FrogNet Research 0.1.0 — Perch 2
A provenance-locked binary frog_present specialist head over frozen 1,536-dimensional Google Perch 2 embeddings.
This repository contains only the FrogNet model family. It shares the frozen Perch 2 feature contract with other field heads and does not use BirdNET logits.
Status: public noncommercial research artifact; this exact checked bundle is privately deployed as a review-only field head. This is not a species classifier, a calibrated probability model, or evidence of ecological occurrence without human review.
This repository is the canonical FrogNet research-family source for the checked model bundle and evidence summaries. The historical frognet-field-probe-v0.1.0 tag and internal model slug are retained as provenance from the archived combined registry. The package contains no source audio, coordinates, private observation locations, uploader identities, private manifests, credentials, private paths, or arbitrary Python serialization. Public dataset identifiers and regional benchmark names are retained where needed for provenance.
Release identity
1Upstream release: frognet-field-probe-v0.1.0
2Model slug: frognet-dev4-field-probe
3Bundle ID: db54359b42526010a2e7782837d2ff8a5e7d98beeebba9214211a2ee83572fa8
4Source commit: 55d42a8c96584e09504b9b61127b0dd9e069cdfc
5Source archive: bc82746ca83b305591073d1aeba8a4598e9860953d9a451d087695ac927d29e1
The Hub package exposes the exact checked JSON/NPZ bundle from that release rather than republishing its outer archive. The runtime verifies the bundle-member checksums and recomputes the bundle ID before scoring. No pickle or joblib loader is required. Renaming the Hub repository does not alter the model bytes, bundle ID, threshold, or deployed inference contract.
Inference contract
| Field | Value |
|---|
| Input audio to Perch | 5 seconds, 32 kHz, mono |
| Input to FrogNet | one 1,536-dimensional Perch 2 embedding |
| Output | uncalibrated ranking score for frog_present |
| Locked field threshold | 0.95 |
| Serialization | checked JSON metadata plus NPZ numeric arrays |
| Perch model-tree SHA-256 | 3fb2d54b3e34534f1130052b25737e54bbb5ebfd340ec040d4510772b64c81ff |
The Perch weights are not included. FrogNet cannot score raw waveforms by itself.
Locked field-probe evidence
At threshold 0.95:
| Challenge lane | Activations | Boundary |
|---|
| Private local frog windows | 1,253 / 1,308 | narrow fixed-recorder and seasonal coverage |
| Permissively licensed regional iNaturalist frog recordings | 106 / 128 | broader taxa and observers, still incomplete |
| Locked insect, chicken, cat, dog, SmartEars, and laying-hen controls | 0 / 3,781 | challenge result, not a general field false-positive rate |
The highest tested confound score was 0.9216794407 on one dog recording, below the locked threshold. Candidate-training grouped test average precision was 0.9645.
These are challenge-set observations. They do not establish prevalence-adjusted precision, field recall, abundance, species identity, or performance at a new recorder or site.
Operational relationship
In the private field system, each 15-second recording is divided into three non-overlapping five-second windows. Perch produces one embedding per window. InsectNet, ChickenNet, and FrogNet score the same embedding independently, while BirdNET remains a separate whole-recording evidence lane. Every output is retained as a model assertion for review rather than a confirmed biological observation.
The operational repository, raw field audio, sensor details, and private review records are intentionally not published here.
Verify the Hub package
1sha256sum --check SHA256SUMS
2cd bundle && sha256sum --check SHA256SUMS
The root inventory covers every public payload file except itself, including the nested inventory. The nested inventory is the original runtime bundle inventory from the tagged prerelease.
Included files
bundle/model.json — checked model metadata and inference contractbundle/weights.npz — numeric arrays; load with allow_pickle=Falsebundle/SHA256SUMS — original bundle-member checksumsrelease/frognet-field-probe-v0.1.0.manifest.json — release identity and rights lanesrelease/field_smoke_report.json — frozen challenge summaries and embedding hashesrelease/source-release.json — upstream tag, commit, and asset identitiesSHA256SUMS — complete Hub payload inventory
Known limits
- Most private local positives came from one late-May chorus at one stationary recorder.
- A strict threshold can miss faint, distant, or mixed frog calls.
- Regional taxa, observers, devices, weather, and seasons remain incompletely represented.
- Existing recordings were not automatically backfilled when the field probe was installed.
- The score is suitable for ranking and review triage, not unattended ecological counts.
Rights and provenance
The training matrix used three rights lanes: private diagnostic material, permissively licensed public material, and research/noncommercial material derived from ESC-50. Because noncommercial source material contributed to training, this prerelease is distributed under CC BY-NC-SA 4.0 for noncommercial research and field evaluation, subject to upstream terms.
No upstream audio is redistributed. This release does not represent the full training corpus as permissively licensed. See the release manifest for the retained hashes and exact public/private boundary.
Source
Load the checked deployed bundle
1from pathlib import Path
2from frognet.bundle import load_bundle, score_embedding
3
4bundle = load_bundle(Path('bundle'))
5raw_score, review_score = score_embedding(bundle, perch2_embedding)
The input is one finite 1,536-dimensional Perch 2 embedding. Do not feed BirdNET logits or waveforms directly into this head.
