yimingzhang's picture
br-subsample
ee5764c
import gzip
import json
import random
from tqdm.auto import tqdm
from concurrent.futures import ProcessPoolExecutor
MONTHS = ["01", "02", "03", "04", "05", "06", "07", "08", "09", "10", "11", "12"]
def determine_format(s):
for format in ["Bullet", "Blitz", "Rapid", "Classical", "Correspondence"]:
if format in s:
return format
else:
return ""
subsample_ratios = {
1700.0: 0.03773727310464546,
1800.0: 0.037873049537948796,
1600.0: 0.03882816595158128,
1900.0: 0.04081466062609689,
1500.0: 0.04225024822020829,
1400.0: 0.04841208365608056,
2000.0: 0.050853060082890485,
1300.0: 0.05542161997395184,
1200.0: 0.0683480281593876,
2100.0: 0.07245851749873197,
1100.0: 0.08971023593792052,
2200.0: 0.1211827435773146,
1000.0: 0.12567550584391102,
900.0: 0.22724690376093626,
2300.0: 0.24479804161566707,
800.0: 0.47438330170777987,
2400.0: 0.5417118093174431,
700.0: 1.0,
2500.0: 1.0,
600.0: 1.0,
2600.0: 1.0,
2700.0: 1.0,
2800.0: 1.0,
2900.0: 1.0
}
def checks(game: dict) -> bool:
format = determine_format(game["event"])
white_elo = int(game["white-elo"])
black_elo = int(game["black-elo"])
elo_diff = abs(white_elo - black_elo)
elo_mean = (white_elo + black_elo) / 2
subsample_ratio = subsample_ratios.get(
(elo_mean) // 100 * 100,
1.0
)
n_plys = len(game["moves-uci"].split())
return format in ("Blitz", "Rapid") and \
elo_diff <= 100 and \
random.random() < subsample_ratio and \
n_plys >= 4
def filter_month(month: str):
train = gzip.open(f"/data/datasets/models/hf_cache/tmp-lichess/lichess-2022/br-subsample/train-2022-{month}.jsonl.gz", "wb")
val = gzip.open(f"/data/datasets/models/hf_cache/tmp-lichess/lichess-2022/br-subsample/val-2022-{month}.jsonl.gz", "wb")
test = gzip.open(f"/data/datasets/models/hf_cache/tmp-lichess/lichess-2022/br-subsample/test-2022-{month}.jsonl.gz", "wb")
with gzip.open(f"/data/datasets/models/hf_cache/tmp-lichess/lichess-2022/all/2022-{month}.jsonl.gz", "rb") as f:
for i, line in tqdm(enumerate(f), desc=month):
d = json.loads(line)
if checks(d):
if random.random() < 1e-4:
val.write((json.dumps(d) + '\n').encode())
elif random.random() < 2e-4:
test.write((json.dumps(d) + '\n').encode())
else:
train.write((json.dumps(d) + '\n').encode())
return 0
ppe = ProcessPoolExecutor(12)
for _ in ppe.map(filter_month, MONTHS):
...
ppe.shutdown()
print("Done!")