insectnet / README.md
TheVortexProject's picture
Align historical baseline with living Perch 2 deployment
a303fc3 verified
|
Raw
History Blame Contribute Delete
7.21 kB
---
license: cc-by-nc-sa-4.0
library_name: sklearn
tags:
- bioacoustics
- audio-classification
- birdnet
- edge-ai
- research
- non-commercial
datasets:
- academic-datasets/InsectSet459
- ESC-50
---
# InsectNet v0.1.0
InsectNet v0.1.0 is a preserved research classifier for scoring insect- and amphibian-related acoustic classes from frozen BirdNET v2.4 output logits.
This repository preserves one canonical v0.1 release identity and also carries separately
versioned Perch 2 research candidates. It does **not** publish live sensor addresses,
deployment credentials, exact collection locations, private media, or an unattended
capture service.
Public discovery and versioned downloads are organized in the
[The Vortex Project — Multi-Taxa Bioacoustics](https://huggingface.co/collections/TheVortexProject/the-vortex-project-multi-taxa-bioacoustics-6a6d92a9e0938f507e931388)
Hugging Face collection. GitHub remains the source, tests, and release-engineering surface.
## Release identity
```text
Release: insectnet-v0.1.0
Artifact: src/insectnet/data/classifier.joblib
SHA-256: 5e6ecfc68d78a2cf2e9e9e47da5cb58d696e8de354fd620cfcccc5db9da48702
Bytes: 474,892
Status: historical research reference
```
The artifact is byte-for-byte preserved from the original v0.1 field prototype.
## Model contract
```text
Audio window: 3.0 seconds
Sample rate: 48,000 Hz mono
Backbone: BirdNET v2.4 FP16 TFLite
Feature space: 6,522 BirdNET output logits
Classifier: StandardScaler → OneVsRest LogisticRegression
Serialization: scikit-learn 1.8.0
Output semantics: independent per-class probabilities
```
Class order:
1. `background`
2. `bee`
3. `cicada_drone`
4. `cricket_katydid`
5. `frog`
6. `grasshopper`
The model file does not embed thresholds, a version number, or its original training snapshot. Those omissions are part of the preserved v0.1 record rather than silently reconstructed metadata.
## What is included
- the exact v0.1 model artifact;
- a machine-readable release manifest;
- artifact and feature-contract verification;
- offline scoring for precomputed BirdNET logit vectors;
- tests that enforce the model checksum, class order, feature dimension, and public privacy boundary;
- documented provenance and limitations.
## What is not included
- live capture or sensor-watching code;
- deployment scripts;
- device addresses or credentials;
- exact collection locations;
- raw or private field audio;
- claims of production readiness.
The original v0.1 live sidecar diverged from this public source during early field experiments and is not a current deployment target. It has been superseded operationally by the separately versioned Perch 2 specialist architecture; the preserved v0.1 artifact remains a historical BirdNET-logit reference.
## Verify the preserved artifact
From a source checkout:
```bash
uv run insectnet verify
```
Expected SHA-256:
```text
5e6ecfc68d78a2cf2e9e9e47da5cb58d696e8de354fd620cfcccc5db9da48702
```
## Score precomputed logits
InsectNet v0.1 expects one finite NumPy vector with shape `(6522,)` extracted from the declared BirdNET backbone:
```bash
uv run insectnet score logits.npy
```
The command returns one probability per declared class. These scores are model assertions for review, not confirmed biological observations.
> **Joblib safety:** joblib artifacts use Python pickle internally. Load only the artifact whose checksum matches the release manifest.
## Known limitations
- Later audits found high false-positive rates on some bird vocalizations.
- Bee and grasshopper had limited training coverage.
- The surviving metrics came from limited public-data evaluation and one private field site; they do not establish general production performance.
- Exact reproduction is blocked because the original per-record training snapshot is unavailable.
- The classifier relies on BirdNET logits and cannot score raw audio by itself.
- There is no validated automated-decision threshold policy in this release.
## Perch 2 research artifacts and field heads
Three provenance-locked specialist lines were trained and audited in July 2026. None
replaces the preserved BirdNET-logit v0.1 artifact. Each line has an explicit public
research boundary and a separate operational status. FrogNet's public mirror contains
the exact checked JSON/NPZ bundle used by its private review-only field head; the public
InsectNet and ChickenNet research artifacts must not be conflated with later private
field bundles.
| Candidate | Status | Key external result |
|---|---|---|
| [ChickenNet Research 0.1.0](models/chickennet-research-0.1.0-perch2/MODEL_CARD.md) | frozen public joblib not deployed; later private JSON/NPZ successor operates as a review-only field head | 33/42 broad-head hits on a locked iNaturalist chicken challenge; 10/1,308 candidate activations on a private local confound set |
| [InsectNet Research 0.2.0](models/insectnet-research-0.2.0-perch2/MODEL_CARD.md) | frozen public joblib not deployed; later private JSON/NPZ successor operates as a review-only field head | 11/26 broad activations on an untouched iNaturalist dog challenge |
| [FrogNet Research 0.1.0](models/frognet-research-0.1.0-perch2/MODEL_CARD.md) | public noncommercial research artifact; exact strict bundle deployed privately for review | 1,253/1,308 local frog-window activations and 0/3,781 tested confound activations at threshold `0.95` |
All three lines consume 1,536-dimensional Google Perch 2 embeddings from five-second,
32 kHz mono windows. They include exact model/data hashes, grouped split reports,
hierarchy contracts, source summaries, and challenge reports. They do not include Perch
weights or source audio.
## Training strategy
[`docs/PERCH2_TRAINING_STRATEGY.md`](docs/PERCH2_TRAINING_STRATEGY.md) records the
provenance-first design used for the Perch 2 specialist heads. The current FrogNet
research and operational contract is frozen in [`docs/FROGNET_FIELD_PROBE.md`](docs/FROGNET_FIELD_PROBE.md).
The dated architecture and artifact-versus-runtime boundary for all three living heads is
recorded in [`docs/LIVING_DEPLOYMENT.md`](docs/LIVING_DEPLOYMENT.md).
## Provenance
The surviving records identify these source families:
- **InsectSet459:** current dataset card states CC BY 4.0, with some source material CC0.
- **ESC-50:** CC BY-NC 3.0.
- **iNaturalist audio:** licenses vary per recording; the original per-record manifest is unavailable.
- **Private field negatives:** not redistributed.
See [`docs/PROVENANCE.md`](docs/PROVENANCE.md) and the release manifest for the exact surviving claims and gaps.
## Privacy and security
The public release intentionally omits exact collection location, network topology, account names, credentials, private paths, and raw evidence. See [`docs/SECURITY_AND_PRIVACY.md`](docs/SECURITY_AND_PRIVACY.md).
## License
The repository and preserved release are distributed under CC BY-NC-SA 4.0, subject to the licenses and terms of the upstream backbone and source media. Source-media rights vary; users are responsible for reviewing those upstream terms for their use case.
This provenance statement is not legal advice.