clipseg/experiments/ablation.yaml

configuration:
  batch_size: 64
  optimizer: torch.optim.AdamW

  lr: 0.001

  trainer: experiment_setup.train_loop
  scorer: experiment_setup.score
  model: models.clipseg.CLIPDensePredT

  lr_scheduler: cosine
  T_max: 20000
  eta_min: 0.0001

  max_iterations: 20000    #  <-##########################################
  val_interval: null

  # dataset
  dataset: datasets.phrasecut.PhraseCut   # <-----------------
  split_mode: pascal_test
  split: train
  mask: text_and_crop_blur_highlight352
  image_size: 352
  negative_prob: 0.2
  mix_text_max: 0.5

  # general
  mix: True # <-----------------
  prompt: shuffle+
  norm_cond: True
  mix_text_min: 0.0
  with_visual: True
  
  # model
  version: 'ViT-B/16'
  extract_layers: [3, 7, 9]
  reduce_dim: 64
  depth: 3
  fix_shift: False            #  <-##########################################

  loss: torch.nn.functional.binary_cross_entropy_with_logits
  amp: True

test_configuration_common:
  normalize: True
  image_size: 352
  batch_size: 32
  sigmoid: True
  split: test
  label_support: True
  
test_configuration: 

  -
    name: pc
    metric: metrics.FixedIntervalMetrics
    test_dataset: phrasecut
    mask: text

  -
    name: pc-vis
    metric: metrics.FixedIntervalMetrics
    test_dataset: phrasecut
    mask: crop_blur_highlight352
    with_visual: True
    visual_only: True


columns: [name, 
pc_fgiou_best, pc_miou_best,  pc_fgiou_0.5, 
pc-vis_fgiou_best, pc-vis_miou_best,  pc-vis_fgiou_0.5, 
duration]


individual_configurations:

- {name: rd64-uni}
- {name: rd64-no-pretrain, not_pretrained: True, lr: 0.0003}
- {name: rd64-no-negatives, negative_prob: 0.0}
- {name: rd64-neg0.5, negative_prob: 0.5}
- {name: rd64-no-visual, with_visual: False, mix: False}
- {name: rd16-uni, reduce_dim: 16}
- {name: rd64-layer3, extract_layers: [3], depth: 1}
- {name: rd64-blur-highlight, mask: text_and_blur_highlight, test_configuration: {mask: blur_highlight}}