SOWA/configs/model/sparc_linear.yaml

_target_: src.models.anomaly_clip_module.AnomalyCLIPModule

optimizer:
  _target_: torch.optim.AdamW
  _partial_: true
  lr: 0.001
  weight_decay: 0.2

scheduler:
  _target_: torch.optim.lr_scheduler.ReduceLROnPlateau
  _partial_: true
  mode: min
  factor: 0.1
  patience: 5

# scheduler:
#   _target_: src.models.components.scheduler.WarmupCosineAnnealingLR
#   _partial_: true
#   warmup_epochs: 10
#   total_epoch: 50

net:
  _target_: src.models.components.sparc.SPARC
  arch: ViT-L/14@336px
  image_size: 336
  temperature: 0.07                     # softmax
  feature_map_idx: [5, 11, 17, 23]      # [0, 12, 23] [6, 12, 18] [5, 11, 17, 23] index of resnetblock in ViT
  prompt_learner:
    _target_: src.models.components.coop.AnomalyPromptLearner
    _partial_: true
    tokenizer:
      _target_: src.models.components.clip.simple_tokenizer.SimpleTokenizer
    prompt_length: 12                     # length of learnable prompts
    context_length: 77                    # defaut 77 for openai clip
    truncate: false
    class_names: ["object"]
    # class_names: ${prompt.class_names}
    # state_template: ${prompt.state_template}
    state_template: 
      normal: ["{}"]
      anomaly: ["damaged {}"]
  text_encoder:
    _target_: src.models.components.text_encoder.TextMapEncoder
    _partial_: true
  adapter:
    # _target_: torch.nn.Linear
    # in_features: 1024                   # clip vit feature dim, defaut 1024 for openai clip
    # out_features: 1024
    # bias: false
    _target_: src.models.components.adapter.Linear
    in_features: 1024                     # clip vit feature dim, defaut 1024 for openai clip
    out_features: 1024
    hidden_features: null                 # set null, same as nn.Linear
    dropout_prob: 0.0
    bias: false
  fusion:
    _target_: src.models.components.cross_modal.DotProductFusion
  embedding_dim: 768                    # clip fusion featrue dim, default 768, only effective for non null

loss:
  cross_entropy:
    _target_: torch.nn.CrossEntropyLoss
  focal:
    _target_: src.models.components.loss.FocalLoss
  dice:
    _target_: src.models.components.loss.BinaryDiceLoss

k_shot: false

filter: true

enable_validation: false

compile: false