Classifier for Plankton data

This repository is for working on a Classifier for images of plankton. The major difficulty is accounting for the different image sizes.

You can download the data from the strand server (strand.fzg.local) at /gpfs/work/machnitz/plankton-dataset

Using this repo

To run the vanilla model just run main.py. A conda environment which works with this project can be installed with conda env create -f environment.yaml. You can get a list of all options using python main.py --help. For more information on configuration see the hydra docs: https://hydra.cc/

This repo contains a submodule, so after cloning it you have to also clone the submodule:

git clone https://github.com/m-dml/plankton-classifier
cd plankton-classifier
git submodule init
git submodule update

Inference with a trained model

What you need:

• The checkpoint file of the trained model some_file.ckpt
• The integer-to-labelname file class_labels.json that was created during training of the model
• A folder containing images of plankton to be classified. This folder is allowed to have subfolders. The images should be ending on ".png".

1. Install the environment with conda env create -f environment.yaml.
2. Activate the environment with conda activate plankton.
3. Make sure the checkpoint and the class-label file into one folder.
4. Run the inference script

   1. Run it locally with: python main.py +experiment=inference/inference load_state_dict=some_file.ckpt output_dir_base_path=/path/to/store/outputs/ datamodule.unlabeled_files_to_append=/path/to/the/image/folder

      • on Windows make sure to also always add datamodule.num_workers=0
      • This command assumes that you have a GPU and are running the program locally. For more control clone the configuration file conf/experiment/inference/inference.yaml and make changes accordingly.
      • To use the CPU add trainer.accelerator="cpu" to the command

   2. To run the inference script on a slurm cluster add -m at the end of the command. Make sure that the right trainer and hydra-launcher are selected in your script.

   3. For easy work on strand use the following command. It will allocate a GPU node and makes the inference: python main.py +experiment=inference/strand load_state_dict=some_file.ckpt output_dir_base_path=/path/to/store/outputs/ datamodule.unlabeled_files_to_append=/path/to/the/image/folder -m

pre-commit

Just use pre-commit run --all-files at the top level of this repo, to let precommit handle the files.