multibench-sc 0.2.0__py3-none-any.whl

multibench/__init__.py

@@ -0,0 +1,46 @@
+"""multibench - run and compare single-cell multimodal integration methods.
+
+Quickstart
+----------
+::
+
+    import multibench as mtb
+
+    mtb.list_categories()            # the four integration scenarios
+    print(mtb.describe_layout())     # how to lay out YOUR dataset
+
+    mtb.scan("MYDATA")               # which methods can run on it (and why not the rest)
+
+    res = mtb.run_all("MYDATA", "vertical", out_dir="out/")
+    print(res.summary)               # per-method status + metrics
+    print(res.failures)              # what did not finish, and why
+    res.plot().savefig("compare.png")
+
+The four ``category`` values are ``vertical``, ``diagonal``, ``mosaic`` and
+``cross`` - see :func:`list_categories` for what each one means, and
+:func:`describe_layout` for the filenames each expects.
+"""
+
+from . import config
+from . import plot
+from . import eval
+from .data import catalog
+from .data.results import load_results, available_datasets
+from .eval.pipeline import evaluate, to_long
+from .engine.runner import run
+from .engine.registry import list_methods, list_tasks
+from .engine.resolve import inputs_for, labels_for
+from .engine import ingest as io
+from .engine import envs as env
+from .discover import find_methods, method_info, params_for
+from .workflow import (scan, run_all, BatchResult, list_categories,
+                       describe_layout, load_batch, runtime_hint, sweep)
+
+__version__ = "0.2.0"
+
+__all__ = ["config", "plot", "eval", "catalog", "load_results", "available_datasets",
+           "evaluate", "to_long", "run", "list_methods", "list_tasks", "inputs_for",
+           "labels_for", "io", "env", "method_info", "find_methods", "params_for",
+           "scan", "run_all", "BatchResult",
+           "list_categories", "describe_layout", "load_batch",
+           "runtime_hint", "sweep", "__version__"]

multibench/__main__.py

@@ -0,0 +1,7 @@
+"""Enable ``python -m multibench`` (equivalent to the ``multibench`` console script)."""
+import sys
+
+from .cli import main
+
+if __name__ == "__main__":
+    sys.exit(main())

