A. Out-of-sample spherical k-means model
-
[DONE] Train on external data.
-
[DONE] Compute predictions on 6 full night recordings.
-
[DONE] Postprocessing: peak-picking, thresholding, event matching.
-
[DONE] Export metrics (n_selected, TP, FP, FN, precision, recall, F) for all 6 units, 10 tolerances, and 100 thresholds, to 6*10=60 CSV files.
-
[DONE] Compute global metrics (n_selected, TP, FP, FN, precision, recall, F and AUPRC) across all 6 units and 10 tolerances. Store in 1 CSV file.
B. Old Bird
-
[DONE] Implement Tseep and Thrush onset detection function (ODF) from Vesper.
-
[DONE] Run Tseep, Thrush on 6 full night recordings. Export into 6 HDF5 containers by chunks. Parallelize over units.
-
[DONE] Apply ad hoc detector, with 100 thresholds and limits on duration, to Tseep and Thrush ODFs, on 6 full night recordings. Export peak times as 26100=1200 CSV files. Parallelize over ODFs (2), units (6), and groups of 10 thresholds (10).
-
[DONE] Run clip suppressor on all CSV files. Export post-processed peak times as 26100=1200 CSV files.
-
[DONE] Merge Thrush and Tseep predictions for every unit and every threshold, both with and without clip suppressor. It results in 26100=1200 CSV files.
-
[DONE] Export metrics (n_selected, TP, FP, FN, precision, recall, F) for all 6 units, 3 detectors (Tseep, Thrush and both), and 10 tolerances, both with and without clip suppressor. It results in 6310*2=360 CSV files.
-
[DONE] Compute global metrics (precision, recall, and F) across all 6 units and 10 tolerances. Store in 6 CSV files, 3 without suppressor, 3 with suppressor.
C. ICASSP model
-
[DONE] Generate BirdVox-70k dataset. Parallelize over units (6).
-
[DONE] Generate JAMS metadata. Parallelize over units (6).
-
[DONE] Augment audio data: 33 augmentations. Parallelize over augmentations (33).
-
[DONE] Store augmented audio into 6*33=198 HDF5 containers. Parallelize over units (6) and augmentations (33).
-
[DONE] Compute log-mel-spectrograms of augmented audio, store into 6*33=198 HDF5 containers. Parallelize over units (6) and augmentations (33).
-
[DONE] Compute log-mel-spectrograms of full night, store into 6 HDF5 containers. Parallelize over units (6).
-
[DONE] Train ICASSP convnet on BirdVox-70k with augmentation, 10 trials. Export 6*10=60 Keras models. Parallelize over folds (6) and trials (10).
-
[DONE] For every fold unit (6), every prediction unit in validation set and test set (3), export BirdVox-70k predictions as CSV files. Parallelize over folds (6), trials (10), and prediction units (3).
-
[DONE] For every fold unit (6), every mode (validation and test), compute metrics with data augmentation.
-
[DONE] For every fold (6), select the 5 trials that achieve the best validation accuracy, along with the corresponding threshold. For every unit, export best five trials, per-trial threshold, and per-trial metrics (n_selected, TP, FP, FN, TPR, TNR, accuracy, precision, recall, and F-measure) in 6 CSV files. Parallelize over units (6).
-
[DONE] For every possible combination of successive trials (5**6=15625), compute global metrics (n_selected, TP, FP, FN, TPR, TNR, accuracy, precision, recall, and F-measure) over the test set, with data augmentation.
-
[DONE] Train ICASSP convnet on BirdVox-70k without augmentation, 10 trials. Export 6*10=60 Keras models. Parallelize over folds (6) and trials (10).
-
[DONE] For every fold unit (6), every prediction unit in validation set and test set (3), export BirdVox-70k predictions as CSV files without data augmentation. Parallelize over folds (6), trials (10), and prediction units (3).
-
[DONE] For every fold unit (6), every mode (validation and test), compute metrics without data augmentation.
-
[DONE] Make a notebook displaying the quantiles of accuracy, both with and without data augmentation. Compute AUC and AUPRC.
-
[DONE] Investigate the locations of 20% of false detections.
-
[DONE] Train ICASSP convnet with "all but noise" augmentation.
-
[DONE] Predict "all but noise" on clips.
-
[IN PROGRESS] Predict "all but noise" on full nights.
-
Threshold "all but noise" on full nights.
D. Snowball on UrbanSound-8K
-
[DONE] Find padding heuristics (repeat vs zero).
-
[DONE] Augment UrbanSound-8K dataset.
-
[DONE] Compute scattering transform of augmented audio.
-
[DONE] Define snowball model and pescador generator.
-
[DONE] Train snowball convnet, one trial. Parallelize across folds (10).
-
[DONE] Compute snowball predictions. Parallelize across test folds (10) and prediction folds (2).
