Supplementary repository: code, configs, and evaluation artifacts for 'Evaluating In-Corpus SimCLR Pretraining for Underwater Acoustic Target Recognition Under Recording-Disjoint Splits'
Year: 2026
Type: supplementary-materials
Abstract: Code, configs, and evaluation artifacts accompanying the preprint 'Evaluating In-Corpus SimCLR Pretraining for Underwater Acoustic Target Recognition Under Recording-Disjoint Splits' (Preprints 2026, DOI: 10.20944/preprints202609.0999.v2). Contents: (1) all 45 per-seed runs of the main analysis (3 corpora x initialization-protocol cells x seeds 42/43/44, one pretrained encoder) with per-run configs, training histories, and summary JSONs; (2) encoder-seed sensitivity for three independently pretrained SimCLR encoders plus the corpus-balanced pretraining control, including the balanced-sampling manifest and aggregate_seeds.json; (3) Section 4.4 extensions: 100/200-epoch budget ladder, channel-varying augmentation control, VICReg objective control, MMSI/hull-disjoint evaluation, and the recording-identity probe; (4) the fixed-label-set macro-F1 recompute (165 runs) with authoritative per-run summaries and per-segment prediction exports; (5) frozen recording-disjoint split manifests, preprocessing manifests (recording-ID/group mapping), seed lists, and experiment configurations; (6) aggregation, training, evaluation, and figure-generation code plus an environment file. Raw audio, model checkpoints, and feature caches are excluded (source dataset licenses; regenerable with the deposited code and frozen splits); VTUAD audio is available at 10.5281/zenodo.22274549. See README.md inside the zip for the full reproduction guide.
Institution: Wuhan University of Science and Technology
Language: English
Subfield: Oceanography
Field: Earth and Planetary Sciences
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213785044