name: feature_extractor
backbone:
  encoder: vgg16
  pretrained: true
  output_dim: ${...latent_dim}
  output_scales: [0]
  num_downsample: 4
  decoder: [512, 256, 256, 128]