seed: 42
model:
  pretrained: /workspace/pretrained_models/sdxl
  type: sdxl
  image_tokens: 4
  include_text_tokens: false
encoder:
  model: /workspace/pretrained_models/clip-vit-base-patch32
  type: hf_clip
  checkpoint: /workspace/SSL-Backdoor/ssl_backdoor/projects/CleanCLIP/logs/clip_10epoch_finetune/checkpoints/epoch.best.pt
  feature_mode: global
data:
  train_manifest: /tmp/defuse_in900_train.jsonl
  val_manifest: /tmp/defuse_in900_val.jsonl
  root: /workspace/dataset/imagenet-1k
  resolution: 224
  center_crop: true
  num_workers: 0
training:
  output_dir: /tmp/defuse_in900_best_encoder_30k_20260818
  batch_size: 1
  learning_rate: 0.0001
  weight_decay: 0.01
  max_steps: 30000
  gradient_accumulation_steps: 11
  mixed_precision: fp16
  checkpoint_interval: 3000
  checkpoint_keep_last: 2
  eval_interval: 2000
  max_grad_norm: null
  resume: null
evaluation:
  dino_model: /workspace/pretrained_models/dinov2-base
  num_samples: 100
  batch_size: 1
  run_at_start: true
  run_at_end: true
  seed: 1234
  inference_steps: 25
  save_images: 4
  dino_threshold: 0.1
reconstruction:
  checkpoint: null
  manifest: /tmp/defuse_in900_val.jsonl
  output_dir: /tmp/defuse_in900_best_encoder_30k_20260818-reconstruct
logging:
  backends:
  - local
  - wandb
  level: INFO
  train_interval: 1
  train_dino_interval: 200
  wandb:
    project: defuse
    entity: jsrdcht
    name: in900-best-encoder-no-grad-clip-30k-20260818
    group: in900-epoch1-regression
    tags:
    - formal-test
    - refactored
    - dual-attention
    - cached-empty-text
    - sdxl
    - clip-backdoor
    - in900
    - train-dino
    - best-encoder
    - no-grad-clip
    - 30k
    mode: online
    run_id: null
    resume: auto