Class: Informers::DonutFeatureExtractor

Inherits:
ImageFeatureExtractor show all
Defined in:
lib/informers/processors.rb

Instance Attribute Summary

Attributes inherited from FeatureExtractor

#config

Instance Method Summary collapse

Methods inherited from ImageFeatureExtractor

#call, #get_resize_output_image_size, #initialize, #preprocess, #rescale, #resize, #thumbnail

Methods inherited from FeatureExtractor

#initialize

Constructor Details

This class inherits a constructor from Informers::ImageFeatureExtractor

Instance Method Details

#pad_image(pixel_data, img_dims, pad_size, **options) ⇒ Object 



394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
 
# File 'lib/informers/processors.rb', line 394

def pad_image(pixel_data, img_dims, pad_size, **options)
  _image_height, _image_width, image_channels = img_dims

  image_mean = @image_mean
  if !image_mean.is_a?(Array)
    image_mean = new Array(image_channels, image_mean)
  end

  image_std = @image_std
  if !image_std.is_a?(Array)
    image_std = new Array(image_channels, image_std)
  end

  constant_values = image_mean.map.with_index { |x, i| -x / image_std[i] }

  super(
    pixel_data,
    img_dims,
    pad_size,
    center: true,
    # Since normalization is done after padding, we need to use certain constant values to ensure the same behaviour is observed.
    # For more information, see https://github.com/huggingface/transformers/blob/main/src/transformers/models/donut/image_processing_donut.py#L433-L451
    constant_values: constant_values,
    **options
  )
end