src/preprocessing/preprocess_17lands.py

import csv
import gzip
import pickle
import hashlib
import os
import shutil
import uuid
import multiprocessing as mp
from tqdm import tqdm
import numpy as np
import lmdb
import requests

from src.utils import utils

ctx = mp.get_context("spawn")  # safer with LMDB/pickle
Process = ctx.Process
Queue = ctx.Queue

def _is_test(draft_id: str, pct=10):
    h = int.from_bytes(hashlib.sha1(draft_id.encode()).digest()[:4], "little")
    return h % 100 < pct

def lmdb_writer(db_path, q: Queue, commit_every=200_000, map_size=int(16e9)):
    env = lmdb.open(db_path, map_size=map_size, subdir=True, lock=True,
                    sync=False, metasync=False, readahead=False, writemap=False)  # writemap=False to avoid big prealloc on some FS
    txn, n = env.begin(write=True), 0
    while True:
        item = q.get()
        if item is None:
            break
        key, value = item
        try:
            txn.put(key, value)
        except lmdb.MapFullError:
            txn.abort()
            env.set_mapsize(env.info()["map_size"] * 2)
            txn = env.begin(write=True)
            txn.put(key, value)
        n += 1
        if n % commit_every == 0:
            txn.commit(); txn = env.begin(write=True)
    txn.commit()
    with env.begin(write=True) as t2:
        t2.put(b"__len__", str(n).encode())
    env.sync(); env.close()

def get_name(obstructed_name):
    if 'pack_card_' in obstructed_name:
        return obstructed_name.replace('pack_card_','')
    elif 'pool_' in obstructed_name:
        return obstructed_name.replace('pool_','')
    else:
        print(obstructed_name)
        raise ValueError('Invalid name')

def encode_sample(sample):
    return pickle.dumps(sample, protocol=pickle.HIGHEST_PROTOCOL)

def stream_csv(csv_path, train_q, test_q):
    with open(csv_path, "r", newline="", buffering=16*1024*1024) as f:
        reader = csv.reader(f)
        header = next(reader)

        pack_idx  = [i for i,c in enumerate(header) if c.startswith("pack_card_")]
        pool_idx  = [i for i,c in enumerate(header) if c.startswith("pool_")]
        pick_idx  = header.index("pick")
        wins_idx  = header.index("event_match_wins")
        loss_idx  = header.index("event_match_losses")
        try:
            user_games = header.index("user_n_games_bucket")
        except: 
            user_games = header.index("user_n_matches_bucket")
        try:
            user_wr = header.index("user_game_win_rate_bucket")
        except:
            try:
                user_wr = header.index("user_match_win_rate_bucket")
            except:
                user_wr = None
        draft_idx = 2  # your draft id column

        pack_names = {i: header[i].replace("pack_card_","") for i in pack_idx}
        pool_names = {i: header[i].replace("pool_","") for i in pool_idx}

        for row in tqdm(reader):
            try:
                draft_id = row[draft_idx]
                positive = row[pick_idx]
            except:
                print(f"Skipping row with missing draft_id or pick: {row}")
                continue

            negatives = []
            for i in pack_idx:
                s = row[i]
                if s and s != "0":
                    cnt = int(s) 
                    name = pack_names[i]
                    if name != positive:
                        negatives.extend([name]*cnt)

            anchor = []
            for i in pool_idx:
                s = row[i]
                if s and s != "0":
                    anchor.extend([pool_names[i]]*int(s))

            wins   = int(row[wins_idx])
            losses = int(row[loss_idx])
            u_g = int(row[user_games])
            u_wr = float(row[user_wr]) if user_wr and row[user_wr] else 0.0

            payload = encode_sample((positive, negatives, anchor, wins, losses, u_g, u_wr))
            key = os.urandom(16)
            (test_q if _is_test(draft_id) else train_q).put((key, payload))




_17LANDS_URL = (
    "https://17lands-public.s3.amazonaws.com/analysis_data/draft_data/"
    "draft_data_public.{set_tag}.{draft_format}.csv.gz"
)

def download_17lands(set_tag, raw_data_folder, draft_format="PremierDraft"):
    url = _17LANDS_URL.format(set_tag=set_tag, draft_format=draft_format)
    out_dir = os.path.join(raw_data_folder, set_tag)
    os.makedirs(out_dir, exist_ok=True)
    csv_path = os.path.join(out_dir, f"{set_tag}_{draft_format}.csv")

    if os.path.exists(csv_path):
        print(f"{csv_path} already exists, skipping download")
        return

    print(f"Downloading {url}")
    response = requests.get(url, stream=True)
    if response.status_code == 404:
        raise FileNotFoundError(
            f"17lands data not found for {set_tag}/{draft_format}.\n"
            f"URL tried: {url}\n"
            f"Check https://17lands.com/public/data for the correct set code and format."
        )
    response.raise_for_status()

    gz_path = csv_path + ".gz"
    total = int(response.headers.get("content-length", 0))
    with open(gz_path, "wb") as f, tqdm(total=total, unit="B", unit_scale=True,
                                         desc=f"Downloading {set_tag}") as bar:
        for chunk in response.iter_content(chunk_size=1 << 17):
            f.write(chunk)
            bar.update(len(chunk))