multibench/cli.py

@@ -0,0 +1,230 @@
+"""Thin CLI over the multibench Python API."""
+from __future__ import annotations
+
+import argparse
+
+from . import discover, load_results
+from . import plot as plot_ns
+
+
+def _cmd_list(args) -> int:
+    from .engine import registry
+    for m in registry.list_methods(category=args.category, task=args.task,
+                                   runnable=args.runnable or None):
+        print(m)
+    return 0
+
+
+def _cmd_find(args) -> int:
+    needs = True if args.needs_labels else None
+    for m in discover.find_methods(category=args.category, task=args.task,
+                                   needs_labels=needs, atac=args.atac,
+                                   runnable=args.runnable or None):
+        print(m)
+    return 0
+
+
+def _cmd_plot(args) -> int:
+    if args.kind != "bubble":
+        raise SystemExit(f"unknown plot kind {args.kind!r}")
+    df = load_results(category=args.category, dataset=args.dataset)
+    metrics = args.metrics.split(",") if args.metrics else None
+    plot_ns.bubble(df, metrics=metrics, aggregate=args.aggregate, save=args.out)
+    print(f"wrote {args.out}")
+    return 0
+
+
+def _parse_inputs(pairs) -> dict:
+    out = {}
+    for p in pairs or []:
+        if "=" not in p:
+            raise SystemExit(f"--input must be role=path, got {p!r}")
+        role, path = p.split("=", 1)
+        if role.strip() == "" or path == "":
+            raise SystemExit(f"--input must be role=path with non-empty parts, got {p!r}")
+        out[role] = path
+    return out
+
+
+def _cmd_run(args) -> int:
+    import multibench
+    res = multibench.run(method=args.method, category=args.category, task=args.task,
+                       inputs=_parse_inputs(args.input), out_dir=args.out,
+                       cmd_template=args.runner)
+    print(f"ran {args.method} -> {res.out_dir}")
+    return 0
+
+
+def _cmd_evaluate(args) -> int:
+    import multibench
+    df = multibench.evaluate(output=args.output, category=args.category, task=args.task,
+                           labels=args.labels, clustering=args.cluster)
+    if args.out:
+        df.to_csv(args.out)
+        print(f"wrote {args.out}")
+    else:
+        print(df.to_string())
+    return 0
+
+
+def _cmd_env(args) -> int:
+    from .engine import envs
+    cmd = args.env_cmd
+    if cmd == "status":
+        for r in envs.status():
+            mark = "x" if r["exists"] else " "
+            tag = r["difficulty"] + ("*" if r["verified_working"] else "")
+            print(f"[{mark}] {r['method']:16} {r['group']:16} {tag}")
+        return 0
+    if cmd == "groups":
+        for name, spec in envs.groups().items():
+            if spec.get("shared"):
+                print(f"{name:16} ({len(spec['members']):2}): {', '.join(spec['members'])}")
+        return 0
+    if cmd == "plan":
+        _mlist = getattr(args, "methods", None)
+        _mlist = [m.strip() for m in _mlist.split(",")] if _mlist else None
+        for p in envs.plan(category=getattr(args, "category", None), methods=_mlist):
+            tag = "shared" if p["shared"] else "own"
+            print(f"{p['env']:16} [{tag:6}] <- {', '.join(p['methods'])}")
+        return 0
+    if cmd == "doctor":
+        _mlist = getattr(args, "methods", None)
+        _mlist = [m.strip() for m in _mlist.split(",")] if _mlist else None
+        rows = envs.doctor(category=getattr(args, "category", None), methods=_mlist)
+        for r in rows:
+            mark = "x" if r["exists"] else ("L" if r["has_lock"] else "!")
+            print(f"[{mark}] {r['env']:18} ({len(r['methods']):2}) <- {', '.join(r['methods'])}")
+        missing = [r for r in rows if not r["exists"]]
+        nolock = [r["env"] for r in missing if not r["has_lock"]]
+        print(f"# {len(rows)} envs needed, {len(missing)} missing"
+              + (f"; NO lockfile for: {', '.join(nolock)}" if nolock else ""))
+        print("# legend: [x]=installed  [L]=missing, lockfile ready (run `multibench env install --run`)  [!]=missing, no lockfile")
+        return 0
+    if cmd == "install":
+        _mlist = getattr(args, "methods", None)
+        _mlist = [m.strip() for m in _mlist.split(",")] if _mlist else None
+        try:
+            rows = envs.create_all(category=getattr(args, "category", None),
+                                   methods=_mlist,
+                                   dry_run=not getattr(args, "run", False))
+        except RuntimeError as e:
+            print(f"error: {e}")
+            return 1
+        for r in rows:
+            state = ("have" if r["exists"]
+                     else ("BUILD" if r["has_lock"] and getattr(args, "run", False)
+                           else "build(dry-run)" if r["has_lock"] else "NO-LOCK"))
+            print(f"{r['env']:18} [{state:14}] <- {', '.join(r['methods'])}")
+        if not getattr(args, "run", False):
+            print("# dry-run — add --run to create the missing envs from their lockfiles")
+        return 0
+    if cmd == "freeze":
+        if getattr(args, "all", False):
+            for env in envs.required_envs(category=getattr(args, "category", None)):
+                try:
+                    print(f"froze {env} -> {envs.freeze(env)}")
+                except Exception as e:  # noqa: BLE001 - report per-env, keep going
+                    print(f"SKIP {env}: {str(e)[:120]}")