-
[DONE] Get validation accuracy, validation MRR, test accuracy, and test MRR, of snowball for one trial.
E. Cross-validated spherical k-means model
-
[DONE] Compute log-mel-spectrograms on 70k recordings. Parallelize across units (6).
-
[DONE] Train PCA on BirdVox-70k clips without augmentation. Save biases and PCA matrices. Parallelize across units (6).
-
[DONE] Train SKM on BirdVox-70k clips. Parallelize across units (6) and trials (10).
-
[DONE] Train SVM on BirdVox-70k. Parallelize across units (6), trials (10), and log10C (5).
-
[DONE] Compute predictions on 6 full night recordings.
-
[DONE] Postprocessing: peak-picking, thresholding, event matching.
-
[DONE] Export metrics (n_selected, TP, FP, FN, precision, recall, F) for all 6 units, 10 tolerances, and 100 thresholds, to 6*10=60 CSV files.
-
[DONE] Compute global metrics (n_selected, TP, FP, FN, precision, recall, F and AUPRC) across all 6 units and 10 tolerances. Store in 1 CSV file.
F. Spectral flux
-
[DONE] Run spectral flux on 6 full night recordings. Export into 6 HDF5 containers by chunks. Parallelize over units.
-
[DONE] Apply spectral flux detector, with 100 thresholds, to spectral flux ODF, on 6 full night recordings. Export peak times as 6*100=600 CSV files. Parallelize over units (6) and groups of 10 thresholds (10).
-
[DONE] Export metrics (n_selected, TP, FP, FN, precision, recall, F) for all 6 units and 10 tolerances. It results in 6*10=60 CSV files.
-
[DONE] Compute global metrics (precision, recall, and F) across all 6 units and 10 tolerances. Store in 60 CSV files.
I. Dataset release
-
[DONE] Write README.
-
[DONE] Have consistent columns across units.
-
[DONE] Publish on Zenodo.
-
Bugfix latitude and longitude.
-
Add a text file with location of clips.
-
Publish on Zenodo under new name.
J. Per-channel energy normalization
-
[DONE] Compute PCEN on clips with augmentation. Parallelize over units (6) and instanced augmentations (31).
-
[DONE] Compute PCEN on full night data. Parallelize over units (6).
-
[DONE] Train PCEN convnets without augmentation. Parallelize over units (6) and trials (10).
-
[DONE] Predict PCEN convnets without augmentation on clips. Parallelize over units (6) and trials (10).
-
[DONE] Predict PCEN convnets without augmentation on full night data. Parallelize over units (6) and trials (10).
-
[DONE] Threshold PCEN convnet predictions without augmentation on full night data. Parallelize over units (6) and trials (10).
-
[DONE] Train PCEN convnets with augmentation. Parallelize over units (6) and trials (10).
-
[DONE] Predict PCEN convnets with augmentation on clips. Parallelize over units (6) and trials (10).
-
[IN PROGRESS] Predict PCEN convnets with augmentation on full night data. Parallelize over units (6) and trials (10).
-
[PENDING 9] Threshold PCEN convnet predictions with augmentation on full night data. Parallelize over units (6) and trials (10).
-
[DONE] Train PCEN convnets with augmentation excepting noise. Parallelize over units (6) and trials (10).
-
[DONE] Predict PCEN convnets with augmentation excepting noise on clips. Parallelize over units (6) and trials (10).
-
[IN PROGRESS] Predict PCEN convnets with augmentation excepting noise on full nights. Parallelize over units (6) and trials (10).
-
[PENDING 13] Threshold PCEN convnet predictions with augmentation excepting noise on full night data. Parallelize over units (6) and trials (10).
K. Dynamic filter networks
-
[DONE] Compute clip background logmelspec summaries. Parallelize over units (6) and instanced augmentations (31).
-
[DONE] Compute full night background logmelspec summaries. Parallelize over units (6).
-
[DONE] Train adaptive threshold convnets without data augmentation. Parallelize over units (6) and trials (10).
-
[DONE] Train adaptive threshold convnets with data augmentation excepting noise. Parallelize over units (6) and trials (10).
-
[DONE] Train adaptive threshold convnets with data augmentation. Parallelize over units (6) and trials (10).
-
[DONE] Train dot-product convnets with data augmentation excepting noise. Parallelize over units (6) and trials (10).
-
[DONE] Train dot-product convnets with data augmentation. Parallelize over units (6) and trials (10).
-
[DONE] Train NTT convnets without data augmentation. Parallelize over units (6) and trials (10).
-
[DONE] Train NTT convnets with data augmentation excepting noise. Parallelize over units (6) and trials (10).
-
[DONE] Train NTT convnets with data augmentation. Parallelize over units (6) and trials (10).
