Skip to content

README.md

Latest commit

 

History

History
37 lines (29 loc) · 1.58 KB

README.md

File metadata and controls

37 lines (29 loc) · 1.58 KB
Raw

Sampling

Trains a model while growing the labeled set with a chosen sampling method. Adapted from TypiClust's deep-al module. Requires a MOSEK license (used by cvxpy for cost-aware optimization).

  1. Edit a config in configs/{dataset}/*.yaml, especially DATASET.ROOT_DIR.
  2. Run from scripts/:
python train_ridge.py \
  --cfg ../configs/usavars/population.yaml \
  --exp-name experiment_1 \
  --sampling_fn greedycost \
  --budget 1000 \
  --seed 42 \
  --cost_func uniform --cost_name uniform \
  --unit_assignment_path ../../0_data/groups/usavars_pop/counties_assignments.pkl \
  --group_assignment_path ../../0_data/groups/usavars_pop/states_assignments.pkl \
  --group_type states

Key optional arguments: --cost_array_path, --unit_assignment_path, --region_assignment_path, --util_lambda, --alpha, --points_per_unit. See train.py --help / train_ridge.py --help for the full list.

Sweep scripts

scripts/shell/ has sweeps configured via environment variables (see the header comment in each file for the full variable list):

Script Sweeps
base.sh method x budget x seed, no initial set
rep_sampling.sh representative sampling (states / image clusters / NLCD), grouped by GROUP_TYPE
multiple_initial_sets.sh initial-set size x cost function
cost_sensitivity.sh alpha x group-assignment 
DATASET=togo CFG=../configs/togo/RIDGE.yaml METHODS="poprisk" BUDGETS="100 500" \
  ./scripts/shell/base.sh