Class: Gsplat::Training::ImageFitter

Inherits:
Object
  • Object
show all
Defined in:
lib/gsplat/training/image_fitter.rb

Overview

A small self-consistency image fitting loop used by examples and E2E tests.

Defined Under Namespace

Classes: MeanSquaredError, Result

Instance Attribute Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(width: 64, height: 64, n_gaussians: 2_000, learning_rate: 20.0, seed: 42, dtype: Numo::SFloat) ⇒ ImageFitter

rubocop:disable Metrics/ParameterLists



33
34
35
36
37
38
39
40
41
42
43
44
 
# File 'lib/gsplat/training/image_fitter.rb', line 33

def initialize(width: 64, height: 64, n_gaussians: 2_000, learning_rate: 20.0,
               seed: 42, dtype: Numo::SFloat)
  # rubocop:enable Metrics/ParameterLists
  @width = width
  @height = height
  @n_gaussians = n_gaussians
  @learning_rate = learning_rate
  @dtype = dtype
  @rng = Random.new(seed)
  validate_options!
  build_scene
end

Instance Attribute Details

#dtype ⇒ Object (readonly)

Returns the value of attribute dtype.



30
31
32
 
# File 'lib/gsplat/training/image_fitter.rb', line 30

def dtype
  @dtype
end

#height ⇒ Object (readonly)

Returns the value of attribute height.



30
31
32
 
# File 'lib/gsplat/training/image_fitter.rb', line 30

def height
  @height
end

#learning_rate ⇒ Object (readonly)

Returns the value of attribute learning_rate.



30
31
32
 
# File 'lib/gsplat/training/image_fitter.rb', line 30

def learning_rate
  @learning_rate
end

#n_gaussians ⇒ Object (readonly)

Returns the value of attribute n_gaussians.



30
31
32
 
# File 'lib/gsplat/training/image_fitter.rb', line 30

def n_gaussians
  @n_gaussians
end

#width ⇒ Object (readonly)

Returns the value of attribute width.



30
31
32
 
# File 'lib/gsplat/training/image_fitter.rb', line 30

def width
  @width
end

Instance Method Details

#fit(steps: 300) ⇒ Object

Raises:

  • (ArgumentError) 


46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
 
# File 'lib/gsplat/training/image_fitter.rb', line 46

def fit(steps: 300)
  raise ArgumentError, "steps must be a positive integer" unless steps.is_a?(Integer) && steps.positive?

  target = render(@teacher_colors)
  colors = initial_colors
  history = [psnr(render(colors), target)]
  steps.times do
    variable = Autograd::Variable.new(colors, requires_grad: true)
    rendered = render(variable)
    MeanSquaredError.apply(rendered, target).backward
    colors = clipped(colors - (learning_rate * variable.grad), 0.0, 1.0)
    history << psnr(render(colors), target)
  end
  Result.new(
    initial_psnr: history.first,
    final_psnr: history.last,
    history: history.freeze,
    colors: colors,
    target: target
  )
end