In order to obtain the predictions based on the class-specific count thresholds, we recommend using a window size of 10s and a hop size of 1s (-w 10 -s 1).
Then apply the postprocess_predictions.py script to obtain the final predictions by specifying the csv path (--path) to the results.csv:
The model has been originally trained on AudioSet by Kong et. al..
Dataset
The model has been further trained (finetuned) on the training set of the EDANSA2019 dataset. The dataset was collected in the North Slope of Alaskan at latitudes between 64◦ and 70◦ N, and longitudes between 139◦ to 150◦ W from a total of 40 devices, each placed in a different location, separated by ca. 20kM from other locations. A subset of the entire dataset has been annotated for 28 labels (tags), of which only the 4 highest level categories were used: anthropophony, biophony, geophony, and silence. The sampling rate was 48kHz.
Features
The EDANSA2019 dataset was resampled to 32kHz, as this was the sampling rate of AudioSet, where the model was originally trained on. Log-Mel spectrograms were then extracted using torchlibrosa using the parameters that the upstream model was trained on.
Training process
The model has been trained for 30 epochs. At the end of each epoch, the model was evaluated on the official validation set. We release the state that achieved the best performance on this validation set. All training hyperparameters can be found inside conf/config.yaml inside the model folder.
Evaluation
The model has only been evaluated on in-domain data. The performance on the official test set reached a 0.87 (weighted) f1-score.
Acknowledgments
Please acknowledge the work which produced the original model and the EDANSA2019 dataset. We would also appreciate an acknowledgment to autrainer.
