ignis-autograd

Reverse-mode automatic differentiation over GPU arrays, on the ignis foundation.

Adds Ignis::AI::Tensor (a differentiable tensor) and an autograd tape. Build computation graphs on the GPU and get exact gradients (verified against finite differences).

require "ignis-autograd"
x = Ignis::AI::Tensor.from_host([1.0, 2.0, 3.0], shape: [3], requires_grad: true)
(x * x).sum.backward!
x.grad.to_host   # => [2.0, 4.0, 6.0]

MIT.