semantic_cityscapes.yaml

# pytorch_lightning==1.8.6
seed_everything: 0

trainer:
  max_epochs: 50
  accelerator: gpu
  strategy: ddp_find_unused_parameters_false
  sync_batchnorm: True
  logger: False
  callbacks:
    - class_path: pytorch_lightning.callbacks.model_checkpoint.ModelCheckpoint
      init_args:
        dirpath: "checkpoints/"
        filename: "semantic_cityscapes"
        every_n_epochs: 10

model:
  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

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: 1
    transform_train:
      - class_path: utils.transforms.ToTensor
      - class_path: utils.transforms.RandomHorizontalFlip
      - class_path: utils.transforms.RandomResizedCrop
        init_args:
          size: ${data_params.image_size_input}
          scale: [0.2, 1.0]
      - class_path: utils.transforms.ColorJitter
        init_args:
          brightness: 0.2
          contrast: 0.2
          saturation: 0.2
          hue: 0.2
      - class_path: utils.transforms.MaskPostProcess
      - class_path: utils.transforms.ImageNormalize
        init_args:
          mean: ${data_params.image_mean}
          std: ${data_params.image_std}
    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: [12, 324, 450, 608, 742, 768, 798, 836, 1300, 2892]
    test_sample_indices: null

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]

ckpt_path: null

optimizer:
  class_path: torch.optim.Adam
  init_args:
    lr: 0.001

lr_scheduler:
  class_path: torch.optim.lr_scheduler.CosineAnnealingLR
  init_args:
    T_max: ${trainer.max_epochs}