OpenPath/eval_configs/monusac.yaml

---
trainer:
  class_path: eva.Trainer
  init_args:
    n_runs: &N_RUNS ${oc.env:N_RUNS, 5}
    default_root_dir: &OUTPUT_ROOT ${oc.env:OUTPUT_ROOT, logs/${oc.env:MODEL_NAME, vit_small_patch16_224_dino}/monusac}
    max_steps: &MAX_STEPS ${oc.env:MAX_STEPS, 2000}
    log_every_n_steps: 6
    checkpoint_type: ${oc.env:CHECKPOINT_TYPE, best}
    devices: ${oc.env:NUM_DEVICES, 1}
    callbacks:
      - class_path: eva.callbacks.ConfigurationLogger
      - class_path: lightning.pytorch.callbacks.TQDMProgressBar
        init_args:
          refresh_rate: ${oc.env:TQDM_REFRESH_RATE, 1}
      - class_path: eva.vision.callbacks.SemanticSegmentationLogger
        init_args:
          log_every_n_epochs: 1
          mean: &NORMALIZE_MEAN ${oc.env:NORMALIZE_MEAN, [0.485, 0.456, 0.406]} 
          std: &NORMALIZE_STD ${oc.env:NORMALIZE_STD, [0.229, 0.224, 0.225]}
      - class_path: lightning.pytorch.callbacks.ModelCheckpoint
        init_args:
          filename: best
          save_last: ${oc.env:SAVE_LAST, false}
          save_top_k: 1
          monitor: &MONITOR_METRIC ${oc.env:MONITOR_METRIC, 'val/MonaiDiceScore'}
          mode: &MONITOR_METRIC_MODE ${oc.env:MONITOR_METRIC_MODE, max}
      - class_path: lightning.pytorch.callbacks.EarlyStopping
        init_args:
          min_delta: 0
          patience: ${oc.env:PATIENCE, 100}
          monitor: *MONITOR_METRIC
          mode: *MONITOR_METRIC_MODE
    logger:
      - class_path: lightning.pytorch.loggers.TensorBoardLogger
        init_args:
          save_dir: *OUTPUT_ROOT
          name: ""
model:
  class_path: eva.vision.models.modules.SemanticSegmentationModule
  init_args:
    encoder:
      class_path: eva.core.models.wrappers.TorchHubModel
      init_args:
        repo_or_dir: facebookresearch/dinov2 
        model_name: dinov2_vitg14_reg 
        pretrained: true
        out_indices: 1
        checkpoint_path: ${oc.env:CHECKPOINT_PATH, ../checkpoints/teacher_epoch250000.pth}

    decoder:
      class_path: eva.vision.models.networks.decoders.segmentation.ConvDecoderWithImage
      init_args:
        in_features: ${oc.env:IN_FEATURES, 1536}
        num_classes: &NUM_CLASSES 5
    criterion:
      class_path: eva.vision.losses.DiceLoss
      init_args:
        softmax: true
        batch: true
        ignore_index: &IGNORE_INDEX 5
    lr_multiplier_encoder: 0.0
    optimizer:
      class_path: torch.optim.AdamW
      init_args:
        lr: ${oc.env:LR_VALUE, 0.002}
    postprocess:
      predictions_transforms:
        - class_path: torch.argmax
          init_args:
            dim: 1
    metrics:
      common:
        - class_path: eva.metrics.AverageLoss
      evaluation:
        - class_path: eva.vision.metrics.defaults.MulticlassSegmentationMetrics
          init_args:
            num_classes: 6
            ignore_index: *IGNORE_INDEX
        - class_path: torchmetrics.ClasswiseWrapper
          init_args:
            metric:
              class_path: eva.vision.metrics.MonaiDiceScore
              init_args:
                include_background: true
                num_classes: 6
                reduction: none
                ignore_index: *IGNORE_INDEX
            labels:
              - background
              - epithelial
              - lymphocyte
              - neutrophil
              - macrophage
              - ambiguous
data:
  class_path: eva.DataModule
  init_args:
    datasets:
      train:
        class_path: eva.vision.datasets.MoNuSAC
        init_args: &DATASET_ARGS
          root: ${oc.env:DATA_ROOT, /block/eva-data/monusac}
          split: train
          download: ${oc.env:DOWNLOAD_DATA, false}
          # Set `download: true` to download the dataset from https://monusac-2020.grand-challenge.org/Data/
          # The MoNuSAC dataset is distributed under the following license: 
          # "Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International"
          # (see: https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode)
          transforms:
            class_path: torchvision.transforms.v2.Compose
            init_args:
              transforms:
                - class_path: torchvision.transforms.v2.RandomResizedCrop
                  init_args:
                    size: ${oc.env:RESIZE_DIM, 224}                    
                - class_path: torchvision.transforms.v2.ToDtype
                  init_args:
                    dtype: torch.float32
                    scale: true
                - class_path: torchvision.transforms.v2.Normalize
                  init_args:
                    mean: *NORMALIZE_MEAN
                    std: *NORMALIZE_STD
      val:
        class_path: eva.vision.datasets.MoNuSAC
        init_args:
          <<: *DATASET_ARGS
          split: test
          transforms:
            class_path: eva.vision.data.transforms.common.ResizeAndCrop
            init_args:
              size: ${oc.env:RESIZE_DIM, 224}
              mean: *NORMALIZE_MEAN
              std: *NORMALIZE_STD
    dataloaders:
      train:
        batch_size: &BATCH_SIZE ${oc.env:BATCH_SIZE, 64}
        num_workers: &N_DATA_WORKERS ${oc.env:N_DATA_WORKERS, 4}
        shuffle: true
      val:
        batch_size: *BATCH_SIZE
        num_workers: *N_DATA_WORKERS