clipseg/experiments/pascal_1shot.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
  mode: train
  mask: text_and_crop_blur_highlight352
  image_size: 352
  normalize: True
  pre_crop_image_size: [sample, 1, 1.5]
  aug: 1new
  with_visual: True
  split: train

  # general
  mix: True
  prompt: shuffle+
  norm_cond: True
  mix_text_min: 0.0
  
  # model
  out: 1
  version: 'ViT-B/16'
  extract_layers: [3, 7, 9]
  reduce_dim: 64
  depth: 3

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

test_configuration_common:
  normalize: True
  image_size: 352
  metric: metrics.FixedIntervalMetrics
  batch_size: 1
  test_dataset: pascal
  sigmoid: True
  # max_iterations: 250

test_configuration: 

  -
    name: pas_t
    mask: text

  -
    name: pas_h
    mask: blur3_highlight01

  -
    name: pas_h2
    mask: crop_blur_highlight352


columns: [name,
pas_t_fgiou_best, pas_t_miou_best,  pas_t_fgiou_ct,
pas_h_fgiou_best, pas_h_miou_best,  pas_h_fgiou_ct,
pas_h2_fgiou_best, pas_h2_miou_best,  pas_h2_fgiou_ct, pas_h2_fgiou_best_t,
train_loss, duration, date
]

individual_configurations:

- {name: rd64-uni-phrasepas5i-0, remove_classes: [pas5i, 0], negative_prob: 0.2, mix_text_max: 0.5, test_configuration: {splits: [0], custom_threshold: 0.24}}
- {name: rd64-uni-phrasepas5i-1, remove_classes: [pas5i, 1], negative_prob: 0.2, mix_text_max: 0.5, test_configuration: {splits: [1], custom_threshold: 0.24}}
- {name: rd64-uni-phrasepas5i-2, remove_classes: [pas5i, 2], negative_prob: 0.2, mix_text_max: 0.5, test_configuration: {splits: [2], custom_threshold: 0.24}}
- {name: rd64-uni-phrasepas5i-3, remove_classes: [pas5i, 3], negative_prob: 0.2, mix_text_max: 0.5, test_configuration: {splits: [3], custom_threshold: 0.24}}


- {name: rd64-phrasepas5i-0, remove_classes: [pas5i, 0], negative_prob: 0.0, test_configuration: {splits: [0], custom_threshold: 0.28}}
- {name: rd64-phrasepas5i-1, remove_classes: [pas5i, 1], negative_prob: 0.0, test_configuration: {splits: [1], custom_threshold: 0.28}}
- {name: rd64-phrasepas5i-2, remove_classes: [pas5i, 2], negative_prob: 0.0, test_configuration: {splits: [2], custom_threshold: 0.28}}
- {name: rd64-phrasepas5i-3, remove_classes: [pas5i, 3], negative_prob: 0.0, test_configuration: {splits: [3], custom_threshold: 0.28}}


# baseline
- {name: bl64-phrasepas5i-0, model: models.clipseg.CLIPDenseBaseline, remove_classes: [pas5i, 0], reduce2_dim: 64, negative_prob: 0.0, test_configuration: {splits: [0], custom_threshold: 0.24}}
- {name: bl64-phrasepas5i-1, model: models.clipseg.CLIPDenseBaseline, remove_classes: [pas5i, 1], reduce2_dim: 64, negative_prob: 0.0, test_configuration: {splits: [1], custom_threshold: 0.24}}
- {name: bl64-phrasepas5i-2, model: models.clipseg.CLIPDenseBaseline, remove_classes: [pas5i, 2], reduce2_dim: 64, negative_prob: 0.0, test_configuration: {splits: [2], custom_threshold: 0.24}}
- {name: bl64-phrasepas5i-3, model: models.clipseg.CLIPDenseBaseline, remove_classes: [pas5i, 3], reduce2_dim: 64, negative_prob: 0.0, test_configuration: {splits: [3], custom_threshold: 0.24}}

# ViT
- {name: vit64-uni-phrasepas5i-0, remove_classes: [pas5i, 0], model: models.vitseg.VITDensePredT, negative_prob: 0.2, mix_text_max: 0.5, lr: 0.0001, test_configuration: {splits: [0], custom_threshold: 0.02}}
- {name: vit64-uni-phrasepas5i-1, remove_classes: [pas5i, 1], model: models.vitseg.VITDensePredT, negative_prob: 0.2, mix_text_max: 0.5, lr: 0.0001, test_configuration: {splits: [1], custom_threshold: 0.02}}
- {name: vit64-uni-phrasepas5i-2, remove_classes: [pas5i, 2], model: models.vitseg.VITDensePredT, negative_prob: 0.2, mix_text_max: 0.5, lr: 0.0001, test_configuration: {splits: [2], custom_threshold: 0.02}}
- {name: vit64-uni-phrasepas5i-3, remove_classes: [pas5i, 3], model: models.vitseg.VITDensePredT, negative_prob: 0.2, mix_text_max: 0.5, lr: 0.0001, test_configuration: {splits: [3], custom_threshold: 0.02}}