blocks

birdnet_stm32.models.blocks

Additional model building blocks for audio classification.

Provides channel alignment for the NPU. Attention pooling lived here too and was removed in 1.2.0; the head uses global average pooling.

Squeeze-and-excite and inverted-residual blocks were removed: A1 showed that backbone could not reach the release parity gates under INT8 (0.739 mean cosine, unrecoverable by QAT), and every shipped model uses plain depthwise separable blocks instead.