/laa

Primary LanguagePython

Aggregating Crowd Wisdoms with Label-aware Autoencoders

Paper

How inference works:

  • Split the dataset into training and validation parts
  • Start training model
  • When loss on validation part starts to increase authors stop training
    • No explicit stopping strategy has been given by the authors
    • So we stop training after max patience step succeeded
  • To find a final solution we should use hyperparameters search using validation loss

Accuracy - validation loss:

dataset LAA-B Majority Vote LAA-B(Paper)
bluebirds 0.8056 0.7593 0.889
syntetic overlap 3 n_classes 3 n_tasks 5000 0.8932 0.9006
syntetic overlap 2 n_classes 3 n_tasks 5000 0.777 0.7892

P.S.

It looks like you could achieve greater results in case of using golden labeled data for the best model search:

Accuracy - validation via golden dataset:

dataset LAA-B Majority Vote
bluebirds 0.907 0.7674
syntetic overlap 3 n_classes 3 n_tasks 5000 0.8998 0.8985
syntetic overlap 2 n_classes 3 n_tasks 5000 0.8085 0.7822

Wandb Sweeps and best run

python -m src inference --dataset-name=classification_dataset_generator  --no-logging --dataset-name=bluebirds --batch-size=33 --d-kl=0.0014471807379961906 --lr=0.1038585651189214 --n-epoch=152 --patience=15 --reg-1=6.420109871402643e-05
python -m src inference --dataset-name=classification_dataset_generator --dataset-kwargs="{\"n_workers\": 100, \"n_tasks\": 5000, \"overlap\": 3, \"n_classes\": 3, \"good_probability\": 0.9, \"good_workers_frac\": 0.6, \"bad_probability\": 0.6}" --batch-size=93 --d-kl=0.007161637022033341 --lr=0.12966202803623433 --n-epoch=78 --patience=2 --reg-1=4.9742867272127485e-05 --no-logging
python -m src inference --dataset-name=classification_dataset_generator --dataset-kwargs="{\"n_workers\": 100, \"n_tasks\": 5000, \"overlap\": 2, \"n_classes\": 3, \"good_probability\": 0.9, \"good_workers_frac\": 0.6, \"bad_probability\": 0.6}" --batch-size=86 --d-kl=0.009459693322090797 --lr=0.003334278174792921 --n-epoch=242 --patience=8 --reg-1=0.0001390032305118152 --no-logging