3 releases

0.1.5-alpha.0 Mar 25, 2023
0.1.4-alpha.0 Mar 3, 2023
0.1.3-alpha.0 Mar 1, 2023

#709 in Machine learning

38 downloads per month
Used in caffe2-histogram


31K SLoC



This Rust crate is a work-in-progress translation of the Caffe2 operator library from C++ to Rust. Some of the function bodies are still in the process of being translated. The crate contains operators for server quantization of neural networks, including fully connected and convolutional layers, elementwise operations, and pooling layers.

The mathematical ideas behind the operators include:

  • Quantization: transforming continuous-valued weights and activations into discrete values to reduce memory requirements and computational complexity.

  • Minimization of quantization error: finding the quantization parameters that minimize the error between the quantized and original values.

  • Low-precision arithmetic: performing computations with reduced numerical precision to further reduce memory requirements and computational complexity.

  • Normalization: scaling weights or activations to have unit variance or to be within a certain range.

  • Pooling: downsampling a feature map by applying a function (such as max or average) over non-overlapping regions of the map.

  • Elementwise operations: applying a function (such as addition or multiplication) to corresponding elements of two or more tensors.

Relevant Mathematical Equations:


The quantization function maps continuous values x to discrete values x_q:

x_q = round(x / delta) * delta,

where delta is the quantization step size.

Minimization of quantization error:

One approach to minimizing the quantization error is to choose the quantization step size delta that minimizes the mean squared error (MSE) between the quantized values x_q and the original values x:

delta = argmin(delta') {1/N * sum[(x - round(x / delta') * delta')^2]},

where N is the number of values being quantized.

Low-precision arithmetic:

Low-precision arithmetic operations involve computing with reduced numerical precision, such as using 8-bit integers instead of 32-bit floats.


One form of normalization is Batch Normalization, which scales and shifts activations by learned parameters, so that they have unit variance and are zero-centered:

y = (x - E[x]) / sqrt(Var[x] + eps) * gamma + beta,

where x is the input, y is the output, E[x] and Var[x] are the mean and variance of the input over the batch dimension, eps is a small constant for numerical stability, and gamma and beta are learned scaling and shifting parameters.


One form of pooling is max pooling, which downsamples a feature map by selecting the maximum value in non-overlapping regions of the map:

y_{i,j,k} = max_{u,v,w} x_{i+su, j+sv, k+s*w},

where x is the input feature map, y is the output feature map, s is the stride, and u, v, w are indices over the spatial dimensions.

Elementwise operations:

An example of an elementwise operation is addition, which adds corresponding elements of two tensors:

y = x_1 + x_2.

42> what is server quantization?

Server quantization is a technique for compressing and optimizing deep neural networks (DNNs) to run efficiently on servers, which are typically equipped with CPUs rather than GPUs or specialized hardware. The goal of server quantization is to reduce the memory and computational requirements of DNNs while maintaining or improving their accuracy. This is achieved by quantizing the weights and activations of the DNN to use lower-precision data types, which reduces the memory footprint and enables faster computations on CPUs.

Server quantization typically involves several steps, such as determining the optimal bit width for the quantized weights and activations, computing scaling factors to preserve the dynamic range of the data, and adjusting the network architecture to accommodate the quantization. Various techniques such as quantization-aware training, per-channel quantization, and sparsity are used to further optimize the quantized networks.

Server quantization is especially useful for large-scale DNNs deployed on cloud servers, which require efficient use of computing resources and energy consumption. It can also enable more widespread deployment of DNNs on devices with limited computational resources, such as smartphones and Internet of Things (IoT) devices.

43> how is this different than regular quantization?

Regular quantization is a technique used to reduce the memory requirements and increase the computational efficiency of neural networks by representing their parameters using fewer bits than the standard floating-point representation. In regular quantization, the parameters of the neural network are quantized to lower precision values, typically using fixed-point or integer arithmetic, which can lead to some loss of accuracy compared to the original floating-point representation.

Server quantization, on the other hand, is a more advanced form of quantization that is designed to minimize the accuracy loss associated with regular quantization. In server quantization, the quantization parameters are learned by minimizing a certain objective function, such as the Kullback-Leibler divergence between the distribution of the original weights and the distribution of the quantized weights. This learning process allows the quantization parameters to be tailored specifically to the weights of the neural network being quantized, leading to improved accuracy compared to regular quantization.

Server quantization can also include techniques such as weight clipping, which limits the range of the weights before quantization, and scaling, which applies a multiplication factor to the weights to ensure that they fall within the range of the quantization. These additional techniques can further improve the accuracy of the quantized model.

Overall, server quantization is a more advanced and sophisticated form of quantization that is designed to minimize the accuracy loss associated with regular quantization, making it a powerful tool for optimizing neural networks for deployment in production environments.

44> which other techniques apply to this quantization method?

Some other techniques that apply to server quantization include:

  1. L1 and L2 Minimization: These techniques involve minimizing the L1 or L2 norm of the difference between the original weights and the quantized weights. L1 minimization is often used for sparse models where many weights are zero, while L2 minimization is more commonly used for dense models.

  2. Kullback-Leibler (KL) Divergence Minimization: This technique involves minimizing the Kullback-Leibler divergence between the original weight distribution and the quantized weight distribution. This can help preserve the statistical properties of the original model.

  3. Histogram-based Quantization: This technique involves creating a histogram of the weight distribution and quantizing the weights based on the histogram. This can help ensure that the quantization levels are distributed more evenly across the weight range.

  4. Dynamic Quantization: This technique involves quantizing the weights dynamically at runtime, rather than pre-quantizing them during training. This can allow for more adaptive quantization and may result in better performance on some models.


~385K SLoC