minigrad

A minimal implementation of autograd and neural networks. Inspired by micrograd and pytorch. This library is meant to be educational and not meant for production code.

Install and dev work:

Install: go get github.com/NoahSchiro/minigrad@latest

Import in your code:

// You may only need one of these depending on your work
import "github.com/NoahSchiro/minigrad/tensor"
import "github.com/NoahSchiro/minigrad/ndarray"
import "github.com/NoahSchiro/minigrad/nn"

Dev work:

Install: git clone git@github.com:NoahSchiro/minigrad.git

Run tests: go test ./...

Benchmark code: go test ./... -bench=.

Example:

There is an example for the xor problem which can be found in cmd/xor/

There is an example for finding the boundary of a circle in cmd/circle/

Features and roadmap:

• Autograd engine
• Linear layers
• RNN, LSTM (targeting v2.1)
• CNN (targeting v2.2)
• Sigmoid
• ReLU
• SoftMax (targeting v2.1)
• SGD optimizer
• Adam optimizer (targeting v2.1)
• CPU parallelism (targeting v2.1)
• CUDA support (targeting v2.1)

SCC Report

The goal is to keep the library under 1k lines of code. This excludes test files and examples.

Didn't really take us long to blow past this goal (57 days since v2.0 was released).

When we add CUDA and C, the goal will be to keep each of those under 500 lines of code combined.

Language	Files	Lines	Blanks	Comments	Code
Go	12	1579	262	137	1180