bridge2vec.py

# /// script
# requires-python = ">=3.12"
# dependencies = [
#     "datasets",
#     "endplay",
#     "huggingface-hub",
#     "jinja2",
#     "matplotlib",
#     "numpy",
#     "safetensors",
#     "torch",
#     "umap-learn",
# ]
# ///
"""Bridge2Vec: embed bridge hands by how they take tricks, learned from double-dummy tables."""

# pylint: disable=too-many-arguments,too-many-instance-attributes,too-many-locals
# pylint: disable=too-many-positional-arguments,too-many-statements

import argparse
import json
import math
import os
import random
import sys
import time
from multiprocessing import Pool
from pathlib import Path

import datasets
import huggingface_hub.utils
import matplotlib.pyplot as plt
import numpy as np
import torch
import umap
from datasets import Dataset, load_dataset
from endplay._dds import SetMaxThreads
from endplay.dds import calc_all_tables
from endplay.types import Deal
from huggingface_hub import (
    HfApi,
    ModelCard,
    ModelCardData,
    PyTorchModelHubMixin,
    snapshot_download,
)
from huggingface_hub.errors import RepositoryNotFoundError
from torch import nn
from torch.nn import functional

REPO = "jgalego/bridge2vec"
DATA = "jgalego/bridge2vec-deals"
HERE = Path(__file__).parent
# Progress bars redraw in place, which shows up as garbage in HF Jobs logs.
PROGRESS = sys.stderr.isatty()
RANKS = "23456789TJQKA"
SEATS = "NESW"
# Suits in PBN order, then notrump: the rows of a double-dummy table.
STRAINS = "SHDCN"
# DDS solves at most 32 tables per call; a test hand gets one call's worth of deals.
CHUNK = 32


def cpus():
    """Return the CPU quota visible to this process."""
    try:
        quota, period = Path("/sys/fs/cgroup/cpu.max").read_text(encoding="utf-8").split()
        if quota != "max":
            return max(1, int(quota) // int(period))
    except OSError:
        pass
    return len(os.sched_getaffinity(0))


def hand_pbn(cards):
    """Write card ids (13 * suit + rank) as a PBN hand, e.g. AKQ32.KJ4.T9.A87."""
    suits = [sorted((c % 13 for c in cards if c // 13 == s), reverse=True) for s in range(4)]
    return ".".join("".join(RANKS[r] for r in suit) for suit in suits)


def parse_hand(text):
    """Card ids of a PBN hand."""
    suits = text.split(".")
    if len(suits) != 4:
        raise ValueError(f"{text!r} needs four suits separated by dots")
    cards = [13 * s + RANKS.index(r) for s, ranks in enumerate(suits) for r in ranks.upper()]
    if len(set(cards)) != 13:
        raise ValueError(f"{text!r} is not 13 different cards")
    return cards


def parse_deal(text):
    """Card ids of a PBN deal with North first, shape (4, 13)."""
    hands = [parse_hand(h) for h in text.removeprefix("N:").split()]
    if len(hands) != 4 or len({c for h in hands for c in h}) != 52:
        raise ValueError(f"{text!r} is not four hands of one deck")
    return hands


def tensors(rows):
    """Cards (n, 4, 13) and double-dummy tables (n, 5, 4) of a split."""
    cards = torch.tensor([parse_deal(d) for d in rows["deal"]])
    return cards, torch.tensor(np.array(rows["dd"]))


def profile(cards):
    """HCP and suit lengths of hands, shape (..., 5)."""
    hcp = (cards % 13 - 8).clamp(min=0).sum(-1, keepdim=True)
    return torch.cat([hcp, functional.one_hot(cards // 13, 4).sum(-2)], -1).float()


def permute_suits(cards, dd):
    """Relabel the suits of each deal at random, moving the table rows with them."""
    perm = torch.rand(len(cards), 4, device=cards.device).argsort(1)
    cards = perm.gather(1, (cards // 13).flatten(1)).view_as(cards) * 13 + cards % 13
    rows = perm.argsort(1)[:, :, None].expand(-1, -1, 4)
    return cards, torch.cat([dd[:, :4].gather(1, rows), dd[:, 4:]], 1)


class Bridge2Vec(nn.Module, PyTorchModelHubMixin):
    """Transformer over the 13 cards of a hand; an MLP on four hands predicts the table."""

