INT8 Quantization of Object Detectors
Type
Work project
From
UncannyVision internship
Headline result
0.5% drop after INT8 quantization
Topics
edge deployment, object detection
Write-up
Description: Quantization of object detectors by leveraging dynamic ReLU activation function for INT-8 calibration and quantizing the weights of the model with the help of NVidia’s TensorRT’s algorithm(standalone implementation) . Dynamic ReLU viz. QReLU helped us to get only 0.5% drop after quantization and 1.2% increase in the overall mAP during training than the baseline.
Design of the quantization process.
Flowchart of INT8 Quantization
