SKETCH-STYLUS

Introduction

This is a hand written and optimized machine learning library written in Rust using Arbitrum's Stylus SDK.

This library implements the following:

• K Nearest Neighbours
• Logistic Regression
• Digit Recognizer ML Model
• Machine Learning Math library

Feautures

• For avoiding the loss of precision in weights and biases, scaling has been implemented.
• Minimal use of std library.
• The size of WASM contract and deployment gas for each is listed below.

KNN

The k-nearest neighbors (KNN) algorithm is a non-parametric, supervised learning classifier, which uses proximity to make classifications or predictions about the grouping of an individual data point.

• Uncompressed size - 48.4 KB
• Compressed size - 15.6 KB
• Deployment gas - 6881292
• Exported ABI

interface IKNN {
   function trainPredict() external;

   function setK(uint256 val) external;

   function getK() external view returns (uint256);
}

Logistic Regression

Logistic regression estimates the probability of an event occurring, such as voted or didn’t vote, based on a given data set of independent variables. This type of statistical model (also known as logit model) is often used for classification and predictive analytics.

• Uncompressed size - 71.1 KB
• Compressed size - 24.0 KB
• Deployment gas -
• Exported ABI

interface ILogReg {
    function train(int128[][] memory x_train, int128[][] memory y_train, uint128 iterations, int128 lr) external returns (bool);
}

ML-math

This is a library for mathematical functions required within the different implementations of machine learning models. This includes various functions such as sigmoid, sqrt, signum, softmax and other matrix operations such as matrix_mul, dot_product, matrix_sub, matrix_sum etc.

• Uncompressed size - 45.2 KB
• Compressed size - 13.1 KB
• Deployment gas - 10062745
• Exported ABI

interface IMlMath {
   function reluDerive(int128[][] memory m) external pure returns (int128[][] memory);

   function elementSumRow(int128[][] memory z) external pure returns (int128[][] memory);

   function scalarDiv(int128[][] memory mat, int128 scalar) external pure returns (int128[][] memory);

   function elementwiseMul(int128[][] memory m1, int128[][] memory m2) external pure returns (int128[][] memory);

   function transpose(int128[][] memory ori) external pure returns (int128[][] memory);

   function softmax(int128[][] memory z) external pure returns (int128[][] memory);

   function relu(int128[][] memory z) external pure returns (int128[][] memory);

   function dotProduct(int128[][] memory m1, int128[][] memory m2) external pure returns (int128[][] memory);

   function sum(int128[][] memory m1, int128[][] memory m2) external pure returns (int128[][] memory);

   function oneHot(int128[][] memory y) external pure returns (int128[][] memory);

   function scalarMul(int128[][] memory mat, int128 scalar) external pure returns (int128[][] memory);
}

Digit Recognizer

Digit Recognition is a computer vision technique to predict the correct digits from pixel values of images. We have employed MNIST dataset for training.

• Uncompressed size - 52 KB
• Compressed size - 16.5 KB
• Deployment gas - 12623680
• Exported ABI

interface ITraining {
   function trainPredict(int128[][] memory x_train, int128[][] memory y_train) external returns (bool);
}