    def __init__(self, dim=256, depth=4, heads=8, embed_dim=128, hidden=1024):
        super().__init__()
        self.suits = nn.Embedding(4, dim)
        self.ranks = nn.Embedding(13, dim)
        layer = nn.TransformerEncoderLayer(
            dim, heads, 4 * dim, dropout=0.0, batch_first=True, norm_first=True
        )
        self.encoder = nn.TransformerEncoder(layer, depth, enable_nested_tensor=False)
        self.norm = nn.LayerNorm(dim)
        self.project = nn.Sequential(nn.Linear(dim, dim), nn.GELU(), nn.Linear(dim, embed_dim))
        self.table = nn.Sequential(
            nn.Linear(4 * embed_dim, hidden),
            nn.GELU(),
            nn.Linear(hidden, hidden),
            nn.GELU(),
            nn.Linear(hidden, 5 * 14),
        )
        self.value = nn.Linear(embed_dim, 5 * 4)
        nn.init.constant_(self.value.bias, 6.5)
        self.shape = nn.Linear(embed_dim, 5)

    def encode(self, cards):
        """Unit-length embeddings of hands, cards shape (n, 13)."""
        x = self.suits(cards // 13) + self.ranks(cards % 13)
        pooled = self.norm(self.encoder(x)).mean(1)
        return functional.normalize(self.project(pooled), dim=-1)

    def forward(self, cards):
        """Hand embeddings, table logits, expected tables and profiles of deals (n, 4, 13).

        Row d of the table comes from the hands in the order declarer d, LHO, partner, RHO,
        so rotating the seats rotates the table. Expected tables are per hand, with
        declarers relative to it: itself, LHO, partner, RHO.
        """
        n = len(cards)
        z = self.encode(cards.flatten(0, 1)).view(n, 4, -1)
        views = torch.stack([z.roll(-d, 1).flatten(1) for d in range(4)], 1)
        logits = self.table(views).view(n, 4, 5, 14).transpose(1, 2)
        return z, logits, self.value(z).view(n, 4, 5, 4), self.shape(z)

    @torch.no_grad()
    def run(self, cards, batch_size=4096):
        """Hand embeddings, double-dummy tables and expected tables for deals (n, 4, 13)."""
        device = next(self.parameters()).device
        starts = range(0, len(cards), batch_size)
        parts = [self(cards[i : i + batch_size].to(device)) for i in starts]
        z, logits, value, _ = (torch.cat(p).float().cpu() for p in zip(*parts))
        return {"hands": z, "tricks": logits.argmax(-1), "value": value}

    @torch.no_grad()
    def embed(self, hands):
        """Embeddings and expected tables for PBN hands."""
        device = next(self.parameters()).device
        z = self.encode(torch.tensor([parse_hand(h) for h in hands], device=device))
        return z.float().cpu(), self.value(z).view(-1, 5, 4).float().cpu()


def solve(job):
    """32 deals with their double-dummy tables. Test deals share North. Seeded per chunk."""
    split, index, seed = job
    rng = random.Random(f"{seed}:{split}:{index}")
    north = rng.sample(range(52), 13) if split == "test" else []
    deals = []
    for _ in range(CHUNK):
        rest = [c for c in range(52) if c not in north]
        rng.shuffle(rest)
        cards = north + rest
        deals.append("N:" + " ".join(hand_pbn(cards[i : i + 13]) for i in range(0, 52, 13)))
    tables = calc_all_tables([Deal(d) for d in deals])
    return [{"deal": d, "dd": t.to_list()} for d, t in zip(deals, tables)]


