A trainable soft-tanh quantizer with learned steepness and clipping improves 1-4 bit network accuracy and yields fast ARM kernels.
Deep learning with low precision by half-wave gaussian quantization
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2019 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Differentiable Soft Quantization: Bridging Full-Precision and Low-Bit Neural Networks
A trainable soft-tanh quantizer with learned steepness and clipping improves 1-4 bit network accuracy and yields fast ARM kernels.