Code release of team "BBracke" on SnakeCLEF2023 challenge

The folders "exp1" to "exp5" contain training scripts for the models according to the experiments we performed in our working notes paper. To ensure reproducibility of the experiments, some hyperparameters may need to be changed, e.g. different image sizes in "exp2", see working notes paper for details.

The training data and the configuration files of the SnakeCLEF23 challenge must be downloaded separately (https://huggingface.co/spaces/competitions/SnakeCLEF2023) and the path to them must be defined in the scripts.

Descriptions of files that are stored in this repository:

• "missing_train_data.csv" contains image files that are missing in the challenge provided datasets and are filtered out during training
• "classDist_HMP_missedRemoved.p" is a pickled dictionary containing the snake species class indices with their corresponding frequency (observation_id level) in the challenge training + additional datasets
• "code_class_mapping_obid.csv" containing the binarised snake species class distribution in relation to the code metadata
• "meta_code_tokens.p" is a pickled dictionary that contains the mapping of code metadata to code tokens (integer value) that are inserted into embedding layers
• "meta_endemic_tokens.p" is a pickled dictionary that contains the mapping of endemic metadata to endemic tokens (integer value) that will be inserted into embedding layers

Descriptions of the "exp" folders, which refer to the experiments in the working notes paper:

• exp1 refers to "Experiment: iNaturalist21 Pre-Training" (iNaturalist pre-trained model weights must be downloaded separately from "https://huggingface.co/BBracke/convnextv2_base.inat21_384")
• exp2 refers to "Experiment: Influence of Image Size" (the image size hyperparameters need to be changed in the code)
• exp3 refers to "Experiment: Leveraging Classification Results with Metadata" (this is the proposed joint feature learning model with embedded metadata)
• exp4 refers to "Experiment: Class Imbalance Learning" (the hyperparameters for focal loss and class weighting need to be changed in the code)
• exp5 refers to the "Experiment: Influence of MIL-Pooling Operators" (only training scripts of the models with attention modules)