/crabs-exploration

A toolkit for detecting and tracking crabs in the field.

Primary LanguagePythonBSD 3-Clause "New" or "Revised" LicenseBSD-3-Clause

crabs-exploration

License CI codecov

A toolkit for detecting and tracking crabs in the field.

Getting Started

Prerequisites

requires Python 3.9 or 3.10 or 3.11.

Installation

Data Structure

We assume the following structure for the dataset directory:

|_ Dataset
    |_ frames
    |_ annotations
        |_ VIA_JSON_combined_coco_gen.json

The default name assumed for the annotations file is VIA_JSON_combined_coco_gen.json. This is used if no input files are passed. Other filenames (or fullpaths) can be passed with the --annotation_files command-line argument.

Running Locally

For training

python train-detector --dataset_dirs {parent_directory_of_frames_and_annotation} {optional_second_parent_directory_of_frames_and_annotation} --annotation_files {path_to_annotation_file.json} {path_to_optional_second_annotation_file.json}

Example (using default annotation file and one dataset):

python train-detector --dataset_dirs /home/data/dataset1

Example (passing the full path of the annotation file):

python train-detector --dataset_dirs /home/data/dataset1 --annotation_files /home/user/annotations/annotations42.json

Example (passing several datasets with annotation filenames different from the default):

python train-detector --dataset_dirs /home/data/dataset1 /home/data/dataset2 --annotation_files annotation_dataset1.json annotation_dataset2.json

For evaluation

python evaluate-detector --model_dir {directory_to_saved_model} --images_dirs {parent_directory_of_frames_and_annotation} {optional_second_parent_directory_of_frames_and_annotation} --annotation_files {annotation_file.json} {optional_second_annotation_file.json}

Example:

python evaluate-detector --model_dir model/model_00.pt --main_dir /home/data/dataset1/frames /home/data/dataset2/frames --annotation_files /home/data/dataset1/annotations/annotation_dataset1.json /home/data/dataset2/annotations/annotation_dataset2.json

For running inference

python crabs/detection_tracking/inference_model.py --model_dir {oath_to_trained_model} --vid_path {path_to_input_video}

MLFLow

We are using MLflow to log our training loss and the hyperparameters used. To run MLflow, execute the following command in your terminal:

mlflow ui --backend-store-uri file:///<path-to-ml-runs>

Replace <path-to-ml-runs> with the path to the directory where you want to store the MLflow output. By default, it's an ml-runs directory under the current working directory.