    print(f"Decompressing {gz_path}")
    with gzip.open(gz_path, "rb") as f_in, open(csv_path, "wb") as f_out:
        shutil.copyfileobj(f_in, f_out)
    os.remove(gz_path)
    print(f"Saved to {csv_path}")


def stream_csv_trajectories(csv_path, train_q, test_q):
    """Buffer all picks by draft_id, then write one trajectory per draft."""
    drafts = {}

    with open(csv_path, "r", newline="", buffering=16*1024*1024) as f:
        reader = csv.reader(f)
        header = next(reader)

        pack_idx  = [i for i,c in enumerate(header) if c.startswith("pack_card_")]
        pick_idx  = header.index("pick")
        wins_idx  = header.index("event_match_wins")
        loss_idx  = header.index("event_match_losses")
        draft_idx = 2
        try:
            user_games = header.index("user_n_games_bucket")
        except:
            user_games = header.index("user_n_matches_bucket")
        try:
            user_wr = header.index("user_game_win_rate_bucket")
        except:
            try:
                user_wr = header.index("user_match_win_rate_bucket")
            except:
                user_wr = None

        if "expansion_pick_number" in header:
            order_idx     = header.index("expansion_pick_number")
            pack_num_idx  = None
            pick_num_idx  = None
        else:
            order_idx     = None
            pack_num_idx  = header.index("pack_number")
            pick_num_idx  = header.index("pick_number")

        maindeck_idx = header.index("pick_maindeck_rate") if "pick_maindeck_rate" in header else None

        pack_names = {i: header[i].replace("pack_card_", "") for i in pack_idx}

        for row in tqdm(reader):
            try:
                draft_id = row[draft_idx]
                positive = row[pick_idx]
            except:
                continue

            step = int(row[order_idx]) if order_idx is not None \
                else int(row[pack_num_idx]) * 15 + int(row[pick_num_idx])

            # pick first, then the rest of the pack
            pack_cards = [positive]
            for i in pack_idx:
                s = row[i]
                if s and s != "0":
                    name = pack_names[i]
                    if name != positive:
                        pack_cards.extend([name] * int(s))

            wins   = int(row[wins_idx])
            losses = int(row[loss_idx])
            u_g    = int(row[user_games])
            u_wr   = float(row[user_wr]) if user_wr and row[user_wr] else 0.0
            in_md  = float(row[maindeck_idx]) if maindeck_idx is not None and row[maindeck_idx] else 0.0
            if positive in ('Plains', 'Island', 'Swamp', 'Mountain', 'Forest'):
                in_md = 0.0

            if draft_id not in drafts:
                drafts[draft_id] = {'steps': {}, 'wins': wins, 'losses': losses,
                                    'u_g': u_g, 'u_wr': u_wr}
            if step not in drafts[draft_id]['steps']:
                drafts[draft_id]['steps'][step] = (pack_cards, in_md)

    for draft_id, data in drafts.items():
        sorted_steps = sorted(data['steps'].items())
        sequence    = [pack for _, (pack, _)  in sorted_steps]
        in_maindeck = [md   for _, (_, md)    in sorted_steps]
        if not sequence:
            continue
        payload = encode_sample((sequence, in_maindeck, data['wins'], data['losses'],
                                 data['u_g'], data['u_wr']))
        (test_q if _is_test(draft_id) else train_q).put((draft_id.encode(), payload))


def all_preprocessing_for_set(set_tag, raw_data_folder, out_folder):
    base = os.path.join(raw_data_folder, set_tag)
    files = [os.path.join(base, f) for f in os.listdir(base) if f.endswith(".csv")]

    out = os.path.join(out_folder, set_tag)
    for name in ('train.lmdb', 'test.lmdb'):
        p = os.path.join(out, name)
        if os.path.exists(p):
            shutil.rmtree(p)
    os.makedirs(out, exist_ok=True)

    train_q = Queue(maxsize=50_000)
    test_q  = Queue(maxsize=50_000)
    train_writer = Process(target=lmdb_writer, args=(os.path.join(out, "train.lmdb"), train_q))
    test_writer  = Process(target=lmdb_writer, args=(os.path.join(out, "test.lmdb"),  test_q))
    train_writer.start(); test_writer.start()

    for csv_path in files:
        print(f"Processing {os.path.basename(csv_path)}")
        stream_csv_trajectories(csv_path, train_q, test_q)

    train_q.put(None); test_q.put(None)
    train_writer.join(); test_writer.join()


if __name__ == "__main__":
    config = utils.load_config('src/configs/config.yaml')
    raw_data_folder = config['raw_data_folder']
    out_folder = config['super_folder']
    os.makedirs(out_folder, exist_ok=True)
    for folder in os.listdir(raw_data_folder):
        if folder not in os.listdir(out_folder) and folder != "cube":
            print(f"Processing set {folder}")
            all_preprocessing_for_set(folder, raw_data_folder, out_folder)