project:
  name: imagenet-project

data:
  raw_dir: ./data/raw/
  captions_file: ./data/captioning/annotations/train.json
  dataset_version: cls_raw-20260525-v2
  # dataset_version: raw-20260509-v1

split:
  train_ratio: 0.7
  val_ratio: 0.15
  test_ratio: 0.15

train:
  seed: 42
  # repeated experiment
  # seed: 7
  # seed: 21
  epochs: 20
  batch_size: 32
  num_workers: 4
  device: cuda
  optimizer: adam

preprocess:
  image_size: 224
  normalize: true

loss:
  name: cross_entropy
  ignore_index: pad_token

evaluate:
  batch_size: 32
  metrics:
    - bleu
    - rouge_l
    - meteor

logging:
  use_wandb: true
  project_name: imagenet-project
  log_interval: 10

outputs:
  base_dir: outputs

demo:
  host: 0.0.0.0
  port: 7860
  share: false
  top_k: 5
  show_gradcam: true
  class_names : [airplane, apple, aster, banana, bicycle, bracelet, bulldog, bus, butterfly, car, carrot, cucumber, cup-cake, daisy, dandelion, dumpling, earrings, elephant, glasses, golden-retriever, hamburger, horse, iris, lavender, lily, marigold, motorcycle, necklace, orange, orchid, pants, pasta, penguin, persian-cat, pizza, rose, salad, sandwich, sheep, siamese-cat, sneakers, squirrel, steak, strawberry, sunflower, sushi, tomato, t-shirt, tulip, waffle]

cnn:
  backbone: resnet18
  pretrained: true
  freeze: true
  output_dim: 512
  dropout: 0.3
  pooling: avg

captioning:

  # encoder: resnet18
  encoder: swin
  # encoder: vit
  decoder: transformer
  # decoder: lstm
  # decoder: gru
  version: final

  epochs: 25
  learning_rate: 0.0001
  batch_size: 32
  optimizer: adamw
  max_caption_length: 30
  train_num_caption: 2

  debug: False

  lstm:
    embed_dim: 256
    hidden_dim: 512
    num_layers: 1

  gru:
    embed_dim: 256
    hidden_dim: 512
    num_layers: 1

  transformer:
    n_layers: 6
    nhead: 8
    d_model: 512
    drop_p: 0.3
    label_smoothing: 0
    weight_decay: 0.001

  data:
    dataset_version: cap_raw-20260524-v1
    train_img: ./data/captioning/raw/train/
    train_caption: ./data/captioning/annotations/train.json
    val_img: ./data/captioning/raw/val/
    val_caption: ./data/captioning/annotations/val.json
    test_img: ./data/captioning/raw/test/
    test_caption: ./data/captioning/annotations/test.json
    
  tokenizer:
    min_freq: 3
    max_vocab_size: 10000
    sp_vocab_size: 2000
    use_subword: False
    sp_model_path: ./src/dataset/sub_tokenizer2000.model

  checkpoint:
    save_dir: ./outputs/captioning
    final_checkpoint: swin-transformer_final_best.pt
    resume: False

  heatmap:
    dec_atten_dir: /workspace/outputs/captioning/heatmap/
    enc_dec_atten_dir: /workspace/outputs/captioning/heatmap/
    layer: 6 # 몇번째 층
    sample: [0, 410, 820, 1230, 1640] # caption & heatmap 몇번째 샘플(batch)

  scheduler:
    use_scheduler: False
    warmup_step: 500
    lr_scale: 0.5

  beam_search:
    use_beam_search: True
    beam_size: 3
    


classification:

  # model_name: resnet18
  # model_name: efficientnet_b0
  # model_name: convnext_tiny
  # model_name: mobilenet_v3_small
  # model_name: vit_b_16
  model_name: swin_t
  # model_name: deit_tiny_patch16_224

  final_checkpoint: ./outputs/classification/cls_swin-t_base_cls_raw-20260525-v2_lr-0005_bs-32_adamw_none_wdc-0.05_ls-0.0_best.pth

  epochs: 50

  learning_rate:

    # baseline
    cnn: 0.001
    transformer: 0.0005

    # hyperparameter tuning
    # cnn: 0.0005
    # transformer: 0.0001

  # optimizer: adam
  # optimizer: sgd
  optimizer: adamw

  # default
  # weight_decay: 0.01

  # tuning
  weight_decay: 0.05

  scheduler:
    use: false

    # use: true
    # name: cosineannealinglr

  augmentation:

    # baseline
    use_aug: false
    type: none

    # mixup
    # use_aug: true
    # type: mixup

    # cutmix
    # use_aug: true
    # type: cutmix

  label_smoothing: 0.0

  # label smoothing experiment
  # label_smoothing: 0.05
  # label_smoothing: 0.1


  metrics:

    train:
      - loss
      - accuracy

    validation:
      - loss
      - accuracy
      - macro_f1

    final_test:
      - accuracy
      - macro_f1
      - precision
      - recall
      - confusion_matrix

  checkpoint:
    save_dir: /workspace/outputs/classification


latent_space:
  data_dir: /workspace/data/raw
  checkpoint: /workspace/outputs/classification/cls_swin-t_base_cls_raw-20260525-v2_lr-0005_bs-32_adamw_none_wdc-0.05_ls-0.0_best.pth
  output_dir: /workspace/outputs/latent_space
  output_umap_npy: cls_swin-t_best_umap_2d_test_nb10_md05
  output_umap_png: cls_swin-t_best_umap_plt_test_nb10_md05
  output_meta_csv: cls_swin-t_best_metadata_test_nb10_md05
  split: test
  batch_size: 32
  num_workers: 4
  device: cuda
  seed: 42
  save_meta: true
  use_wandb: true
  wandb_name: latent_space_umap

  umap:
    n_neighbors: 10
    min_dist: 0.5
    metric: cosine