+        else:
+            print(f"froze {args.env} -> {envs.freeze(args.env)}")
+        return 0
+    if cmd == "create-group":
+        cmds = envs.create_group(args.group, dry_run=not getattr(args, "run", False))
+        if not getattr(args, "run", False):
+            print("# dry-run — add --run to execute:")
+            for c in cmds:
+                print(" ".join(c))
+        else:
+            print(f"created group env {args.group}")
+        return 0
+    method = getattr(args, "method", None)
+    name = getattr(args, "name", None)
+    if cmd == "recipe":
+        for c in envs.create_commands(method, env_name=name):
+            print(" ".join(c))
+        return 0
+    if cmd == "yml":
+        y = envs.environment_yml(method, env_name=name)
+        out = getattr(args, "out", None)
+        if out:
+            with open(out, "w") as f:
+                f.write(y)
+            print(f"wrote {out}")
+        else:
+            print(y, end="")
+        return 0
+    if cmd == "create":
+        cmds = envs.create(method, env_name=name, dry_run=not getattr(args, "run", False))
+        if not getattr(args, "run", False):
+            print("# dry-run — add --run to execute:")
+            for c in cmds:
+                print(" ".join(c))
+        else:
+            print(f"created environment for {method}")
+        return 0
+    raise SystemExit(
+        "usage: multibench env {status|groups|plan|doctor|install|freeze|recipe|yml|create|create-group}")
+
+
+def build_parser() -> argparse.ArgumentParser:
+    p = argparse.ArgumentParser(prog="multibench")
+    sub = p.add_subparsers(dest="command", required=True)
+
+    pl = sub.add_parser("list"); pl.add_argument("--category"); pl.add_argument("--task")
+    pl.add_argument("--runnable", action="store_true",
+                    help="only methods with a declared variant (usable by run)")
+    pl.set_defaults(func=_cmd_list)
+
+    pf = sub.add_parser("find")
+    pf.add_argument("--category"); pf.add_argument("--task")
+    pf.add_argument("--needs-labels", action="store_true")
+    pf.add_argument("--atac", choices=["peak", "gene_activity"],
+                    help="filter by ATAC representation the method consumes")
+    pf.add_argument("--runnable", action="store_true",
+                    help="only methods with a declared variant (usable by run)")
+    pf.set_defaults(func=_cmd_find)
+
+    pp = sub.add_parser("plot")
+    pp.add_argument("kind")
+    pp.add_argument("--category", required=True); pp.add_argument("--dataset")
+    pp.add_argument("--metrics"); pp.add_argument("--aggregate", default="dataset")
+    pp.add_argument("--out", required=True)
+    pp.set_defaults(func=_cmd_plot)
+
+    pr = sub.add_parser("run")
+    pr.add_argument("--method", required=True); pr.add_argument("--category", required=True)
+    pr.add_argument("--task", default="clustering")
+    pr.add_argument("--input", action="append", help="role=path (repeatable)")
+    pr.add_argument("--out", required=True); pr.add_argument("--runner")
+    pr.set_defaults(func=_cmd_run)
+
+    pe = sub.add_parser("evaluate")
+    pe.add_argument("--output", required=True)
+    pe.add_argument("--category", help="reserved; not used by v1 metrics")
+    pe.add_argument("--task", default="clustering")
+    pe.add_argument("--labels"); pe.add_argument("--cluster"); pe.add_argument("--out")
+    pe.set_defaults(func=_cmd_evaluate)
+
+    pv = sub.add_parser("env", help="per-method environment recipes")
+    ev = pv.add_subparsers(dest="env_cmd", required=True)
+    ev.add_parser("status").set_defaults(func=_cmd_env)
+    ev.add_parser("groups").set_defaults(func=_cmd_env)
+    er = ev.add_parser("recipe"); er.add_argument("method"); er.add_argument("--name"); er.set_defaults(func=_cmd_env)
+    ey = ev.add_parser("yml"); ey.add_argument("method"); ey.add_argument("--name"); ey.add_argument("--out"); ey.set_defaults(func=_cmd_env)
+    ec = ev.add_parser("create"); ec.add_argument("method"); ec.add_argument("--name"); ec.add_argument("--run", action="store_true"); ec.set_defaults(func=_cmd_env)
+    ep = ev.add_parser("plan"); ep.add_argument("--category"); ep.add_argument("--methods", help="comma-separated method names; only their envs"); ep.set_defaults(func=_cmd_env)
+    eg = ev.add_parser("create-group"); eg.add_argument("group"); eg.add_argument("--run", action="store_true"); eg.set_defaults(func=_cmd_env)
+    edoc = ev.add_parser("doctor", help="preflight: which envs are present / need building"); edoc.add_argument("--category"); edoc.add_argument("--methods", help="comma-separated method names; only their envs"); edoc.set_defaults(func=_cmd_env)
+    ei = ev.add_parser("install", help="build every needed env from its lockfile"); ei.add_argument("--category"); ei.add_argument("--methods", help="comma-separated method names; only their envs"); ei.add_argument("--run", action="store_true"); ei.set_defaults(func=_cmd_env)
+    ef = ev.add_parser("freeze", help="capture an env (or --all) to a committed lockfile"); ef.add_argument("env", nargs="?"); ef.add_argument("--all", action="store_true"); ef.add_argument("--category"); ef.set_defaults(func=_cmd_env)
+    return p
+
+
+def main(argv=None) -> int:
+    args = build_parser().parse_args(argv)
+    return args.func(args)
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

