Neural Network for MNIST in Cairo 1.0

Implementation of a Neural Network from scratch using Cairo 1.0 for MNIST predictions.

The NN has a simple two-layer architecture:

• Input layer 𝑎[0] will have 784 units corresponding to the 784 pixels in each 28x28 input image.
• A hidden layer 𝑎[1] will have 10 units with ReLU activation.
• Output layer 𝑎[2] will have 10 units corresponding to the ten digit classes with softmax activation.

Functionalities implemented in Cairo 1.0:

• Vector implementation with operations: sum, max, min, argmax.
• Matrix implementation with operations: get, dot, add, len.
• Tensor implementation.
• 8-bit weight quantization based in ONNX quantization.
• ReLU activation.
• Forward propagation of NN.
• Predict method for NN.
• Pseudo-softmax activation optimized for quantized values.
• Weight loading into Cairo NN from trained Tensorflow NN.
• MNIST inferences using Cairo NN.

Built with auditless/cairo-template

Working with the project

Currently supports building and testing contracts.

Build

Build the contracts.

$ make build

Test

Run the tests in src/test:

$ make test

Format

Format the Cairo source code (using Scarb):

$ make fmt

Credits

• cairo-ml by Raphael Doukhan.