CameraTrapDetectoR Model Training

This repository contains the Python code used to train new versions of the CameraTrapDetectoR models.

CameraTrapDetectoR Model Deployment

We also provide documentation and assistance for deploying our models via Python script on the command line. The deploy_model.py script provides similar output to the deploy_model function in the CameraTrapDetectoR package. This approach allows users to take advantage of a GPU-enabled machine or run the models via batch mode on a remote instance, for example on a high-performance computing (HPC) cluster.

Installation

1. System Prerequisites

You will need Anaconda and Git installations to set up your model environment. If your environment can access a Pytorch-compatible GPU, make sure you have an updated NVIDIA driver installed as well.

2. Clone Git Repository

Open your Anaconda prompt and navigate to the directory where you'd like to store your project. Clone this repository:

cd /path/to/project/dir
git clone https://github.com/CameraTrapDetectoR/model_training.git

3. Create Python Environment

Navigate into the model_training directory and locate a file called deploy_model_env.yml. This file contains the setup info and required Python packages needed to run the models. You can create your environment using the YAML file:

cd model_training
conda env create -n ctd-deploy-model -f deploy_model_env.yml

or you can manually create a Python environment and install the listed packages using the package manager of your choosing.

This step may take a while! Go get a cup of coffee or a yummy snack, and start downloading the models while you wait.

Check that your environment was successfully created, and add the repository to your Python path:
For Windows:

conda activate ctd-deploy-model
set PYTHONPATH=%PYTHONPATH%;\path\to\project\dir\model_training

For Mac/Linux:

conda activate ctd-deploy-model
export PYTHONPATH="$PYTHONPATH:$path/to/project/dir/model_training

4. Download Models

Download the model architecture, model weights, and auxiliary information needed to run each desired model. Model versions 2 and later are supported by this script:

• General V2 Model
• Family V2 Model
• Species V2 Model

The [user tutorial] assumes you are putting the model folder in the same project directory as this repo; you can place it anywhere you like, just adjust accordingly. Once you download the model folder to its final location, unzip it. DO NOT rename, modify, or delete anything inside this folder or your model will not run. Inside each folder are three files specific to running that model, titled cameratrapdetector_metadata.cfg, model_args.txt, and model_checkpoint.pth. Do not change anything! Just unzip it and leave it.

Now you are ready to run the CameraTrapDetectoR models!

Use

1. Activate Python Environment

Activate your Python environment and, if necessary, export the path to the model_training repo to your Python path:
For Windows:

conda activate ctd-deploy-model
set PYTHONPATH=%PYTHONPATH%;\path\to\project\dir\model_training

For Mac/Linux:

conda activate ctd-deploy-model
export PYTHONPATH="$PYTHONPATH:$path/to/project/dir/model_training

2. The deploy_model.py script

Inside the main folder of the model_training repo, you will see a file titled deploy_model.py; you will pass your user arguments to this script through the command line. This script provides a similar output to the deploy_model function in the CameraTrapDetectoR package. Navigate into the repo and run the help command to view descriptions of the inputs:

cd /path/to/project/dir/model_training
python deploy_model.py --help

The required arguments are both absolute paths:

• model_folder: the absolute path to the folder containing all model files that was downloaded earlier. If you put the model folders in the same directory as the model_training repo, then an example path to the species_v2 model folder would read: path/to/project/dir/species_v2_cl_0
• image_dir: the absolute path to the parent directory containing your images. The model script scans recursively through this directory, so if you want to restrict a model run to a specific folder, you will need to specify that folder. Currently, only .jpg files are accepted. If you require additional file types to run through the model, please let us know

The optional arguments give the user some control over results placement, checkpointing, etc. and are detailed in the help. Here is a sample command to deploy the species version 2 model on images from Project A, saving the results to a separate results directory, with a score threshold of 0.5, an overlap threshold of 0.8, a checkpoint frequency of 100, and resuming from a previous checkpoint:

python deploy_model.py /path/to/model_folder/species_v2 /path/to/images/ProjectA --output_dir /path/to/results/ProjectA --score_threshold 0.5 --overlap_threshold 0.8 --checkpoint_frequency 100 --resume_from_checkpoint /path/to/results/ProjectA/species_v2_checkpoint_20231010082035.csv