multibench/config.py

@@ -0,0 +1,98 @@
+"""Filesystem paths and on-disk token maps for multibench.
+
+The benchmark's result tree uses space-named category folders and a singular
+`scib_metric` top-level dir. Callers use clean tokens (e.g. "vertical"); this
+module translates them to the real folder names.
+"""
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from pathlib import Path
+
+__all__ = ["Config", "DEFAULT", "category_folder", "metric_set_dir"]
+
+# token -> on-disk space-named folder
+_CATEGORY_FOLDERS = {
+    "vertical": "vertical integration",
+    "diagonal": "diagonal integration",
+    "mosaic": "mosaic integration",
+    "cross": "cross integration",
+}
+
+# metric-set token -> top-level result dir. Only "scib" is wired in v1
+# (load_results raises NotImplementedError otherwise); other metric sets are not
+# yet exposed here rather than advertising tokens with no working loader.
+_METRIC_SET_DIRS = {
+    "scib": "scib_metric",
+}
+
+_ROOT = Path(__file__).resolve().parent.parent  # <ROOT>
+# Where the package lives decides where large client-side artefacts go. In a
+# repository checkout (source tree / editable install) they sit next to the
+# package, as always; installed as a wheel, _ROOT lands inside site-packages,
+# which must not accumulate datasets or clones - use a per-user cache dir.
+import os as _os
+_IN_REPO = (_ROOT / "pyproject.toml").is_file()
+_BASE = _ROOT if _IN_REPO else (
+    Path(_os.environ.get("XDG_CACHE_HOME", str(Path.home() / ".cache")))
+    / "multibench")
+
+
+def category_folder(token: str) -> str:
+    """Map a category token to its space-named result folder."""
+    try:
+        return _CATEGORY_FOLDERS[token]
+    except KeyError:
+        raise ValueError(
+            f"unknown category {token!r}; valid: {sorted(_CATEGORY_FOLDERS)}"
+        ) from None
+
+
+def metric_set_dir(token: str) -> str:
+    """Map a metric-set token to its top-level result dir name."""
+    try:
+        return _METRIC_SET_DIRS[token]
+    except KeyError:
+        raise ValueError(
+            f"unknown metric_set {token!r}; valid: {sorted(_METRIC_SET_DIRS)}"
+        ) from None
+
+
+@dataclass
+class Config:
+    """Resolved paths. Override fields to point at custom data locations."""
+
+    result_path: Path = field(default_factory=lambda: _ROOT / "multibench" / "result")
+    files_path: Path = field(default_factory=lambda: _ROOT / "multibench" / "files")
+    repo_path: Path = field(default_factory=lambda: _BASE / "scMultiBench_ref")
+    data_path: Path = field(default_factory=lambda: _BASE / "data")
+
+
+# module-level default instance; callers may replace its fields
+DEFAULT = Config()
+
+
+def ensure_repo(path=None):
+    """Return a directory that contains ``tools_scripts/``, provisioning it if needed.
+
+    Resolution order: the given (or configured) ``repo_path``; the package root
+    itself (the merged-repository layout, where ``tools_scripts/`` sits next to
+    ``multibench/``); otherwise a one-time shallow clone of the public
+    scMultiBench repository into the configured location - which is what makes
+    method execution work on a fresh machine or Colab, where the wrapper's
+    clone does not carry the 3 GB of upstream method scripts.
+    """
+    import subprocess
+    from pathlib import Path as _P
+
+    p = _P(path) if path else DEFAULT.repo_path
+    if (p / "tools_scripts").is_dir():
+        return p
+    if (_ROOT / "tools_scripts").is_dir():
+        return _ROOT
+    print(f"method scripts not found - fetching PYangLab/scMultiBench (once) into {p} ...",
+          flush=True)
+    subprocess.run(["git", "clone", "--depth", "1",
+                    "https://github.com/PYangLab/scMultiBench.git", str(p)],
+                   check=True)
+    return p

