A deep learning training framework built with MoonBit, featuring tape-based autograd and PyTorch-like API (CPU backend)

deep-learning

resnet

neural-network

autograd

tensor

Download zip

Author

chnlkw

Version

0.1.1

License

MIT

Repository

github.com/moonxi-net/moonxi-net

Last updated

4 months ago

Downloads

Dependencies

moonbitlang/x0.4.43

#moonxi-net

A deep learning training framework built with MoonBit, featuring tape-based autograd and a PyTorch-like API on the CPU backend (NpArray). Write your model once using the Tensor trait, and it works with both CPU and GPU backends via tagless-final style.

#Installation

moon add chnlkw/moonxi-net

Note: This package works with the standard MoonBit toolchain — no CUDA or GPU required.

#Quick Example

A minimal linear regression: learn y = 3x + 1 from data.

///|
fn[T : @tensor.Tensor + @tensor.BlasTensor] train_linear() -> Array[T] {
  let w = @grad.no_grad(T::zeros([1, 1]))
  let b = @grad.no_grad(T::zeros([1, 1]))
  let x = @grad.no_grad(
    T::from_host(FixedArray::makei(5, i => Float::from_int(i + 1)), [5, 1]),
  )
  let y = @grad.no_grad(
    T::from_host(FixedArray::makei(5, i => Float::from_int((i + 1) * 3 + 1)), [
      5, 1,
    ]),
  )
  let params : Array[@grad.Grad[T]] = [w, b]
  for _epoch in 0500 {
    @grad.clear_tape()
    for p in params { p.grad = Some(None) }
    let loss = x.matmul(w).add(b).sub(y).square().mean()
    loss.backward()
    for p in params {
      match p.grad {
        Some(Some(g)) => p.value = p.value.sub(g.scale(0.001))
        _ => ()
      }
    }
  }
  params.map(g => g.value)
}

moon run moonxi-net/examples/linear --target native
# Learned: w=3.00, b=1.00

#Usage

#Import in moon.pkg

{
  "import": [
    "chnlkw/moonxi-net" @tensor,
    "chnlkw/moonxi-net/nparray" @nparray,
    "chnlkw/moonxi-net/grad" @grad,
    "chnlkw/moonxi-net/model" @model,
    "chnlkw/moonxi-net/optimizer" @optimizer,
    "chnlkw/moonxi-net/loss" @loss,
    "chnlkw/moonxi-net/train" @train,
    "chnlkw/moonxi-net/dataloader" @dl
  ]
}

#Core Packages

Package	Alias	Description
moonxi-net	@tensor	Core traits: Tensor, BlasTensor, ImageTensor, ImageBackwardOps
moonxi-net/nparray	@nparray	CPU tensor backend (NpArray implements all tensor traits)
moonxi-net/grad	@grad	Tape-based autograd engine (Grad[T], backward(), clear_tape())
moonxi-net/model	@model	Neural network layers: Linear[T], Conv2d[T], ResNet18[T], MLP[T]
moonxi-net/optimizer	@optimizer	Optimizers: Momentum SGD, Adam, RMSprop (with gradient clipping)
moonxi-net/loss	@loss	Loss functions: cross-entropy, MSE
moonxi-net/train	@train	Training loop utilities, Experiment config, run_cpu
moonxi-net/dataloader	@dl	DataLoader with Fisher-Yates shuffle and mini-batch iteration
moonxi-net/datasets/mnist	—	MNIST dataset loader
moonxi-net/datasets/cifar10	—	CIFAR-10 dataset loader

#Basic Training Workflow

///|
fn main {
  // Create tensors using NpArray (CPU backend)
  let w = @grad.no_grad(@nparray.NpArray::zeros([10, 10]))
  let b = @grad.no_grad(@nparray.NpArray::zeros([10]))

  // Build computation graph, compute loss, backprop
  @grad.clear_tape()
  let loss = /* ... your forward pass ... */
  loss.backward()

  // Access gradients
  match w.grad {
    Some(Some(g)) => w.value = w.value.sub(g.scale(0.01))
    _ => ()
  }
}

#Build & Test

# From this directory
moon test --target native

# Build
moon build --target native

moonxi-net

#moonxi-net

#Installation

#Quick Example

#Usage

#Import in moon.pkg

#Core Packages

#Basic Training Workflow

#Build & Test

#License

#BlasTensor

#ImageBackwardOps

#ImageTensor

#Tensor

#CudaError

#IoError

#ShapeError

#UnsupportedError

#ShapeTensor

#ShapeTensor::dim

#ShapeTensor::from_array

#ShapeTensor::linear

#ShapeTensor::ndim

#ShapeTensor::pool2d

#ShapeTensor::to_array

#TensorData

Source Files

BlasTensor

ImageBackwardOps

ImageTensor

Tensor

CudaError

IoError

ShapeError

UnsupportedError

ShapeTensor

ShapeTensor::dim

ShapeTensor::from_array

ShapeTensor::linear

ShapeTensor::ndim

ShapeTensor::pool2d

ShapeTensor::to_array

TensorData