ML2SC

ML2SC: Machine Learning to Smart Contract

Introduction

ML2SC is a primitive PyTorch to Solidity translator that can automatically translate multi-layer perceptron (MLP) models written in Pytorch to Solidity smart contract versions.

Pytorch Support

As of now ML2SC supports translating the following pytorch syntax

• classes inheriting the nn.Module
• init() and forward() member functions
• super(class_name, self).__init() constructor
• declaring nn.Linear layers in init(*)
• passing input along nn.Linear layers in forward(*)
• Applying an activation function on layer outputs in forward(*)
• Returning the activation output in forward(*)

Dev Environment

Install Ganache with npm install -g ganache-cli and Truffle with npm install -g truffle

Adding & Translating Models

src/data contaings training and testing data for ML models

To use ML2SC:

1. first clone the repository
2. define your pytorch models in src/models.py.
3. Add translate_model() calls in src/translate.py for your defined models
4. run src/translate.py. This will use our translator module to generate solidity contract code in contracts/classifiers/.

Example

Here is a sample input/output pair for interpreting a single layer perceptron.

Input a PyTorch to Solidity translator that can automati-

cally translate multi-layer perceptron (MLP) models written in Pytorch to Solidity smart contract versions

import torch
from torch import nn
import torch.nn.functional as F


class MLP_2L_1n(nn.Module):
    def __init__(self, input_dim):
        super(MLP_2L_1n, self).__init__()
        self.fc1 = nn.Linear(input_dim, 1)  # Input to Hidden Layer with 1 neuron
        self.fc2 = nn.Linear(1, 1)  # Hidden Layer to Output

    def forward(self, x):
        return torch.sigmoid(self.fc2(F.relu(self.fc1(x))))

Output

// SPDX-License-Identifier: UNLICENSED
pragma solidity >=0.4.22 <0.9.0;
import {SD59x18, convert, sd} from "../lib/prb-math/src/SD59x18.sol";

contract MLP_2L_1n {
    int256[][] public weights_layer1;
    int256[][] public weights_layer2;
    int256[][] public biases;
    int256[][] public training_data;
    int public correct_Count;

    //relu activation function
    function relu(SD59x18 x) public pure returns (SD59x18) {
        int256 zero = 0;
        SD59x18 zero_cvt = convert(zero);
        if (x.gte(zero_cvt)) {
            return x;
        }
        return zero_cvt;
    }

    function sigmoid(SD59x18 x) public pure returns (SD59x18) {
        int256 one = 1;
        SD59x18 one_cvt = convert(one);
        return (one_cvt).div(one_cvt.add((-x).exp()));
    }

    constructor(uint256 input_dim) {
        biases = new int256[][](layer_num);
        biases[0] = new int256[](1);
        biases[1] = new int256[](1);
    }

    function set_Biases(uint256 layer, int256[] calldata b) external {
        require(
            b.length == biases[layer].length,
            "Size of input biases does not match neuron number"
        );
        biases[layer] = b;
    }

    function set_Weights(uint256 layer, int256[] calldata w) external {
        require(layer < 2, "Layer index out of bounds");
        int256[] memory temp_w = new int256[](w.length);
        for (uint256 i = 0; i < w.length; i++) {
            temp_w[i] = w[i];
        }
        if (layer == 0) {
            weights_layer1.push(temp_w);
        } else if (layer == 1) {
            weights_layer2.push(temp_w);
        }
    }

    function view_dataset_size() external view returns (uint256 size) {
        size = training_data.length;
    }

    function set_TrainingData(int256[] calldata d) external {
        int256[] memory temp_d = new int256[](d.length);
        for (uint256 i = 0; i < d.length; i++) {
            temp_d[i] = d[i];
        }
        training_data.push(temp_d);
    }

    function classify() public view returns (int) {
        int correct = 0;
        for (uint256 j = 0; j < 50; j++) {
            int256[] memory data = training_data[j];
            int256 label = data[0];

            SD59x18[] memory neuronResultsLayer1 = new SD59x18[](
                weights_layer1.length
            );
            for (uint256 n = 0; n < weights_layer1.length; n++) {
                neuronResultsLayer1[n] = SD59x18.wrap(biases[0][n]);
                for (uint256 i = 1; i < data.length; i++) {
                    neuronResultsLayer1[n] = neuronResultsLayer1[n].add(
                        SD59x18.wrap(data[i]).mul(
                            SD59x18.wrap(weights_layer1[n][i - 1])
                        )
                    );
                }
                neuronResultsLayer1[n] = relu(neuronResultsLayer1[n]);
            }

            SD59x18[] memory neuronResultsLayer2 = new SD59x18[](
                weights_layer2.length
            );
            for (uint256 n = 0; n < weights_layer2.length; n++) {
                neuronResultsLayer2[n] = SD59x18.wrap(biases[1][n]);
                for (uint256 i = 0; i < weights_layer1.length; i++) {
                    neuronResultsLayer2[n] = neuronResultsLayer2[n].add(
                        neuronResultsLayer1[i].mul(
                            SD59x18.wrap(weights_layer2[n][i])
                        )
                    );
                }
                neuronResultsLayer2[n] = sigmoid(neuronResultsLayer2[n]);
            }

            int256 classification;
            SD59x18 point_five = sd(0.5e18);
            if (neuronResultLayer2[0].gte(point_five)) {
                classification = int256(1e18);
            } else {
                classification = int256(0e18);
            }

            if (label == classification) {
                correct++;
            }
        }
        return correct;
    }
}

Deploying Models

src/weights_biases weights and biases of the inferences

To deploy model, follow the below step.

1. Open up Ganache and make a workspace
2. Run truffle migrate in terminal

Then then contract will show up under the Contract tab in Ganache GUI

Evaluating Accuracy and Gas Cost

Once the contracts are running, run the according javascript test files under the test folder to evaluate a model's accuracy and gas cost.