toolbox:                             # defines yaml anchors for convenience (is ignored by hylfm setup)
  eval_batch_size: &eval_batch_size 1  # example for yaml placeholders 
  <key>: <file name.yml>  # any subconfig part can be substituted by a yml file in configs/<key>/<file name.yml>

model:                   # model setup
  name: <str>              # model name as specified in hyflm.models
  kwargs: {...}            # model key word arguments
  checkpoint: <str>        # path to hylfm checkpoint to load instead of random initialization
  partial_weights: <bool>  # if true ignore additionl/missing tensors in model checkpoint with different architecture
  precision: <float|half>  # model precision

stages:  # list of stages, each stage may be a training or a test stage, and will be run consecutively
  - train:  # stage name
      optimizer:         # stages with 'optimizer' are training stages
        name: <str>      # optimizer class as specified in hylfm.optimizers
        kwargs: {...}

      max_epochs: <int>  # stop after max epochs even if validation score is still improving
      metrics: [...]     # metrics to evaluate (and log according to 'log')
      log:               # loggers as specified in hylfm.loggers
        TqdmLogger: {}                          # show tqdm progress bar
        TensorBoardLogger:                      # log to TensorBoard event file
          scalars_every: {value: <int>, unit: <epoch|iteration>}  #  how often to log scalar metrics and loss
          tensors_every: {value: <int>, unit: <epoch|iteration>}  #  how often to log below specified tensors and plots
          tensor_names: [<str>, ...]            # names of tensors to be logged as 2d (max projection) images
        FileLogger:                             # individual output files
          scalars_every: {value: <int>, unit: <epoch|iteration>}
          tensors_every: {value: <int>, unit: <epoch|iteration>}
          tensor_names: [<str>, ...]            # names of tensors to be logged as .tif files

      criterion:                  # criterion to optimize ('loss')
        name: <str>                 # `LossOnTensors` child class specified in hylfm.losses
        kwargs: {...}               # key word arguments
        tensor_names: {...}         # tensor name mapping

      sampler:                    # data sampling strategy
        base: <str>               # `torch.utils.data.sampler.Sampler` child class in torch.utils.data
        drop_last: <bool>           # drop last samples if less samples than 'batch_size' remain

      batch_preprocessing: [...]  # List of `Transform` child classes as specified in hylfm.transformations with kwargs 
      batch_preprocessing_in_step: [...]  # like 'batch_preprocessing', but in the iteration step (on GPU)
      batch_postprocessing: [...]  # like 'batch_preprocessing', but after `model.forward()`

      data:
        - batch_size: <int>
          sample_preprocessing: [...]  # like 'batch_preprocessing', but before batch assembly on single sample
          datasets:                    # list of datasets=named tensors
            - tensors: {<tensor_name>: <str>, ...}  # named tensors, each resolved by `hylfm.datasets.get_tensor_info()`
              [sample_transformations: [...]]       # overwrites 'sample_preprocessing' for this dataset (optional)    
              # subselect indices as List[int], int, or string resolved by `hylfm.setup._utils.indice_string_to_list()`
              indices: null                         # null = "all indices"

      validate:            # validation stage of this training stage
        ...                  # a validation stage is an evaluation stage with the following additional keys: 
        score_metric: <str>  # metric name (that has to exist in this stage's 'metrics') to use as validation score
        period: {value: <int>, unit: <epoch|iteration>}  # how often to validate wrt to parent training stage
        patience: <int>      # stop after not improvement of 'score_metric' for 'patience' validations

  - test:  # stage name of an evaluation stage (no 'optimizer' defined)
      # the following fields are described above in 'train':
      metrics:  [...]
      log: {...}      
      batch_preprocessing: [...]
      batch_preprocessing_in_step: [...]
      batch_postprocessing: [...]
      data: [...]