From 11e8e31c049bf674d604bcebb89aeb17acf8abe4 Mon Sep 17 00:00:00 2001 From: Jayson Lin Date: Tue, 31 Oct 2023 13:48:19 -0400 Subject: [PATCH] typo fix --- optimizations.qmd | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/optimizations.qmd b/optimizations.qmd index cb38f24f..02c33084 100644 --- a/optimizations.qmd +++ b/optimizations.qmd @@ -506,7 +506,7 @@ Entropy: Use KL divergence to minimize information loss between the original flo Percentile: Set the range to a percentile of the distribution of absolute values seen during calibration. For example, 99% calibration would clip 1% of the largest magnitude values. -![Histogram of input activatsions to layer 3 in ResNet50 and calibrated ranges (@intquantfordeepinf).](images/efficientnumerics_calibrationcopy.png) +![Histogram of input activations to layer 3 in ResNet50 and calibrated ranges (@intquantfordeepinf).](images/efficientnumerics_calibrationcopy.png) Importantly, the quality of calibration can make a difference between a quantized model that retains most of its accuracy and one that degrades significantly. Hence, it's an essential step in the quantization process. When choosing a calibration range, there are two types: symmetric and asymmetric.