PERCH 2 PyTorch

PyTorch implementation of the complete PERCH 2 waveform model for bioacoustic classification and 1536-dimensional audio embeddings.

Only need embeddings? perch2-torch-embed has the same frontend and backbone without the label head, as a single torch.export file (47.6 MB instead of 408 MB). It loads with PyTorch alone, and comes with a small helper that returns avg-pooled, max-pooled and temporal embeddings for audio of any length.

Load From Hugging Face

import torch

model = torch.hub.load("janclemenslab/perch2_torch", "perch_v2").eval()
waveform = torch.zeros(5 * 32_000)  # Mono 32 kHz audio.

with torch.no_grad():
    outputs = model(waveform)

index = outputs["label"][0].argmax()
print(model.labels[index])

This call downloads perch_v2_torch.pt from this Hugging Face repository and caches it locally.

Inputs are mono, 32 kHz waveforms with shape (time,) or (batch, time). Audio longer than five seconds is processed in overlapping windows and pooled.

Outputs

• embedding: (batch, 1536) global embedding
• spatial_embedding: (batch, time, frequency, 1536) unpooled embedding
• spectrogram: log-mel spectrogram
• label: (batch, 14795) uncalibrated class logits

model.labels[index] maps a logit index to its embedded class name. Thresholds should be calibrated for the target data.

Reproducibility

The conversion code, demo notebook, and verification command are available in janclemenslab/perch2_torch.

Attribution And License

Derived from the Apache-2.0 PERCH ONNX model and Apache-2.0 wrice EfficientNet-B3 checkpoint. The class taxonomy comes from the Apache-2.0 PERCH model release.