A training-time method that reshapes weights and activations toward uniform distributions and splits filters into groups with separate scales, yielding strong low-bit quantization results on classification, detection, and segmentation.
Title resolution pending
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
-
GDRQ: Group-based Distribution Reshaping for Quantization
A training-time method that reshapes weights and activations toward uniform distributions and splits filters into groups with separate scales, yielding strong low-bit quantization results on classification, detection, and segmentation.