def data(args):
    """Deal random hands, solve them double dummy; save as parquet, optionally push."""
    start = time.time()
    jobs = [("test", i) for i in range(args.test_hands)]
    jobs += [("train", i) for i in range(args.deals // CHUNK)]
    jobs = jobs[args.shard :: args.shards]
    rows = {"train": [], "test": []}
    step = max(1, len(jobs) // 20)
    with Pool(cpus(), initializer=SetMaxThreads, initargs=(1,)) as pool:
        tasks = [(split, i, args.seed) for split, i in jobs]
        for n, ((split, _), part) in enumerate(zip(jobs, pool.imap(solve, tasks)), 1):
            rows[split] += part
            if n % step == 0 or n == len(jobs):
                rate = n * CHUNK / (time.time() - start)
                print(json.dumps({"chunks": n, "of": len(jobs), "deals_per_s": round(rate, 1)}),
                      flush=True)
    out = Path(args.output)
    out.mkdir(parents=True, exist_ok=True)
    for split, part in rows.items():
        if not part:
            continue
        name = f"{split}-{args.shard:05d}-of-{args.shards:05d}.parquet"
        Dataset.from_list(part).to_parquet(out / name)
        if args.push:
            HfApi().upload_file(
                path_or_fileobj=out / name,
                path_in_repo=f"data/{name}",
                repo_id=args.repo,
                repo_type="dataset",
                commit_message=f"Add {name}",
            )
    stats = {split: len(part) for split, part in rows.items()}
    print(json.dumps({**stats, "cpus": cpus(), "minutes": round((time.time() - start) / 60, 1)}))


def table(source, split):
    """A split from the Hub or from a local data folder."""
    if Path(source).is_dir():
        files = str(Path(source) / f"{split}-*.parquet")
        return load_dataset("parquet", data_files=files, split="train")
    return load_dataset(source, split=split)


def schedule(step, warmup, total):
    """Linear warmup, then cosine decay to zero."""
    if step < warmup:
        return (step + 1) / warmup
    return 0.5 * (1 + math.cos(math.pi * (step - warmup) / max(1, total - warmup)))


def metric_loss(z, value, temperature, target_temperature):
    """Make hands that play alike close: soft targets from the distance between expected tables."""
    z = z.float().flatten(0, 1)
    value = value.detach().float().flatten(0, 1).flatten(1)
    self_pair = torch.eye(len(z), dtype=torch.bool, device=z.device)
    distance = torch.cdist(value, value, p=1) / value.shape[1]
    target = (-distance / target_temperature).masked_fill(self_pair, -1e9).softmax(1)
    logits = (z @ z.T / temperature).masked_fill(self_pair, -1e9)
    return -(target * logits.log_softmax(1)).sum(1).mean()


def push_result(repo, name, result, revision=None):
    """Upload a result as results/<name>.json in the model repo."""
    HfApi().upload_file(
        path_or_fileobj=json.dumps(result, indent=1).encode(),
        path_in_repo=f"results/{name}.json",
        repo_id=repo,
        revision=revision,
        commit_message=f"Add {name} results",
    )


