instance_cityscapes.yaml

# pytorch_lightning==1.8.6
seed_everything: 0

trainer:
  accelerator: gpu
  strategy: ddp_find_unused_parameters_false
  logger: False

model:
  semantic_model:
    class_path: semantic_fine_tuning.SemanticFineTuner
    init_args:
      dinov2_vit_model: "vitb14"
      num_classes: ${data.init_args.num_classes}
      train_output_size: ${data_params.image_size_input}
      upsample_factor: 14.0
      head: "mlp"
      ignore_index: 255
      top_k_percent_pixels: 0.2
      test_output_size: ${data_params.image_size_original}
      test_multi_scales: [ 1, 2, 3 ]
      test_plot: False
      test_save_dir: null
  semantic_model_ckpt: "checkpoints/semantic_cityscapes.ckpt"
  boundary_model:
    class_path: boundary_fine_tuning.BoundaryFineTuner
    init_args:
      dinov2_vit_model: "vitb14"
      mode: "direct"
      upsample_factor: 4.0
      head: "mlp"
      neighbor_radius: 1.5
      threshold_boundary: 0.93
      num_boundary_neighbors: 1
      test_output_size: ${data_params.image_size_original}
      test_multi_scales: [3, 4, 5]
      test_plot: False
  boundary_model_ckpt: "checkpoints/boundary_cityscapes.ckpt"
  structure_connectivity: [[0, 1, 0], [1, 1, 1], [0, 1, 0]]
  instance_min_pixel: 500
  erosion_structure: [[1, 1, 1], [1, 1, 1], [1, 1, 1]]
  erosion_iterations: 0
  output_size: ${data_params.image_size_original}
  test_plot: False
  test_save_dir: "results/cityscapes"
  test_save_vis: True

data:
  class_path: datasets.cityscapes.CityscapesDataModule
  init_args:
    cfg_dataset:
      name: "cityscapes"
      path: "" # SET THE PATH TO THE CITYSCAPES DATASET
      feed_img_size: ${data_params.image_size_original}
      offsets: [0]
      remove_classes: []
    num_classes: 19
    batch_size: 1
    num_workers: 2
    transform_train: []
    transform_test:
      - class_path: utils.transforms.ToTensor
      - class_path: utils.transforms.Resize
        init_args:
          size: ${data_params.image_size_input}
      - class_path: utils.transforms.MaskPostProcess
      - class_path: utils.transforms.ImageNormalize
        init_args:
          mean: ${data_params.image_mean}
          std: ${data_params.image_std}
    label_mode: "cityscapes_19"
    train_sample_indices: []
    test_sample_indices: null
    test_set: "val"

data_params:
  image_size_original: [1024, 2048]
  image_size_input: [1008, 2016]
  image_mean: [0.485, 0.456, 0.406]
  image_std: [0.229, 0.224, 0.225]