-
[DONE] Predict NTT convnets on clips without augmentation. Parallelize over units (6) and trials (10).
-
[DONE] Predict NTT convnets on clips with augmentation excepting noise. Parallelize over units (6) and trials (10).
-
[DONE] Predict NTT convnets on clips with augmentation. Parallelize over units (6) and trials (10).
-
[DONE] Predict adaptive threshold convnets on clips without augmentation. Parallelize over units (6) and trials (10).
-
[DONE] Predict adaptive threshold convnets on clips with augmentation excepting noise. Parallelize over units (6) and trials (10).
-
[DONE] Predict adaptive threshold convnets on clips with augmentation. Parallelize over units (6) and trials (10).
-
[DONE] Predict NTT convnets on full nights without augmentation. Parallelize over units (6) and trials (10).
-
[DONE] Predict NTT convnets on full nights with augmentation excepting noise. Parallelize over units (6) and trials (10). (should run trials 1 to 9 included)
-
[DONE] Predict NTT convnets on full nights with augmentation. Parallelize over units (6) and trials (10). (should re-run trials 2 to 6 included + 039_aug-all_test-unit01_trial-7_predict-unit07)
-
[PENDING 17] Threshold NTT convnets on full nights without augmentation. Parallelize over units (6) and trials (10).
-
[PENDING 18] Threshold NTT convnets on full nights with augmentation excepting noise. Parallelize over units (6) and trials (10).
-
[PENDING 19] Threshold NTT convnets on full nights with augmentation. Parallelize over units (6) and trials (10).
-
Predict adaptive threshold convnets on full nights without augmentation. Parallelize over units (6) and trials (10).
-
Predict adaptive threshold convnets on full nights with augmentation excepting noise. Parallelize over units (6) and trials (10).
-
Predict adaptive threshold convnets on full nights with augmentation. Parallelize over units (6) and trials (10).
-
Threshold adaptive threshold convnets on full nights without augmentation. Parallelize over units (6) and trials (10).
-
Threshold adaptive threshold convnets on full nights with augmentation excepting noise. Parallelize over units (6) and trials (10).
-
Threshold adaptive threshold convnets on full nights with augmentation. Parallelize over units (6) and trials (10).
L. Dynamic filter networks with PCEN
-
[DONE] Compute clip background PCEN summaries. Parallelize over units (6) and instanced augmentations (31).
-
[DONE] Compute full night background PCEN summaries. Parallelize over units (6).
-
[DONE] Train NTT PCEN convnets without data augmentation. Parallelize over units (6) and trials (10).
-
[DONE] Train NTT PCEN convnets with data augmentation excepting noise. Parallelize over units (6) and trials (10).
-
[DONE] Train NTT PCEN convnets with data augmentation. Parallelize over units (6) and trials (10).
-
[DONE] Train PCEN convnets with adaptive threshold, without data augmentation. Parallelize over units (6) and trials (10).
-
[DONE] Train PCEN convnets with adaptive threshold, with augmentation excepting noise. Parallelize over units (6) and trials (10).
-
[DONE] Train PCEN convnets with adaptive threshold, with augmentation. Parallelize over units (6) and trials (10).
-
[DONE] Predict NTT PCEN convnets with data augmentation on clips. Parallelize over units (6) and trials (10).
-
[READY] Predict NTT PCEN convnets without data augmentation on full nights. Parallelize over units (6) and trials (10).
-
[READY] Predict NTT PCEN convnets with data augmentation excepting noise on full nights. Parallelize over units (6) and trials (10).
-
[READY] Predict NTT PCEN convnets with data augmentation on full nights. Parallelize over units (6) and trials (10).
-
[PENDING 12] Threshold NTT PCEN convnets without data augmentation on full nights. Parallelize over units (6) and trials (10).
-
[PENDING 13] Threshold NTT PCEN convnets with data augmentation excepting noise on full nights. Parallelize over units (6) and trials (10).
-
[PENDING 14] Threshold NTT PCEN convnets with data augmentation on full nights. Parallelize over units (6) and trials (10).
-
[READY] Predict PCEN convnets with adaptive threshold without data augmentation on clips. Parallelize over units (6) and trials (10).
-
[READY] Predict PCEN convnets with adaptive threshold with data augmentation excepting noise on clips. Parallelize over units (6) and trials (10).
-
[READY] Predict PCEN convnets with adaptive threshold with data augmentation on clips. Parallelize over units (6) and trials (10).
-
[READY] Predict PCEN convnets with adaptive threshold without data augmentation on full nights. Parallelize over units (6) and trials (10).
-
[READY] Predict PCEN convnets with adaptive threshold with data augmentation excepting noise on full nights. Parallelize over units (6) and trials (10).
-
[READY] Predict PCEN convnets with adaptive threshold with data augmentation on full nights. Parallelize over units (6) and trials (10).