Auto Plankton Training

The mbari-auto-plankton-training package allows to both train a new model and to classify new images for bootstrapping a dataset.

To install the required packages, set up a python environment and then pip into a terminal:

pip install -r requirements.txt

Training

Training Data needs to be in a file called "Training_Data"

mkdir "Training_Data"

In order to train a model, match the training categories in classification.py with the categories in the dir "Training_Data". Once this is done and training data exists, type into the terminal that is in the directory with both "Training_Data" and classification.py

python classification.py -t

The -t flag will set the programing into training mode to create a new model if no previous model exists. It will both look and save a model called HM_model.pth. If a model does not exist, then classification.py will create it.

Classification

In order to with HM_model.pth to classify new data for bootstrapping, images must be in a file called "New_Data"

mkdir "New_Data

Dump all images that are to be categorized into this directory. Do not place other directories into this directory as classification.py will not see those images.

To start to classification process, type into the terminal that is in the directory with both "New_Data" and classification.py

classification.py -c

The -c flag will set the programing into classification mode. A model must exist called HM_model.pth and be in the same directory as classification.py. All classifed images will be placed into a new directory called "Categorized Data". The program will make this directory and does not need to exist before running classification.py. All images will be placed into directories with their associated categorized name as specified in classification.py.

History

Every completed run of classification.py will result in a History directory to be created. This directory will included previous models created, CSVs of images that the models were trained on, and CSVs of images that where categorized depending on the command line arguments given to classification.py