class

GS::Autograd::Variable

Inherits Reference < Object

Variable wraps a Tensor and tracks gradients

Constructors

from_tensor(tensor : Tensor, requires_grad : Bool = false) : Variable

Create from existing data

new(data : Tensor, requires_grad : Bool = false)

Create variable from tensor

ones(*dims : Int32, requires_grad : Bool = false, device : Tensor::Device = Tensor::Device::GPU) : Variable

rand(*dims : Int32, requires_grad : Bool = false, device : Tensor::Device = Tensor::Device::GPU) : Variable

randn(*dims : Int32, requires_grad : Bool = false, device : Tensor::Device = Tensor::Device::GPU) : Variable

zeros(*dims : Int32, requires_grad : Bool = false, device : Tensor::Device = Tensor::Device::GPU) : Variable

Factory methods that mirror Tensor

Instance methods

*(other : Variable) : Variable

Multiplication

*(scalar : Float32) : Variable

+(other : Variable) : Variable

Addition

+(scalar : Float32) : Variable

-(other : Variable) : Variable

Subtraction

/(other : Variable) : Variable

Division

backward(grad_output : Tensor | Nil = nil) : Nil

Backward pass - compute gradients

clone

Clone with gradient tracking

data

detach

Detach from computation graph (returns new variable with same data, no grad tracking)

grad

grad=(grad : Tensor | Nil)

grad_fn

grad_fn=(grad_fn : GradFn | Nil)

inspect(io : IO) : Nil

Appends a String representation of this object which includes its class name, its object address and the values of all instance variables.

class Person
  def initialize(@name : String, @age : Int32)
  end
end

Person.new("John", 32).inspect # => #<Person:0x10fd31f20 @name="John", @age=32>

is_leaf?

item

Item (for scalar tensors)

matmul(other : Variable) : Variable

Matrix multiplication

mean

Mean reduction

ndim

numel

relu

ReLU activation

requires_grad?

reshape(*dims : Int32) : Variable

Reshape

shape

Shape delegation

sigmoid

Sigmoid activation

sum

Sum reduction

t

Transpose

to_s(io : IO) : Nil

String representation

transpose

Transpose (swap last two dims)

zero_grad!

Zero gradient