A Hessian-guided packing scheme for post-training quantization that jointly reconstructs groups of blocks and adds pack-based mixed precision, claiming state-of-the-art low-bit accuracy.
Training data-efficient image transformers & distillation through attention
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
REJECT 1representative citing papers
citing papers explorer
-
Pack-PTQ: Advancing Post-training Quantization of Neural Networks by Pack-wise Reconstruction
A Hessian-guided packing scheme for post-training quantization that jointly reconstructs groups of blocks and adds pack-based mixed precision, claiming state-of-the-art low-bit accuracy.