multibench/data/__init__.py

@@ -0,0 +1,2 @@
+
+from .fetch import fetch  # noqa: F401

multibench/data/catalog.py

@@ -0,0 +1,132 @@
+"""Catalog: parse files/*.csv into typed tables with canonical names."""
+from __future__ import annotations
+
+import re
+from pathlib import Path
+
+import pandas as pd
+
+from .. import config
+
+__all__ = ["methods", "datasets", "metrics", "canonical_id", "canonical_metric"]
+
+# --- canonical method id + aliases -----------------------------------------
+# Canonical id = the registry token. Map known display / result-dir spellings.
+_ALIASES = {
+    "seurat v3": "Seurat_v3",
+    "seurat_v3": "Seurat_v3",
+    "seurat.v3": "Seurat_v3",
+    "seurat v4": "Seurat_v4",
+    "seurat_v4": "Seurat_v4",
+    "seurat.v4": "Seurat_v4",
+    "seurat v5": "Seurat_v5",
+    "seurat_v5": "Seurat_v5",
+    "seurat.v5": "Seurat_v5",
+    "seurat(wnn)": "Seurat_WNN",
+    "seurat_wnn": "Seurat_WNN",
+    "seurat.wnn": "Seurat_WNN",
+    "mofa+": "MOFA2",
+    "mofa2": "MOFA2",
+    "online inmf": "online_iNMF",
+    "online_inmf": "online_iNMF",
+    "online.inmf": "online_iNMF",
+    "ipolng": "iPOLNG",
+}
+
+
+def canonical_id(name: str) -> str:
+    """Return the canonical method id for any known spelling."""
+    key = name.strip().lower()
+    if key in _ALIASES:
+        return _ALIASES[key]
+    # default: collapse separators to underscore, keep original casing token
+    return re.sub(r"[ .]+", "_", name.strip())
+
+
+# --- metric code canonicalization ------------------------------------------
+_METRIC_CANON = {
+    "kbet": "kBET",
+    "ifi": "iF1",
+    "if1": "iF1",
+    "ari": "ARI",
+    "nmi": "NMI",
+    "asw": "ASW",
+    "iasw": "iASW",
+    "clisi": "cLISI",
+    "ilisi": "iLISI",
+    "gc": "GC",
+    "asw_batch": "ASW_batch",
+    "pcr": "PCR",
+    # raw scIB long-names -> canonical codes (some metric.csv use these)
+    "ari_cluster/label": "ARI",
+    "nmi_cluster/label": "NMI",
+    "asw_label": "ASW",
+    "isolated_label_f1": "iF1",
+    "isolated_label_silhouette": "iASW",
+}
+
+
+def canonical_metric(code: str) -> str | None:
+    """Canonicalize a raw metric short-code; None for blank/unknown-empty."""
+    if code is None:
+        return None
+    key = str(code).strip().lower()
+    if key == "" or key == "nan":
+        return None
+    return _METRIC_CANON.get(key, str(code).strip())
+
+
+# --- tables ----------------------------------------------------------------
+def _split_multivalue(cell: object) -> list[str]:
+    if pd.isna(cell):
+        return []
+    parts = re.split(r"[\n;,]+", str(cell))
+    out = []
+    for p in parts:
+        t = p.strip().lower().replace(" integration", "").replace("integration", "").strip()
+        if t:
+            out.append(t)
+    return out
+
+
+def methods(files_dir: Path | str | None = None) -> pd.DataFrame:
+    """Return the methods table with normalized columns and list-valued cats/tasks."""
+    if files_dir is None:
+        files_dir = config.DEFAULT.files_path
+    raw = pd.read_csv(Path(files_dir) / "method.csv")
+    # normalize column access by stripping whitespace/newlines
+    cols = {c: c.strip().replace("\n", " ").strip() for c in raw.columns}
+    raw = raw.rename(columns=cols)
+    out = pd.DataFrame()
+    out["method"] = raw["Methods"].astype(str).str.strip()
+    out["canonical_id"] = out["method"].map(canonical_id)
+    out["language"] = raw["Programming Language"].astype(str).str.strip().str.lower()
+    out["deep_learning"] = raw["Deep Learning"].astype(str).str.strip()
+    out["atac"] = raw["Peak/Gene Activity"].astype(str).str.strip()
+    out["output"] = raw["Output"].astype(str).str.strip()
+    out["needs_labels"] = (
+        raw["CellType Information Required"].astype(str).str.strip().str.lower().isin(["yes", "y", "true"])
+    )
+    out["categories"] = raw["Integration Categories"].map(_split_multivalue)
+    out["tasks"] = raw["Task Categories"].map(_split_multivalue)
+    return out
+
+
+def datasets(files_dir: Path | str | None = None) -> pd.DataFrame:
+    """Return the datasets table with a derived `simulated` flag."""
+    if files_dir is None:
+        files_dir = config.DEFAULT.files_path
+    raw = pd.read_csv(Path(files_dir) / "dataset.csv")
+    raw = raw.rename(columns={c: c.strip() for c in raw.columns})
+    raw = raw.dropna(how="all", axis=1).dropna(how="all", axis=0)
+    name_col = "dataset name"
+    raw["simulated"] = raw[name_col].astype(str).str.strip().str.upper().str.startswith("SD")
+    return raw
+
+
+def metrics(files_dir: Path | str | None = None) -> pd.DataFrame:
+    """Return the metric details table."""
+    if files_dir is None:
+        files_dir = config.DEFAULT.files_path
+    raw = pd.read_csv(Path(files_dir) / "metric_full.csv")
+    return raw.rename(columns={c: c.strip() for c in raw.columns})

