Class: Gsplat::Training::Losses::StructuralSimilarity

Inherits:

Autograd::Function

• Object
• Autograd::Function
• Gsplat::Training::Losses::StructuralSimilarity

Defined in:

lib/gsplat/training/losses.rb

Overview

Structural similarity with the standard 11x11, sigma 1.5 window.

Class Method Summary

• .backward(context, grad_output) ⇒ Object private

  Propagates the SSIM score gradient.

• .forward(context, image_a, image_b, layout:) ⇒ Object private

  Evaluates mean SSIM and stores convolution intermediates.

Methods inherited from Autograd::Function

apply

Class Method Details

.backward(context, grad_output) ⇒ Object

This method is part of a private API. You should avoid using this method if possible, as it may be removed or be changed in the future.

Propagates the SSIM score gradient.

# File 'lib/gsplat/training/losses.rb', line 48

def backward(context, grad_output)
  Math::Ssim.backward(context.saved_values.first, grad_output)
end

.forward(context, image_a, image_b, layout:) ⇒ Object

This method is part of a private API. You should avoid using this method if possible, as it may be removed or be changed in the future.

Evaluates mean SSIM and stores convolution intermediates.

# File 'lib/gsplat/training/losses.rb', line 40

def forward(context, image_a, image_b, layout:)
  score, cache = Math::Ssim.forward(image_a, image_b, layout: layout)
  context.save(cache)
  image_a.class.cast(score)
end