clipseg/experiments/coco.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.coco_wrapper.COCOWrapper
  # split_mode: pascal_test
  split: train
  mask: text_and_blur3_highlight01
  image_size: 352
  normalize: True
  pre_crop_image_size: [sample, 1, 1.5]
  aug: 1new

  # general
  mix: True
  prompt: shuffle+
  norm_cond: True
  mix_text_min: 0.0
  
  # model
  out: 1
  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
  # max_iterations: 10
  batch_size: 8
  sigmoid: True
  test_dataset: coco
  metric: metrics.FixedIntervalMetrics

test_configuration: 

  -
    name: coco_t
    mask: text
    
  -
    name: coco_h
    mask: blur3_highlight01

  -
    name: coco_h2
    mask: crop_blur_highlight352


columns: [i, name,
coco_t_fgiou_best, coco_t_miou_best,  coco_t_fgiou_0.5, 
coco_h_fgiou_best, coco_h_miou_best,  coco_h_fgiou_0.5,
coco_h2_fgiou_best, coco_h2_miou_best,  coco_h2_fgiou_0.5, coco_h2_fgiou_best_t,
train_loss, duration, date
]

individual_configurations:


- {name: rd64-7K-vit16-cbh-coco-0, version: 'ViT-B/16', fold: 0, reduce_dim: 64, mask: text_and_crop_blur_highlight352, T_max: 7000, max_iterations: 7000}
- {name: rd64-7K-vit16-cbh-coco-1, version: 'ViT-B/16', fold: 1, reduce_dim: 64, mask: text_and_crop_blur_highlight352, T_max: 7000, max_iterations: 7000}
- {name: rd64-7K-vit16-cbh-coco-2, version: 'ViT-B/16', fold: 2, reduce_dim: 64, mask: text_and_crop_blur_highlight352, T_max: 7000, max_iterations: 7000}
- {name: rd64-7K-vit16-cbh-coco-3, version: 'ViT-B/16', fold: 3, reduce_dim: 64, mask: text_and_crop_blur_highlight352, T_max: 7000, max_iterations: 7000}


- {name: rd64-7K-vit16-cbh-neg0.2-coco-0, version: 'ViT-B/16', negative_prob: 0.2, fold: 0, reduce_dim: 64, mask: text_and_crop_blur_highlight352, T_max: 7000, max_iterations: 7000}
- {name: rd64-7K-vit16-cbh-neg0.2-coco-1, version: 'ViT-B/16', negative_prob: 0.2, fold: 1, reduce_dim: 64, mask: text_and_crop_blur_highlight352, T_max: 7000, max_iterations: 7000}
- {name: rd64-7K-vit16-cbh-neg0.2-coco-2, version: 'ViT-B/16', negative_prob: 0.2, fold: 2, reduce_dim: 64, mask: text_and_crop_blur_highlight352, T_max: 7000, max_iterations: 7000}
- {name: rd64-7K-vit16-cbh-neg0.2-coco-3, version: 'ViT-B/16', negative_prob: 0.2, fold: 3, reduce_dim: 64, mask: text_and_crop_blur_highlight352, T_max: 7000, max_iterations: 7000}


# ViT
- {name: vit64-7K-vit16-cbh-coco-0, version: 'ViT-B/16', model: models.vitseg.VITDensePredT, fold: 0, reduce_dim: 64, mask: text_and_crop_blur_highlight352, T_max: 7000, max_iterations: 7000, lr: 0.0001}
- {name: vit64-7K-vit16-cbh-coco-1, version: 'ViT-B/16', model: models.vitseg.VITDensePredT, fold: 1, reduce_dim: 64, mask: text_and_crop_blur_highlight352, T_max: 7000, max_iterations: 7000, lr: 0.0001}
- {name: vit64-7K-vit16-cbh-coco-2, version: 'ViT-B/16', model: models.vitseg.VITDensePredT, fold: 2, reduce_dim: 64, mask: text_and_crop_blur_highlight352, T_max: 7000, max_iterations: 7000, lr: 0.0001}
- {name: vit64-7K-vit16-cbh-coco-3, version: 'ViT-B/16', model: models.vitseg.VITDensePredT, fold: 3, reduce_dim: 64, mask: text_and_crop_blur_highlight352, T_max: 7000, max_iterations: 7000, lr: 0.0001}


# BASELINE
- {name: bl64-7K-vit16-cbh-neg0.2-coco-0, model: models.clipseg.CLIPDenseBaseline, reduce2_dim: 64, version: 'ViT-B/16', negative_prob: 0.2, fold: 0, reduce_dim: 64, mask: text_and_crop_blur_highlight352, T_max: 7000, max_iterations: 7000}
- {name: bl64-7K-vit16-cbh-neg0.2-coco-1, model: models.clipseg.CLIPDenseBaseline, reduce2_dim: 64, version: 'ViT-B/16', negative_prob: 0.2, fold: 1, reduce_dim: 64, mask: text_and_crop_blur_highlight352, T_max: 7000, max_iterations: 7000}
- {name: bl64-7K-vit16-cbh-neg0.2-coco-2, model: models.clipseg.CLIPDenseBaseline, reduce2_dim: 64, version: 'ViT-B/16', negative_prob: 0.2, fold: 2, reduce_dim: 64, mask: text_and_crop_blur_highlight352, T_max: 7000, max_iterations: 7000}
- {name: bl64-7K-vit16-cbh-neg0.2-coco-3, model: models.clipseg.CLIPDenseBaseline, reduce2_dim: 64, version: 'ViT-B/16', negative_prob: 0.2, fold: 3, reduce_dim: 64, mask: text_and_crop_blur_highlight352, T_max: 7000, max_iterations: 7000}