multibench/data/fetch.py

@@ -0,0 +1,40 @@
+"""Download reference datasets from the repository's release assets."""
+from __future__ import annotations
+
+import tarfile
+import urllib.request
+from pathlib import Path
+
+from .. import config
+
+RELEASE_URL = "https://github.com/DSichang/scMultiBench/releases/download/data-v1"
+
+# datasets published as release assets, with approximate download sizes
+AVAILABLE = {"D11": "11 MB", "D28": "137 MB", "D45": "290 MB",
+             "D46": "97 MB", "D52": "179 MB"}
+
+
+def fetch(*datasets: str, data_path=None, quiet: bool = False) -> Path:
+    """Ensure the named reference datasets exist locally, downloading if needed.
+
+    Idempotent: datasets already present under the data root are left alone.
+    Returns the data root. The full 65-dataset collection is linked from the
+    scMultiBench README; this helper covers the tutorial reference sets.
+    """
+    root = Path(data_path) if data_path is not None else config.DEFAULT.data_path
+    for ds in datasets:
+        if (root / ds).is_dir():
+            continue
+        if ds not in AVAILABLE:
+            raise ValueError(
+                f"{ds!r} is not in the release assets ({', '.join(sorted(AVAILABLE))}); "
+                "see 'Get the data' in the installation guide for the full collection")
+        root.mkdir(parents=True, exist_ok=True)
+        if not quiet:
+            print(f"downloading {ds} ({AVAILABLE[ds]}) ...", flush=True)
+        tgz, _ = urllib.request.urlretrieve(f"{RELEASE_URL}/{ds}.tar.gz")
+        with tarfile.open(tgz) as t:
+            t.extractall(root)
+        if not (root / ds).is_dir():
+            raise RuntimeError(f"downloaded archive did not contain {ds}/")
+    return root