huggingface

Classes

WaveformTempoNet(model_name[, dropout, ...])

Any HuggingFace waveform encoder with a tempo regression head.
class WaveformTempoNet(model_name, dropout=0.1, freeze_backbone=True)[source]

Bases: Module

Any HuggingFace waveform encoder with a tempo regression head.

Loads any model via AutoModel.from_pretrained, mean-pools its last hidden state, and feeds it through a small regression head. Expects raw waveforms at the sample rate required by the chosen backbone (typically 16 kHz). The backbone is frozen by default.

Input: (B, 1, T) Output: (B,)

Parameters:

  • model_name (str) – HuggingFace model identifier.

  • dropout (float) – Dropout probability in the regression head.

  • freeze_backbone (bool) – Whether to freeze the backbone weights.

forward(wav)[source]
Parameters:

wav (Tensor)

Return type:

Tensor