File size: 3,320 Bytes
666d084 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 | ---
language:
- en
license: mit
tags:
- bioacoustics
- audio-classification
- bird-sound-identification
- pytorch
- vision-transformers
- protoclr
- umap
- hdbscan
pipeline_tag: audio-classification
library_name: pytorch
metrics:
- accuracy
---
# 🦅 Edge-Optimized Bioacoustic Atlas & Real-Time Avian Classifier
[](https://colab.research.google.com/drive/1EL5VS_vAKvojPf5UPuQVFbK5gkgP51hB?usp=sharing)
[](https://huggingface.co/sukriramli/tiny-bird-diffusion)
[](https://pytorch.org/)
[]()
[](https://opensource.org/licenses/MIT)
A decoupled, ultra-lightweight machine learning pipeline designed for low-latency edge deployment, automated avian species tracking, and interactive audio streaming.
By leveraging **Prototypical Contrastive Learning (ProtoCLR)** backbones combined with advanced topological manifold compression (**UMAP & HDBSCAN**), this system projects complex audio waveforms onto a dense, 2D geometric map. The resulting production architecture handles **168 unique biological species** across **149 autonomous eco-acoustic clusters** natively in a client browser window with sub-second latency—completely bypassing the need for compute-heavy cloud inferencing heads.
---
## 🛠️ The System Architecture Problem & Our Solution
### Our Decoupled Geometric Solution
This repository implements a **decoupled mathematical pattern**. Heavy feature extraction is processed upfront. The complex high-dimensional latent space is then permanently compressed into a frozen geometric lookup coordinate plane. The client device only runs low-compute spatial distance algorithms, achieving zero-lag edge inference.
---
## 🔬 Core Engineering Pillars
### 1. High-Ratio Manifold Compression
Instead of forcing edge hardware to hold dense classification layer weights, we isolate the 512-dimensional floating-point latent vectors generated by the transformer. We utilize **UMAP (Uniform Manifold Approximation and Projection)** to topology-map this high-dimensional array down to a highly constrained 2D coordinate vector (X, Y). This slashes the database RAM footprint by over **99%** while preserving semantic biological boundaries.
### 2. Acoustic Domain Shift Mitigation (augment.py)
Pre-trained foundation models are typically trained on pristine, studio-grade wildlife audio recordings, causing them to fail frequently in noisy consumer spaces. To bridge this gap, our data preparation pipeline routes clean data shards through a custom acoustic corruption environment mimicking real-world conditions.
---
## 📁 Modular Codebase Layout
* `augment.py`: Digital Signal Processing (DSP) environment warping functions (Noise, Echo, Muffling filters).
* `pipeline.py`: Low-latency engineering pipeline managing model configuration and UMAP coordinate projection.
* `api.py`: Clean, production-ready prediction endpoint designed for real-time app integration.
|