train_config-res.yaml

version: v1.0.0
random_seed: 133
port: 11111

dataset:
  type: custom

  image_reader:
    type: npy
    kwargs:
      image_dir: ./chunks_scaled/

  train:
    meta_file: ./chunks_scaled/metadata/train_metadata.json
    rebalance: False
    hflip: True
    vflip: True
    rotate: True

  test:
    meta_file: ./chunks_scaled/metadata/test_metadata.json

  input_size: [224,224] # [h,w]
  batch_size: 32
  normals: [0]
  workers: 10 # number of workers of dataloader for each process


criterion:
  - name: FeatureMSELoss
    type: FeatureMSELoss
    kwargs:
      weight: 1.0

trainer:
  max_epoch: 250
  clip_max_norm: 0.1
  val_freq_epoch: 10
  print_freq_step: 200
  tb_freq_step: 500
  lr_scheduler:
    type: StepLR
    kwargs:
      step_size: 100
      gamma: 0.5
  optimizer:
    type: AdamW
    kwargs:
      lr: 0.0001
      betas: [0.9, 0.999]
      weight_decay: 0.0001

saver:
  auto_resume: False
  always_save: False
  load_path: UniAD/stitch-o_checkpoint/ckpt.pth.tar
  save_dir: UniAD/stitch-o_checkpoint/
  log_dir: UniAD/log/

evaluator:
  save_dir: result_eval_temp
  key_metric: mean_std_auc
  metrics:
    auc:
      - name: std
      - name: max
        kwargs:
          avgpool_size: [16, 16]
      - name: mean


frozen_layers: [backbone]

net:
    - name: backbone
      type: models.backbones.resnet50
      frozen: True
      kwargs:
        pretrained: True
        # select outlayers from: resnet [1,2,3,4], efficientnet [1,2,3,4,5]
        # empirically, for industrial: resnet [1,2,3] or [2,3], efficientnet [1,2,3,4] or [2,3,4]
        outlayers: [1,2,3]
    - name: neck
      prev: backbone
      type: models.necks.MFCN
      kwargs:
        outstrides: [16]
    - name: reconstruction
      prev: neck
      type: models.reconstructions.UniAD
      kwargs: 
        pos_embed_type: learned
        hidden_dim: 256
        nhead: 8
        num_encoder_layers: 4
        num_decoder_layers: 4
        dim_feedforward: 1024
        dropout: 0.0
        activation: relu
        normalize_before: False
        feature_jitter:
          scale: 0.0
          prob: 0.0
        neighbor_mask:
          neighbor_size: [7,7]
          mask: [True, True, True] # whether use mask in [enc, dec1, dec2]
        save_recon: False
        initializer:
          method: xavier_uniform