def train(args):
    """Predict each deal's table from its four hand embeddings, plus per-hand heads."""
    device = "cuda" if torch.cuda.is_available() else "cpu"
    random.seed(args.seed)
    torch.manual_seed(args.seed)
    torch.set_num_threads(cpus())
    cards, dd = (t.to(device) for t in tensors(table(args.data, "train")))
    model = Bridge2Vec(
        dim=args.dim, depth=args.depth, embed_dim=args.embed_dim, hidden=args.hidden
    ).to(device)
    scale = torch.tensor([10.0, 4, 4, 4, 4], device=device)
    optimizer = torch.optim.AdamW(model.parameters(), lr=args.lr, weight_decay=0.05)
    steps = args.max_steps
    lr_schedule = torch.optim.lr_scheduler.LambdaLR(
        optimizer, lambda step: schedule(step, min(args.warmup, steps // 10 + 1), steps)
    )
    parameters = sum(p.numel() for p in model.parameters())
    print(json.dumps({"deals": len(cards), "parameters": parameters}), flush=True)
    start, log = time.time(), {}
    model.train()
    for step in range(steps):
        batch = torch.randint(len(cards), (args.batch_size,), device=device)
        hands, tricks = permute_suits(cards[batch], dd[batch])
        with torch.autocast(device, dtype=torch.bfloat16, enabled=device == "cuda"):
            z, logits, value, shape = model(hands)
        expected = torch.stack([tricks.roll(-d, 2) for d in range(4)], 1).float()
        table_loss = functional.cross_entropy(logits.float().flatten(0, 2), tricks.flatten())
        value_loss = functional.mse_loss(value.float(), expected)
        shape_loss = functional.mse_loss(shape.float(), profile(hands) / scale)
        metric = metric_loss(z, value, args.temperature, args.target_temperature)
        loss = (table_loss + args.value_weight * value_loss + args.aux_weight * shape_loss
                + args.metric_weight * metric)
        optimizer.zero_grad(set_to_none=True)
        loss.backward()
        torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
        optimizer.step()
        lr_schedule.step()
        if step % 100 == 0 or step == steps - 1:
            log = {
                "step": step,
                "loss": round(loss.item(), 4),
                "table": round(table_loss.item(), 4),
                "value": round(value_loss.item(), 4),
                "shape": round(shape_loss.item(), 4),
                "metric": round(metric.item(), 4),
                "exact": round((logits.argmax(-1) == tricks).float().mean().item(), 3),
                "minutes": round((time.time() - start) / 60, 1),
            }
            print(json.dumps(log), flush=True)

    model.eval()
    model.save_pretrained(args.output)
    (Path(args.output) / "README.md").unlink(missing_ok=True)
    if args.push:
        api = HfApi()
        if args.revision:
            api.create_branch(args.repo, branch=args.revision, exist_ok=True)
        api.upload_folder(
            folder_path=args.output,
            repo_id=args.repo,
            revision=args.revision,
            commit_message="Upload model",
        )
        api.upload_file(
            path_or_fileobj=__file__,
            path_in_repo="bridge2vec.py",
            repo_id=args.repo,
            revision=args.revision,
        )
        push_result(
            args.repo,
            "train",
            {
                **log,
                "data": args.data,
                "deals": len(cards),
                "steps": steps,
                "batch_size": args.batch_size,
                "learning_rate": args.lr,
                "value_weight": args.value_weight,
                "aux_weight": args.aux_weight,
                "metric_weight": args.metric_weight,
                "temperature": args.temperature,
                "target_temperature": args.target_temperature,
                "dim": args.dim,
                "depth": args.depth,
                "embed_dim": args.embed_dim,
                "hidden": args.hidden,
                "parameters": parameters,
                "runtime_s": round(time.time() - start),
                "device": torch.cuda.get_device_name() if device == "cuda" else "cpu",
            },
            args.revision,
        )


def nearest(score, groups, chunk=1024):
    """For each row, the best-scoring row of another group; score(rows) gives a block."""
    out = []
    for start in range(0, len(groups), chunk):
        rows = slice(start, start + chunk)
        block = score(rows).float()
        block[groups[rows, None] == groups[None]] = -math.inf
        out.append(block.argmax(1))
    return torch.cat(out)


def retrieval(tables, groups, embedding, features):
    """Mean table distance to the nearest neighbour by embedding, HCP and shape, and chance."""
    methods = {
        "embedding": lambda rows: embedding[rows] @ embedding.T,
        "hcp_shape": lambda rows: (torch.rand(len(groups))[None] * 1e-3
                                   - torch.cdist(features[rows], features, p=1)),
        "random": lambda rows: torch.rand(len(groups[rows]), len(groups)),
    }
    flat = tables.flatten(1).float()
    return {
        name: round((flat[nearest(score, groups)] - flat).abs().mean().item(), 3)
        for name, score in methods.items()
    }


def hard_pairs(embedding, expected, features, margin=0.5):
    """Triplet accuracy among hands the HCP and shape heads cannot tell apart.

