configs/detector2.yaml

TASK: custom_laanet
PRECISION: float32
METRICS_BASE: binary
SEED: 317
DATA_RELOAD: False
Resume: True
begin_epoch: 100

MODEL:
  # PRETRAINED_PATH: '' 
  type: PoseEfficientNet
  model_name: efficientnet-b4
  num_layers: B4
  include_top: False
  include_hm_decoder: True
  head_conv: 64
  use_c2: False
  use_c3: True
  use_c4: True
  use_c51: True
  efpn: True
  tfpn: False
  se_layer: False
  heads:
    hm: 1
    cls: 1
    cstency: 256
  INIT_WEIGHTS:
    pretrained: True
    advprop: True

DATASET:
  type: BinaryFaceForensic
  FROM_FILE: False
  PIN_MEMORY: True
  NUM_WORKERS: 7
  COLOR_NORM: 'simple'  
  mean: [0.485, 0.456, 0.406]
  std: [0.229, 0.224, 0.225]
  IMAGE_SUFFIX: png
  COMPRESSION: c0
  IMAGE_SUFFIX: png
  IMAGE_SIZE: [384, 384]
  HEATMAP_SIZE: [96, 96] #[IMAGE_SIZE//4, IMAGE_SIZE//4]
  SIGMA: 2
  ADAPTIVE_SIGMA: True
  HEATMAP_TYPE: gaussian
  SPLIT_IMAGE: False
  DATA:
    TYPE: frames
    SAMPLES_PER_VIDEO: 
      ACTIVE: True
      TRAIN: 8
      VAL: 8
      TEST: 32
    TRAIN:
      NAME: custom_dataset
      ROOT: ./datasets/train/
      FROM_FILE: False
      FAKETYPE: [fake]
      LABEL_FOLDER: [real, fake]
      
    VAL:
      NAME: custom_dataset
      ROOT: ./datasets/test/
      FROM_FILE: False
      FAKETYPE: [fake]
      LABEL_FOLDER: [real, fake]
    TEST:
      NAME: custom_dataset
      ROOT: ./datasets/test/
      FROM_FILE: False
      FAKETYPE: [fake]
      LABEL_FOLDER: [real, fake]
  TRANSFORM:
    geometry:
      type: GeometryTransform
      resize: [384, 384, 0] #h, w, p=probability. If no affine transform, set p=1
      normalize: 0
      horizontal_flip: 0.5
      cropping: [0.15, 0.5] #Format: [crop_limit, prob]
      scale: [0.15, 0.5] #Format: [scale_limit, prob]
      rand_erasing: [0.5, 1] #Format: [p, max_count]
    color:
      type: ColorJitterTransform
      clahe: 0.0
      colorjitter: 0.3
      gaussianblur: 0.3
      gaussnoise: 0.3
      jpegcompression: [0.5, 40, 100] # prob, lower and upper quality respectively
      rgbshift: 0.3
      randomcontrast: 0.0
      randomgamma: 0.5
      randombrightness: 1
      huesat: 1
    normalize:
      mean: [0.5, 0.5, 0.5]
      std: [0.5, 0.5, 0.5]

TRAIN:
  resume: True
  gpus: [0]
  pretrained_model: './logs/27-03-2025/PoseEfficientNet_custom_laanet_model_final.pth'
  batch_size: 32
  lr: 0.00005
  epochs: 150
  begin_epoch: 100
  warm_up: 6
  every_val_epochs: 1
  loss:
    type: CombinedFocalLoss
    use_target_weight: False
    cls_lmda: 1
    dst_hm_cls_lmda: 0
    offset_lmda: 0
    hm_lmda: 100
    cstency_lmda: 100
    mse_reduction: sum
    ce_reduction: mean
  optimizer: SAM
  distributed: False
  tensorboard: False
  resume: True
  lr_scheduler:
    # type: MultiStepLR
    milestones: [5, 15, 20, 25]
    gamma: 0.5
  freeze_backbone: True
  debug:
    active: False
    save_hm_gt: True
    save_hm_pred: True

TEST:
  gpus: [0]
  subtask: 'eval'
  test_file: ''
  vis_hm: True
  threshold: 0.5
  flip_test: True
  video_level: True
  pretrained: './training/weights/final_model.pth'