"""Constants shared by every step. Nothing here reads data.""" from pathlib import Path ROOT = Path(__file__).resolve().parents[2] DATA_RAW = ROOT / "data" / "raw" REPORTS = ROOT / "reports" # Training data, pinned to one commit of the Hub dataset (sha256 checked on load). HF_DATASET = "CMU-OSPO/protostar-climate-train-to-2018" HF_REVISION = "bbbdebdab2a18e00f97ef87bd3f35a71e222180b" FILES = { # local name -> (path in the Hub repo, sha256) "annual_indices.parquet": ( "annual_indices/train-00000-of-00001.parquet", "0b5f5fa9b55ab59602dcd9884b97c74c7bec7e89ea9d23f4219703046030e411"), "city_info.parquet": ( "city_info/train-00000-of-00001.parquet", "6b71424c7c30a1b8027ba47a92ccb9aab8553147950af491538ea0912bbb955d"), "coverage_and_stations.parquet": ( "coverage_and_stations/train-00000-of-00001.parquet", "d52dfcacd8eae99e5bd6c807f34dfe16f8f22e7688e5718a1ba93c8d93ca07e6"), "daily.parquet": ( # 49 MB: audit only, not committed, fetched on demand "daily/train-00000-of-00001.parquet", "1a85791055019f25f759ba23f33b326675d46a28278d4eb56a7e9845fe494361"), } # The data rule: nothing dated after this year may be a forecast input. TRAIN_END = 2018 FORECAST_YEARS = tuple(range(2019, 2026)) N_CITIES = 210 # metric code -> (column in annual_indices, missing-day columns of the 10-day rule) METRICS = { "annual_avg_temp": ("Avg Temp", ("Miss Tmax", "Miss Tmin")), "total_precip_in": ("ToT Prcp", ("Miss Prcp",)), "hottest_day_temp": ("Max Tmax", ("Miss Tmax", "Miss Tmin")), "cdd": ("cdd", ("Miss Tmax", "Miss Tmin")), } UNITS = {"annual_avg_temp": "°F", "total_precip_in": "in", "hottest_day_temp": "°F", "cdd": "°F-days"} # Organizers' rules (Space "About" tab, dataset card, example notebook). MAX_MISSING_DAYS = 10 # a year is valid if each variable the metric uses misses <= 10 days MIN_VALID_YEARS = 10 # a city-metric has a baseline (is scored) with >= 10 valid years in 30 BASELINE_WINDOW = 30 # "30-year average" and "linear trend" use the last 30 years # Backtest folds: (origin = last training year, first test year, last test year). # Selection folds have the same 7-year horizon as 2018 -> 2019-2025. SELECTION_FOLDS = ((2011, 2012, 2018), (2004, 2005, 2011)) # Veto only: these folds can reject a candidate (if it loses to the 30-year average on them), # they never rank candidates. CONFIRMATION_FOLDS = ((1997, 1998, 2004), (1990, 1991, 1997)) # The organizers' practice run (train <= 2010, test 2011-2018): used to verify the harness. PRACTICE_FOLD = (2010, 2011, 2018) SEED = 20181231