    Hands share a bucket when they have the same suit-length pattern and HCP within a band
    of three. For an anchor and two bucket mates whose expected tables differ from the
    anchor's by more than the margin, the embedding must rank the closer table first.
    Chance is 0.5, and so is anything that sees only HCP and shape.
    """
    keys = [(int(f[0]) // 3, *sorted(f[1:].int().tolist())) for f in features]
    flat = expected.flatten(1)
    right = total = 0
    for key in set(keys):
        mates = torch.tensor([i for i, k in enumerate(keys) if k == key])
        if len(mates) < 3:
            continue
        near = torch.cdist(flat[mates], flat[mates], p=1) / flat.shape[1]
        close = 1 - embedding[mates] @ embedding[mates].T
        gap = near[:, :, None] - near[:, None, :]
        same = torch.eye(len(mates), dtype=torch.bool)
        different = (gap.abs() > margin) & ~same[:, :, None] & ~same[:, None, :]
        agree = (close[:, :, None] - close[:, None, :]) * gap > 0
        right += (agree & different).sum().item() / 2
        total += different.sum().item() / 2
    return {"triplets": int(total), "accuracy": round(right / max(total, 1), 3)}


def hand_map(embedding, expected, path):
    """UMAP of the test hands' embeddings, coloured by expected notrump tricks."""
    reducer = umap.UMAP(metric="cosine", n_neighbors=min(15, len(embedding) - 1), random_state=0)
    xy = reducer.fit_transform(embedding.numpy())
    fig, ax = plt.subplots(figsize=(8, 7))
    points = ax.scatter(*xy.T, c=expected, cmap="viridis", s=4, linewidths=0)
    fig.colorbar(points, ax=ax, label="Expected notrump tricks with North declaring", shrink=0.7)
    ax.set_axis_off()
    fig.savefig(path, dpi=150, bbox_inches="tight")
    plt.close(fig)


def evaluate(args):
    """Score predicted tables and expected tables; retrieve hands and deals that play alike."""
    device = "cuda" if torch.cuda.is_available() else "cpu"
    model = Bridge2Vec.from_pretrained(args.model, revision=args.revision).to(device).eval()
    test = table(args.data, "test")
    norths = [d.removeprefix("N:").split()[0] for d in test["deal"]]
    _, groups = np.unique(norths, return_inverse=True)
    if args.limit:
        test = test.select(np.flatnonzero(groups < args.limit))
        groups = groups[groups < args.limit]
    groups = torch.tensor(groups)
    cards, dd = tensors(test)
    out = model.run(cards)
    error = (out["tricks"] - dd).abs()
    first = torch.tensor(np.unique(groups.numpy(), return_index=True)[1])
    hands = len(first)
    expected = torch.zeros(hands, 5, 4).index_add_(0, groups, dd.float())
    expected /= torch.bincount(groups, minlength=hands)[:, None, None]
    value = out["value"][first, 0]
    result = {
        "model": args.model,
        "tables": {
            "deals": len(dd),
            "mae": round(error.float().mean().item(), 3),
            "exact": round((error == 0).float().mean().item(), 3),
            "within_one": round((error <= 1).float().mean().item(), 3),
            "table_exact": round((error == 0).flatten(1).all(1).float().mean().item(), 3),
            "mae_by_strain": {
                s: round(error[:, i].float().mean().item(), 3) for i, s in enumerate(STRAINS)
            },
        },
        "hands": {
            "hands": hands,
            "deals_per_hand": round(len(dd) / hands, 1),
            "expected_mae": round((value - expected).abs().mean().item(), 3),
            "constant_mae": round((expected.mean(0) - expected).abs().mean().item(), 3),
            "retrieval": retrieval(
                expected, torch.arange(hands), out["hands"][first, 0], profile(cards[first, 0])
            ),
            "hard_pairs": hard_pairs(
                out["hands"][first, 0], expected, profile(cards[first, 0])
            ),
        },
        "deal_retrieval": retrieval(
            dd, groups, functional.normalize(out["hands"].flatten(1), dim=-1),
            profile(cards).flatten(1),
        ),
    }
    print(json.dumps(result, indent=1))
    Path(args.output).mkdir(parents=True, exist_ok=True)
    hand_map(out["hands"][first, 0], expected[:, 4, 0], Path(args.output) / "map.png")
    if args.push:
        push_result(args.repo, "eval", result, args.revision)
        HfApi().upload_file(
            path_or_fileobj=Path(args.output) / "map.png",
            path_in_repo="results/map.png",
            repo_id=args.repo,
            revision=args.revision,
        )


