config_effdet2_l.yaml

callbacks:
  model_checkpoint:
    _target_: pytorch_lightning.callbacks.model_checkpoint.ModelCheckpoint
    monitor: "val/Competition_metric/"  #"val/mean_auc"
    save_top_k: 0
    auto_insert_metric_name: False
    save_last: True
    mode: "max"
    dirpath: "checkpoints/"
    filename: "epoch-{epoch}_step-{step}_loss-{1/val/Competition_metric/:.2f}"

  SWA:
    _target_: pytorch_lightning.callbacks.stochastic_weight_avg.StochasticWeightAveraging
    swa_epoch_start: 15 
    annealing_epochs: 5
    annealing_strategy: 'cos'

datamodule:
  _target_: src.datamodules.Datamodules.Nodule21
  cfg:
    name: Nodule21
    model: ${model.cfg.name}
    path:
      train:
        images: 
        labels: 
  
      val:  ## part of train set
        images: 
        labels: 

    sample_set: False
    preload : False

    imbalancedSampling: True
    imbalancedSamplingTest: True
    new_shape: 1024   ##length of the square image
    replacement: True 
    num_gpus : 1
    batch_size: 4
    num_workers: 2
    lr : 1e-4
    test: False
    colorJitter: False
    brightness: 0.05 # colorjitter brightness
    contrast: 0
    saturation: 0
    hue: 0
    randomCrop: False
    cutout: False
    cropsize: 900
    invertIntensity: False
    vinDR_augment: True
    effdet: True
    all_sets_mixed: True # use v3 split!
    final_set: True
    
    cropStrat: downsample 

trainer: 
  _target_: pytorch_lightning.Trainer
  gpus: [0] # -1
  min_epochs: 1
  max_epochs: 20 #20
  # limit_train_batches: 0.5
  gradient_clip_val: 3.0
  # accelerator: ddp
  sync_batchnorm: True
  log_every_n_steps: 50
  weights_summary: null
  progress_bar_refresh_rate: 25
  # resume_from_checkpoint: null
  # profiler : "simple"
  # val_check_interval: 0
  # overfit_batches : 0.001
  precision : 16
  num_sanity_val_steps : 0 # This does not work with dp, only with ddp
  # val_check_interval: 1
  check_val_every_n_epoch: 10
  benchmark: False
  deterministic: False
  replace_sampler_ddp: True

model:
  _target_: src.models.Detector_effdet.Detector
  cfg:
    name: EfficientDet
    version: tf_efficientdet_d2
    decision_thres: 0.5
    lr: ${datamodule.cfg.lr}
    LR_Scheduler: custom
    sched_factor: 0.5
    sched_patience: 4
    weight_decay: 1e-5
    batch_size: ${datamodule.cfg.batch_size}
    new_shape: ${datamodule.cfg.new_shape}
    optim : Adam
    multiCrop: False
    combineCrops: 'mean'
    cropAttention: False
    pretrained: False
    pretrained_backbone: False
    replace_head: True
    num_classes: 1
    prediction_confidence_threshold: 0.05
    vindr_weights: True
    custom_optim: False
    LR_warmup: False
    vindr_path: /opt/algorithm/F1_E79_ModelX_v4_T0.325_V0.410.ckpt