pasqualedem/DistillFSS_WeedMap_DCAMA_5shot

DistillFSS-DCAMA is a distilled version of the DCAMA model for a specific downstream segmentation task. The DistillFSS framework allows to distill large few-shot segmentation models into smaller and more efficient ones, while improving or maintaining their performance on the target task.

• Code: https://github.com/pasqualedem/DistillFSS
• Paper: https://arxiv.org/abs/2512.05613

How to use this model: Clone the repository:

git clone https://github.com/pasqualedem/DistillFSS.git

Install the required dependencies as specified in the repository.

Load the model using the following code snippet:

from distillfss.models.dcama.distillator import DistilledDCAMA
model = DistilledDCAMA.from_pretrained("pasqualedem/DistillFSS_WeedMap_DCAMA_5shot")

YAML configuration:

dataloader:
  num_workers: 0
dataset:
  datasets:
    test_weedmap:
      prompt_images: 5
      test_root: data/weedmap/0_rotations_processed_003_test/RedEdge/003
      train_root: data/weedmap/0_rotations_processed_003_test/RedEdge/000
  preprocess:
    image_size: 384
    mean:
    - 0.485
    - 0.456
    - 0.406
    std:
    - 0.229
    - 0.224
    - 0.225
model:
  name: distillator
  params:
    student:
      name: conv_distillator
      num_classes: 2
    teacher:
      backbone: swin
      backbone_checkpoint: checkpoints/swin_base_patch4_window12_384.pth
      concat_support: false
      image_size: 384
      model_checkpoint: checkpoints/swin_fold0_pascal_modcross_soft.pt
      name: dcama
push_to_hub:
  repo_name: pasqualedem/DistillFSS_WeedMap_DCAMA_5shot
refinement:
  hot_parameters:
  - model.conv1
  - model.conv2
  - model.conv3
  - model.mixer1
  - model.mixer2
  - model.mixer3
  - student
  iterations_is_num_classes: false
  loss:
    name: refine_distill
  lr: 0.001
  max_iterations: 500
  subsample: 1
  substitutor: paired
test:
  prompt_to_use: null
tracker:
  cache_dir: tmp
  group: WeedMap
  log_frequency: 1
  project: FSSWeed
  tags:
  - WeedMap
  - Distill
  test_image_log_frequency: 10
  tmp_dir: tmp
  train_image_log_frequency: 25