def strains(tricks):
    """A table (5, 4) as {strain: {seat: tricks}}."""
    return {s: dict(zip(SEATS, row)) for s, row in zip(STRAINS, tricks.tolist())}


def embed(args):
    """Print a hand's embedding, expected tricks and look-alikes, or a deal's table."""
    model = Bridge2Vec.from_pretrained(args.model).eval()
    if args.deal:
        cards = torch.tensor([parse_deal(args.deal)])
        out = model.run(cards)
        truth = calc_all_tables([Deal("N:" + args.deal.removeprefix("N:"))])[0].to_list()
        result = {
            "predicted": strains(out["tricks"][0]),
            "double_dummy": strains(torch.tensor(truth)),
            "embedding": [round(x, 4) for x in out["hands"][0].flatten().tolist()],
        }
    else:
        z, value = model.embed([args.hand])
        deals = table(args.data, "test")["deal"]
        gallery = sorted({d.removeprefix("N:").split()[0] for d in deals})
        similarity = z @ model.embed(gallery)[0].T
        top = similarity[0].topk(min(args.top, similarity.shape[1]))
        hcp, *lengths = profile(torch.tensor(parse_hand(args.hand))).int().tolist()
        result = {
            "hcp": hcp,
            "lengths": dict(zip(STRAINS, lengths)),
            "expected_tricks": {
                who: {s: round(t, 1) for s, t in zip(STRAINS, value[0, :, j].tolist())}
                for j, who in ((0, "this hand declares"), (2, "partner declares"))
            },
            "nearest": [
                {"hand": gallery[i], "cosine": round(s, 3)}
                for s, i in zip(top.values.tolist(), top.indices.tolist())
            ],
            "embedding": [round(x, 4) for x in z[0].tolist()],
        }
    print(json.dumps(result, indent=1))


def card(args):
    """Render card.jinja into card/README.md with the results stored in the model repo."""
    try:
        folder = Path(snapshot_download(args.repo, allow_patterns="results/*.json"))
        paths = folder.glob("results/*.json")
    except RepositoryNotFoundError:
        paths = []
    results = {path.stem: json.loads(path.read_text(encoding="utf-8")) for path in paths}
    meta = ModelCardData(
        model_name=args.repo.split("/")[1],
        datasets=[DATA],
        license="mit",
        library_name="pytorch",
        pipeline_tag="feature-extraction",
        tags=["contract-bridge", "double-dummy", "embeddings", "weird2vec"],
    )
    rendered = ModelCard.from_template(
        meta,
        template_path=HERE / "card.jinja",
        repo=args.repo,
        data=DATA,
        train=results.get("train"),
        eval=results.get("eval"),
    )
    (HERE / "card").mkdir(exist_ok=True)
    rendered.save(HERE / "card" / "README.md")


