planktonclass: FlowCam

Author: Wout Decrop (VLIZ) Related publication: Automated image classification workflow for phytoplankton monitoring Resources: Documentation PyPI package Package downloads Projects: iMagine planktonclass is a toolkit for training, evaluating, and serving phytoplankton image classifiers! It was originally developed for FlowCam data, and has also been retrained or adapted in separate branches for other instruments and datasets: FlowCam / main branch Zooscan branch Cyto branch PI10 branch If you want the full repository with Docker, OSCAR, AI4OS, packaged deployment assets, and broader project explanation, see: phyto-plankton-classification

+. +: :==. % .#. #:*==* *= -+**+*####. +********%%. +*******#**#+ ********#%%####+ .*====+==::=#%%* -%** --::=-:. +=#. -:::+. -+*++: +. +:::* :+. .+- ==: +::::* =- == ::-+*+:::::*##- .+. :+-.-====-:::::+%#. ===*: :++::::-=:++*#= -#. -+**:::=*++**%##+ .=+-= ##*:**#*%******= .=**+ =*++#************#- .++*****++++++++*##+ :+*+#%++++++++*+. *** :###- ::#**. +**+ .%@+.: --@@@% :.

Install

Install with Python 3.12 and pip:

pip install planktonclass

For notebook support:

pip install "planktonclass[notebooks]"

Choose Your Path

1. I want to train locally

Use:

planktonclass train my_project

This is the best choice if you already know where your image folder is and want a direct local workflow.

2. I want to use a browser UI / API

Use:

planktonclass api my_project

Then open:

• http://127.0.0.1:5000/ui
• http://127.0.0.1:5000/api#/

This is the best choice if you want to interact through the DEEPaaS UI or integrate with an external service.

3. I want notebooks

Use:

planktonclass notebooks my_project

This copies the packaged notebooks into my_project/notebooks/. It is the best choice for exploration, augmentation experiments, prediction analysis, and explainability.

pip install planktonclass installs the package dependencies used by the notebooks, including TensorFlow, plotting, and reporting libraries. For local notebook use, install the notebook extra instead:

pip install "planktonclass[notebooks]"

Quick Start

Option A: Use it locally

Read the Docs site

pip install planktonclass

Then create a project:

planktonclass init my_project

Or create a runnable demo project:

planktonclass init my_project --demo

OPTIONAL: Validate the generated config:

planktonclass validate-config my_project

Local training:

planktonclass train my_project

For a quick smoke test on the demo project:

planktonclass train my_project --quick

OPTIONAL: Download a published pretrained model into the project:

planktonclass pretrained my_project --model FlowCam

Available published pretrained model names currently include FlowCam, FlowCyto, and PI10. Only the actual model directory is extracted into my_project/models, even when the downloaded archive contains a full exported project tree.

OPTIONAL: Build an inference Docker image from your trained model run:

planktonclass docker my_project

For the published FlowCam pretrained model, the packaged checkpoint is currently final_model.h5. The FlowCyto and PI10 published models are expected to use best_model.keras. New training runs created by planktonclass train save best_model.keras when validation is enabled. If you train without validation, the run saves final_model.keras instead.

Report generation after training:

planktonclass report my_project

If you leave out --timestamp, planktonclass report suggests the most recent run, lists the available timestamps, and lets you choose another one by number. It also lets you choose between quick and full mode. quick is the default and creates the core figures only; full also generates the threshold-based plots in the results/ subfolders.

Option B: Use api

Read the Docs site

pip install planktonclass

Then create a project:

planktonclass init my_project

Local API:

planktonclass api my_project

Option C: I want notebooks

For local notebook use:

pip install "planktonclass[notebooks]"

Then create a project:

planktonclass init my_project

Copy notebooks into the project:

planktonclass notebooks my_project

In the model-based notebooks (3.0, 3.1, and 3.2), the first variables to check are TIMESTAMP and MODEL_NAME. They are prefilled for the published pretrained model so the notebooks work out of the box, but when you want to inspect a model from your own training run you should change those two values first.

Project Structure

After planktonclass init, your project looks like this:

my_project/
  config.yaml
  data/
    images/
    dataset_files/
  models/
  notebooks/

What is required?

The only mandatory input is the image directory:

• data/images/
• or the directory pointed to by images_directory in config.yaml

If data/dataset_files/ is empty, training can generate dataset splits automatically from the image-folder structure.

If you provide your own dataset metadata files, the expected files are:

• custom-split required: classes.txt, train.txt
• optional: val.txt, test.txt, info.txt, aphia_ids.txt

The split files map image paths to integer labels starting at 0.

Configuration

The main user config is a project-local config.yaml.

