Size of a quantized DNN

Hello,

I am using the deep network quantizer to quantize a deep CNN. I followed instructions given at https://www.mathworks.com/help/deeplearning/ref/deepnetworkquantizer-app.html . I used the Export button and selected Export Quantizer to export my quantized network. But when I compare the size of both networks (the original one before quantization and the quantized one stored in the dlquantizer object), I see they both have the same number of bytes in the workspace.

Is my way for checking the size of network correct and resonable?

The app shows 73% size reduction which I can not find when comparing the networks!

Any help os appreciated.

Answer 1

[Jacob Mathew]

Hey @Ali Al-Saegh,

Quantizing a neural network does not necessarily lead to reduction in the physical storage size needed to represent the model. The goal is to replace floating point computations with fixed point or look up table based activation functions as well as reducing the bits needed for activation function calculation. A detailed definition and workflows that go behind quantizing a network can be found in the documentation below:

https://www.mathworks.com/help/deeplearning/ug/quantization-of-deep-neural-networks.html

As such the goal is to reduce execution memory requirements and specialised hardware requirements allowing ML and DL models to run on lower power hardware