Documentation
¶
Overview ¶
Package autodiff provides automatic differentiation capabilities.
This package implements reverse-mode automatic differentiation (backpropagation) using a gradient tape. It wraps any backend to add autodiff capabilities.
Example:
import (
"github.com/born-ml/born/autodiff"
"github.com/born-ml/born/backend/cpu"
"github.com/born-ml/born/tensor"
)
func main() {
// Wrap CPU backend with autodiff
base := cpu.New()
backend := autodiff.New(base)
// Use for training
x := tensor.Zeros[float32](tensor.Shape{2, 3}, backend)
y := x.Add(x) // Operations recorded on tape
// Compute gradients
grads := backend.Backward(y.Raw())
}
Index ¶
- func Backward[T tensor.DType, B BackwardCapable](t *tensor.Tensor[T, B], backend B) map[*tensor.RawTensor]*tensor.RawTensor
- func NoGrad[B tensor.Backend](backend *Backend[B], fn func())
- func ReleaseGradients(grads map[*tensor.RawTensor]*tensor.RawTensor, backend ...tensor.Backend)
- type Backend
- type BackwardCapable
- type GradientTape
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func Backward ¶
func Backward[T tensor.DType, B BackwardCapable](t *tensor.Tensor[T, B], backend B) map[*tensor.RawTensor]*tensor.RawTensor
Backward computes gradients via backpropagation.
func NoGrad ¶ added in v0.3.0
NoGrad disables gradient recording for operations within the function. Use this for inference or operations that shouldn't be differentiated.
Example:
backend := autodiff.New(cpu.New())
autodiff.NoGrad(backend, func() {
// Operations here won't be recorded for gradients
output := model.Forward(input)
})
func ReleaseGradients ¶ added in v0.9.1
ReleaseGradients releases GPU buffers for all gradient tensors in the map. Call after optimizer.Step(grads) to free GPU memory immediately. Pass the backend so the releaser can schedule GPU buffer deferred destruction. Without a backend the function is a no-op (backwards-compatible default).
Types ¶
type Backend ¶
type Backend[B tensor.Backend] = autodiff.AutodiffBackend[B]
Backend is the autodiff-enabled backend.
type BackwardCapable ¶
type BackwardCapable = autodiff.BackwardCapable
BackwardCapable interface for backends that support backpropagation.
type GradientTape ¶
type GradientTape = autodiff.GradientTape
GradientTape records operations for automatic differentiation.
func NewGradientTape ¶
func NewGradientTape() *GradientTape
NewGradientTape creates a new gradient tape.
Source Files