation and test sets.
• 03b_ensemble.ipynb - The notebook contains a pipeline for creating ensemble model predictions by applying majority voting strategy.
• 04_evaluating.ipynb - The notebook contains a pipeline for evaluating predictions of a single CNN. This includes classification report with precision, recall and F1 scores, and examples of correct and incorrect predictions with high or low confidence.
• 05_submission_results.ipynb - The notebook contains classification scores of CNNs submitted to the SnakeCLEF 2021 challenge.

Additionally, the repository contains Jupyter Notebooks with experiments comparing prediction scores under various setups:

• 01_models.ipynb - The notebook compares residual networks ResNet, ResNeXt and ResNeSt with 50 and 101 layers.
• 02_loss_functions.ipynb - The notebook compares loss functions Cross Entropy, Weighted Cross Entropy and F1 Loss.
• 03_optimizers.ipynb - The notebook compares optimization algorithms SGD with momentum and Adam with and without one cycle schedule policy.
• 04_mixed_precision.ipynb - The notebook compares the single precision training, used by default, and the mixed precision training.

Getting Started

Dataset

The training dataset was provided by the organizers of SnakeCLEF 2021 challenge, and it is not part of this repository. The dataset was obtained through AIcrowd.

Package Dependencies

The proposed methods were developed using Python=3.7 with PyTorch=1.7 machine learning framework and a fastai=2.3 framework build on top of PyTorch. All models were fine-tuned from pre-trained PyTorch Image Models, timm=0.4. Additionally, the repository requires packages: numpy, pandas, scikit-learn, matplotlib and seaborn.

To install required packages with PyTorch for CPU run:

pip install -r requirements.txt

For PyTorch with GPU run:

pip install -r requirements_gpu.txt

The requirement files do not contain jupyterlab nor any other IDE. To install jupyterlab run

pip install jupyterlab

The Jupyter Notebooks use pycodestyle magic function which checks Python code against the PEP 8 style conventions allowing to maintain a well readable and clean code. To install pycodestyle for jupyterlab run:

pip install flake8 pycodestyle_magic

Contact

Rail Chamidullin - [email protected] - Github account