It is created by:

planktonclass init my_project

Most users only need to adjust a small number of fields:

• general.base_directory
• general.images_directory
• model.modelname
• pretrained.use_pretrained
• pretrained.name
• pretrained.version
• training.epochs
• training.batch_size
• training.use_validation
• training.use_test
• monitor.use_tensorboard

Internal-only values such as model-specific preprocessing are now derived automatically and are not meant to be edited by users.

Local CLI Workflow

The package installs a planktonclass command with these main subcommands:

• planktonclass init [DIR]
• planktonclass init [DIR] --demo
• planktonclass validate-config [DIR]
• planktonclass train [DIR]
• planktonclass report [DIR] [--timestamp TS]
• planktonclass api [DIR]
• planktonclass docker [DIR]
• planktonclass pretrained [DIR]
• planktonclass list-models [DIR]
• planktonclass notebooks [DIR]

The pretrained command accepts a published model name and version, for example:

planktonclass pretrained my_project --model FlowCyto --version latest

The list-models command now shows published pretrained models with extra metadata such as architecture, version, and checkpoint name, while local timestamped runs still appear as plain folder names.

Typical local workflow:

planktonclass init my_project
planktonclass notebooks my_project
planktonclass validate-config my_project
planktonclass train my_project
planktonclass docker my_project
planktonclass report my_project

For a faster package smoke test with the demo data:

planktonclass init my_project --demo
planktonclass train my_project --quick
planktonclass report my_project

API Workflow

Start the API with:

planktonclass init my_project
planktonclass api my_project

Then open:

• http://127.0.0.1:5000/ui
• http://127.0.0.1:5000/api#/

You can also start DEEPaaS directly after a repo install:

$env:planktonclass_CONFIG = (Resolve-Path .\my_project\config.yaml)
$env:DEEPAAS_V2_MODEL = "planktonclass"
deepaas-run --listen-ip 0.0.0.0

Important notes:

• 0.0.0.0 is a bind address, not the browser URL
• open 127.0.0.1 in the browser
• for prediction, the browser UI supports file uploads for image and zip
• for training, images_directory is a path field, so it must point to a folder visible to the machine running the API

Notebook Workflow

Copy the packaged notebooks into your project with:

planktonclass init my_project
planktonclass notebooks my_project

The copied notebooks auto-detect the nearest project config.yaml, so they use the paths inside your local project folder rather than the installed package directory. They also copy data/data_transformation/start, reference_style, and end for the transformation notebook.

Notebook overview:

• 1.0-Dataset_exploration.ipynb
• 1.1-Image_transformation.ipynb
• 1.2-Image_augmentation.ipynb
• 2.0-Model_training.ipynb
• 3.0-Computing_predictions.ipynb
• 3.1-Prediction_statistics.ipynb
• 3.2-Saliency_maps.ipynb

For 1.1-Image_transformation.ipynb:

• put your new raw images in data/data_transformation/start/
• keep one or more reference images in data/data_transformation/reference_style/
• the transformed outputs are written to data/data_transformation/end/

Outputs

Each training run creates a timestamped folder under models/:

models/<timestamp>/
  ckpts/
  conf/
  logs/
  stats/
  dataset_files/
  predictions/
  results/

Useful outputs include:

• checkpoints like best_model.keras
• stats.json
• saved prediction JSON files
• saved test metrics JSON files with top-k accuracy, precision, recall, and F1 summaries
• report images and CSV summaries under results/

For a portable inference runtime after training, you can package a selected model run into a Docker image:

planktonclass docker my_project

This builds an image from the local package source and bundles the latest trained timestamp by default. You can choose a specific run or checkpoint with:

planktonclass docker my_project --timestamp 2026-04-21_120000 --ckpt-name best_model.keras --tag my-plankton-api:latest

To generate performance plots after training:

planktonclass report my_project

If you keep the standard project layout created by planktonclass init, these commands automatically use my_project/config.yaml. Use --config PATH only when your config file lives somewhere else.

More Documentation

The full documentation is available here:

• Read the Docs site
• Documentation entry page

Main documentation pages:

• Installation
• Quickstart
• API usage
• Python usage
• Notebooks
• Reference

For Docker, OSCAR, AI4OS, and the broader deployment-oriented repository, see:

• https://github.com/ai4os-hub/phyto-plankton-classification

Development

Choose this only if you want to work on the package itself.

git clone https://github.com/lifewatch/planktonclass
cd phyto-plankton-classification
python -m venv .venv
.venv\Scripts\activate
pip install -U pip
pip install -e .
pip install -e ".[dev]"
python -m pytest

Acknowledgements

If you use this project, please consider citing:

Decrop, W., Lagaisse, R., Mortelmans, J., Muñiz, C., Heredia, I., Calatrava, A., & Deneudt, K. (2025). Automated image classification workflow for phytoplankton monitoring. Frontiers in Marine Science, 12. https://doi.org/10.3389/fmars.2025.1699781