VLCQ
VLCQ is a post-training quantization method based on variable-length coding. It abandons the fixed-point integer coding constraint and represents quantized values with variable-length codes, balancing high compression rates and accuracy without requiring training data or fine-tuning.
Timeline (last 90 days)
The paper "VLCQ: Post-training quantization for deep neural networks using variable length coding" proposes a method for post-training quantization using variable length coding, which does not require access to training data or model fine-tuning.
VLCQ is a fully data-free PTQ that does not require calibration data, making it suitable for scenarios where training data cannot be obtained due to privacy, copyright, or collection costs.
VLCQ represents quantized values through variable length coding, so that regions with dense value distributions can use finer codes and sparse regions can use shorter codes.
VLCQ was experimented on three types of tasks: image classification, segmentation, and object detection; in per-channel quantization it surpassed FP32 precision and PWLQ on most models, achieving up to 6x compression compared to FP32 and up to 1.7x compression compared to PWLQ.
When compression priority is higher and limited precision loss is tolerated, VLCQ can achieve up to 12.25x compression compared to FP32, with performance degradation kept within 4%.
In per-tensor quantization, VLCQ performs comparably to Data-Free Quantization (DFQ), but provides greater flexibility across different tasks and models.
Neither PWLQ nor DFQ breaks out of the premise of fixed code length, whereas VLCQ changes the encoding mechanism itself.
Variable length coding brings increased encoding and decoding complexity, and decoding overhead may offset storage gains; the paper does not discuss the degree of hardware friendliness in detail.
All Facts (8)
The paper "VLCQ: Post-training quantization for deep neural networks using variable length coding" proposes a method for post-training quantization using variable length coding, which does not require access to training data or model fine-tuning.
50%UnverifiedVLCQ is a fully data-free PTQ that does not require calibration data, making it suitable for scenarios where training data cannot be obtained due to privacy, copyright, or collection costs.
50%UnverifiedVLCQ represents quantized values through variable length coding, so that regions with dense value distributions can use finer codes and sparse regions can use shorter codes.
50%UnverifiedVLCQ was experimented on three types of tasks: image classification, segmentation, and object detection; in per-channel quantization it surpassed FP32 precision and PWLQ on most models, achieving up to 6x compression compared to FP32 and up to 1.7x compression compared to PWLQ.
50%UnverifiedWhen compression priority is higher and limited precision loss is tolerated, VLCQ can achieve up to 12.25x compression compared to FP32, with performance degradation kept within 4%.
50%UnverifiedIn per-tensor quantization, VLCQ performs comparably to Data-Free Quantization (DFQ), but provides greater flexibility across different tasks and models.
50%UnverifiedNeither PWLQ nor DFQ breaks out of the premise of fixed code length, whereas VLCQ changes the encoding mechanism itself.
50%UnverifiedVariable length coding brings increased encoding and decoding complexity, and decoding overhead may offset storage gains; the paper does not discuss the degree of hardware friendliness in detail.
50%