def main():
    """Parse arguments and run a command."""
    parser = argparse.ArgumentParser(description=__doc__)
    commands = parser.add_subparsers(dest="command", required=True)

    data_parser = commands.add_parser("data")
    data_parser.add_argument("--deals", type=int, default=200_000, help="train deals")
    data_parser.add_argument("--test-hands", type=int, default=1000, help=f"{CHUNK} deals each")
    data_parser.add_argument("--shard", type=int, default=0)
    data_parser.add_argument("--shards", type=int, default=1)
    data_parser.add_argument("--seed", type=int, default=0)
    data_parser.add_argument("--output", default="out/data")
    data_parser.add_argument("--repo", default=DATA)
    data_parser.add_argument("--push", action="store_true")
    data_parser.set_defaults(run=data)

    train_parser = commands.add_parser("train")
    train_parser.add_argument("--data", default=DATA, help="dataset repo or local data folder")
    train_parser.add_argument("--max-steps", type=int, default=50_000)
    train_parser.add_argument("--batch-size", type=int, default=1024, help="deals per step")
    train_parser.add_argument("--lr", type=float, default=3e-4)
    train_parser.add_argument("--warmup", type=int, default=1000)
    train_parser.add_argument("--value-weight", type=float, default=0.1)
    train_parser.add_argument("--aux-weight", type=float, default=0.1)
    train_parser.add_argument("--metric-weight", type=float, default=0.0,
                              help="weight of the loss that ties embedding to table distance")
    train_parser.add_argument("--temperature", type=float, default=0.1)
    train_parser.add_argument("--target-temperature", type=float, default=0.1)
    train_parser.add_argument("--dim", type=int, default=256)
    train_parser.add_argument("--depth", type=int, default=4)
    train_parser.add_argument("--embed-dim", type=int, default=128)
    train_parser.add_argument("--hidden", type=int, default=1024)
    train_parser.add_argument("--seed", type=int, default=0)
    train_parser.add_argument("--output", default="out/model")
    train_parser.add_argument("--repo", default=REPO)
    train_parser.add_argument("--revision", default=None, help="branch to push to")
    train_parser.add_argument("--push", action="store_true")
    train_parser.set_defaults(run=train)

    eval_parser = commands.add_parser("eval")
    eval_parser.add_argument("--model", default=REPO, help="model repo or local folder")
    eval_parser.add_argument("--data", default=DATA)
    eval_parser.add_argument("--limit", type=int, default=None, help="number of test hands")
    eval_parser.add_argument("--repo", default=REPO, help="where --push stores the result")
    eval_parser.add_argument("--revision", default=None, help="model branch")
    eval_parser.add_argument("--output", default="out/eval", help="where the hand map goes")
    eval_parser.add_argument("--push", action="store_true")
    eval_parser.set_defaults(run=evaluate)

    embed_parser = commands.add_parser("embed")
    target = embed_parser.add_mutually_exclusive_group(required=True)
    target.add_argument("--hand", help="PBN hand, spades first, e.g. AKQ32.KJ4.T9.A87")
    target.add_argument("--deal", help="PBN deal, North first, e.g. N:AKQ32.KJ4.T9.A87 ...")
    embed_parser.add_argument("--model", default=REPO)
    embed_parser.add_argument("--data", default=DATA, help="test hands to search")
    embed_parser.add_argument("--top", type=int, default=5)
    embed_parser.set_defaults(run=embed)

    card_parser = commands.add_parser("card")
    card_parser.add_argument("--repo", default=REPO)
    card_parser.set_defaults(run=card)

    args = parser.parse_args()
    if not PROGRESS:
        datasets.disable_progress_bars()
        huggingface_hub.utils.disable_progress_bars()
    args.run(args)


if __name__ == "__main__":
    main()