What is model quantization?

Model quantization is the process of converting the weights and activations of a neural network from a floating-point format (e.g., float32) to a more compact format with fewer bits (e.g., int8 or int16). The goal of quantization is to reduce the model size and speed up its operation, especially on resource-constrained devices (mobile phones, embedded systems).

Advantages of quantization:

• Reduced memory footprint required to store the model.
• Faster computations by using simpler integer operations.
• Lower energy consumption.

Disadvantages:

• Possible loss of model accuracy due to rounding and reduced bit width.

Example: when quantizing weights from float32 to int8, the weight values are scaled and rounded, allowing the use of 8-bit operations instead of 32-bit.

Popular libraries like TensorFlow Lite or PyTorch have built-in tools for model quantization, which can perform this process automatically or manually.