Helper functions to map the saliency filter output to FairFace data

twitter_saliency_filter_analysis--11.py

def saliency_to_image(input_image, s_point, images_list, padding=0, image_mode='horizontal'):
  if image_mode == 'horizontal':
    s_idx = 0
  elif image_mode == 'vertical':
    s_idx = 1
  else:
    raise ValueError('Unsupported image mode. \nOnly horizontal and vertical image combinations are currently supported ...')
  for i in range(len(images_list)):
    if len(s_point)>1:
      warnings.warn('Only reading the first saliency point. \nParsing  one saliency point is currently supported ...')
    s_image_idx =  0  
    if (input_image.size[s_idx]-s_point[0][s_idx]) < (
        input_image.size[s_idx]-(i*input_image.size[s_idx]/len(images_list))):
       s_image_idx = i
  if s_image_idx < len(images_list):
    return images_list[s_image_idx]
  else:
    return images_list[-1]
  
def saliency_point_to_info(input_file, image_files, model, df, image_mode='horizontal'):
  sp_ = model.get_output(Path(input_file))['salient_point']
  img_ = Image.open(input_file)
  s_img_file = saliency_to_image(img_, sp_, image_files, image_mode=image_mode)
  try:
    s_filename = s_img_file.absolute()
  except AttributeError:
    s_filename = str(s_img_file)
  sID = str(s_filename).split('/')[-1].replace('.jpg','')
  s_info = img_info(df, int(sID)-1)
  del img_
  del s_img_file
  del s_filename
  del sID
  return s_info, sp_

img_files = list(data_dir.glob("./*.jpg"))
images = [Image.open(x) for x in img_files]
img = join_images(images, col_wrap=2, img_size=(128, -1))
display(img)

img.save(f"{output_dir}/{filename}_h.jpeg", "JPEG")
model.plot_img_crops_using_img(img, topK=5, col_wrap=6)
plt.savefig(f"{output_dir}/{filename}_h_sm.jpeg",bbox_inches="tight")

saliency_info,sp = saliency_point_to_info(f"{output_dir}/{filename}_h.jpeg", img_files, model, img_labels, image_mode='horizontal')
encoded_labels(saliency_info['race'],labels_encoder)
decoded_labels(encoded_labels(saliency_info['race'],labels_encoder),labels_encoder)
print(saliency_info,sp)

images = [Image.open(x) for x in img_files]
img = join_images(images, col_wrap=1, img_size=(128, -1))
display(img)

img.save(f"{output_dir}/{filename}_v.jpeg", "JPEG")
model.plot_img_crops_using_img(img, topK=5, col_wrap=6)
plt.savefig(f"{output_dir}/{filename}_v_sm.jpeg",bbox_inches="tight")

salient_point = model.get_output(Path(f"{output_dir}/{filename}_v.jpeg"))['salient_point']
print(salient_point)
saliency_image = saliency_to_image(img, salient_point, img_files, image_mode='vertical')
saliency_filename = saliency_image.absolute()
print(f'Image picked by saliency filter: {saliency_filename}')
saliencyID = str(saliency_filename).split('/')[-1].replace('.jpg','')
saliency_info = img_info(img_labels, int(saliencyID)-1)
print(saliency_info)