example.yml.example

training:
    # ---------------------------------------------  Data  ------------------------------------------------
    # Shape of the global patch
    global_crop_size: [16, 112, 112]
    # Whether to take into account the brain-bkd ratio also when cropping the gobal patch
    wise_crop: False

    # Number of small local views to generate. Set this parameter to 0 to disable multi-crop
    local_crops_number: 0
    # Shape of the local patch (not used)
    local_crop_size: null

    # Number of iterations to use during training (number of batches to sample)
    n_iters: 9600
    # Number of iterations to use during validation (number of batches to sample)
    n_val_iters: 3200

    # Whether the images provided are already cropped to the brain bbox or not
    already_cropped: True
    # If the images are not cropped to the brain bboxes, then either pass this path to a json
    # containing the bbox for each image. This file must be generated by meas of preprocessing pipeline.
    bboxes_paxth: 'example_files/bbox_info.json'
    # How thick the slab sampled in z axes will be
    slab_thickness: 24
    # This was added to use the ais label used for the visualizations during training
    all_datasets_csv_path: 'ssl/example_files/dataset.csv'
    # When using difference image as extra channel, this should be True
    multichannel_input: False
    
    # Transformations. Default:
    transformations_cfg:
      # Used for large global view cropping. When disabling multi-crop (local_crops_number: 0)
      global_scale: [0.8, 1.2]
      # Used for small local view cropping of multi-crop. Not used if local_crops_nr = 0
      local_scale: null
      # whether to use mirror data augmentation with a probabbility of 0.8 across the three axes
      mirror: True
      # Probability to use with symmetry data aumentation, if provided the dataset used should be
      # selected accordingly
      symmetry: 0.7
      # Batch of standard augmentations
      i_global:
        g_noise: [1.0, 1.0]
        g_blur: [0.8, 0.8]
        mult_bright: [0.8, 1.0]
        brightness: [0.8, 0.8]
        contrast_augm: [1.0, 1.0]
        gamma: [0.8, 1.0]
        per_channel: False
      i_local:
        g_noise: null
        g_blur: null
        mult_bright: null
        brightness: null
        contrast_augm: null
        gamma: null
        per_channel: null

    # ------------------------------------- Model Parameters --------------------------------------
    # Dimensionality of the projection head output.
    out_dim: 65536
    # Whether or not to weight normalize the last layer of the DeSD head.
    norm_last_layer: True
    # Base EMA parameter for teacher update. The value is increased to 1 during training
    # with cosine schedule. 
    momentum_teacher: 0.996
    # Whether to use batch normalizations in projection head
    use_bn_in_head: True
    # Initial value for the teacher temperature.
    warmup_teacher_temp: 0.04
    # Final value (after linear warmup) of the teacher temperature.
    teacher_temp: 0.07
    # Number of warmup epochs for the teacher temperature.
    warmup_teacher_temp_epochs: 10

    #------------------------------ Training/Optimization parameters -----------------------------------
    # Whether or not to use half precision for training.
    use_fp16: True
    # Initial value of the weight decay
    weight_decay: 0.04
    # Final value of the weight decay. We use a cosine schedule for WD
    weight_decay_end: 0.000001
    # Maximal parameter gradient norm if using gradient clipping. 0 for disabling.
    clip_grad: 0
    # Batch-size. Default: 32
    batch_size: 64
    # Epochs. Default: 100
    epochs: 100
    # Number of epochs during which we keep the output layer fixed. 
    freeze_last_layer: 1
    # Learning rate at the end of linear warmup (highest LR used during training).
    lr: 0.1
    # Number of epochs for the linear learning-rate warm up.
    warmup_epochs: 10
    # This parameter is used to configure the learning rate initial value ooriginally set by
    # a fized rule. This should be eaqual to the batch size
    lr_sch_den: 64
    # Target LR at the end of optimization. We use a cosine LR schedule with linear warmup.
    min_lr: 0.01
    # Options: 'adamw', 'sgd', 'lars'. Type of optimizer.
    optimizer: 'adamw'
    # How the loss obtained at the different stages are combined. Costant means the wight is
    # the same for all. Other options are: 'exp', 'fraction', 'linear', 'last'. Decreasing
    # from bottleneck to base.
    weights: 'constant'

    # ------------------------------ Misc -------------------------------------------------------
    # Path to save logs and checkpoints.
    output_dir: '<SOME_PATH>'
    # Save checkpoint every x epochs.
    saveckp_freq: 50
    # Random seed.
    seed: 0
    # Number of data loading workers
    num_workers: 6

    # -------------------------------  Metrics_cfg -----------------------------------------------
    metrics_cfg:
      # Over which set to compute the metrics at each epoch's end. 'train' and/or 'val'
      over: ['val']
      # Frequency to log the RankMe meassurment, null if avoided
      rank_me_freq: 1
      # Frequency to log the dimensionality reduced projection scatter plot, null if avoided
      projection_freq: 5

    # -------------------------------  Pretrained SSL encoder -------------------------------------
    # learning rate of the encoder weights (not projection head)
    encoder_lr: 0.001
    # minimum learning rate of the encoder weights (not projection head)
    min_encoder_lr: 0.0001
    # path to the pretrained weights
    encoder_chkpt: '<SOME_PATH>/checkpoint.pth'

    # ---------------------------------- nnUnet ---------------------------------------------------
    # name of the nnUNet dataset to use
    dataset: 'Dataset044_AIS'
    # name of the configuration that was used to build the model
    configuration: '3d_fullres'
    # partition to use during training
    fold: 0
    # nnUNet trainer name
    trainer: 'nnUNetTrainerCfg'
    # nnUNet experiment planner name
    exp_planner: 'nnUNetPlansSSL'
    # whether to use gpu
    device: 'cuda'
    # This configurations are used to generate the config file that is later going to be used to
    # finetune nnUNet
    # Number of epochs to finetune nnUNet afterwards
    num_epochs: 100
    # Unfreezing policy to use in supervised nnUNet finetuning
    unfreeze_lr: null
    unfreeze_epoch: null
    # This flag indicates nnUNet pipeline that the model uses pretrained weights
    ssl_pretrained: True
    # Path from where to extract the partitions
    split_path: '<SOME_PATH>/splits_final.json'