qat
9 memberslucid.nn.qat`lucid.nn.qat` — quantization-aware training modules.
Float-trainable layers that fake-quantize their weights and outputs (via the
straight-through estimator) so a network learns to be robust to int8
inference. Produced by lucid.quantization.prepare_qat; turned into
real quantized inference layers by lucid.quantization.convert.
Classes
Conv1d3 methodsQuantization-aware 1-D convolution — trainable float kernel, fake-quant per forward.
Conv2d3 methodsQuantization-aware 2-D convolution — trainable float kernel, fake-quant per forward.
Conv3d3 methodsQuantization-aware 3-D convolution — trainable float kernel, fake-quant per forward.
ConvReLU1d2 methodsQuantization-aware fused 1-D conv + ReLU — trainable, fake-quant per forward.
ConvReLU2d2 methodsQuantization-aware fused 2-D conv + ReLU — trainable, fake-quant per forward.
ConvReLU3d2 methodsQuantization-aware fused 3-D conv + ReLU — trainable, fake-quant per forward.
Linear3 methodsQuantization-aware Linear — trainable float weight, fake-quant every forward.
LinearReLU2 methodsQuantization-aware fused Linear + ReLU — trainable, fake-quant every forward.
Embedding3 methodsQuantization-aware embedding — trainable float table, fake-quant every forward.