leaderboard / scripts /import_benchmark_data.py
dbonet's picture paumateo's picture
Publish TabBioMed leaderboard
d189f56
Raw History Blame Contribute Delete
31.3 kB
#!/usr/bin/env python3
"""Build the static leaderboard payload using TabBioMed's paper analysis package."""
import argparse
from collections import Counter
from concurrent.futures import ProcessPoolExecutor
from datetime import datetime, timezone
import hashlib
import json
from pathlib import Path
import numpy as np
import pandas as pd
from scipy.stats import rankdata
from tabbiomed.analysis.common import PRIMARY_METRIC, load_benchmark_models_config
from tabbiomed.analysis.datasets import load_benchmark, load_arrays
from tabbiomed.analysis.elo.missing import apply_missing_policy
from tabbiomed.analysis.elo.splitwise import rate
from tabbiomed.analysis.figures.elo_domains import (
MIN_BOOTSTRAP_GROUPS,
cluster_rate,
summarize as summarize_domains,
)
from tabbiomed.analysis.figures.pareto import collect_timings, summarize_costs
from tabbiomed.analysis.figures.source_dependence import load_groups, partition
from tabbiomed.analysis.figures.win_rates import pairwise_matrix
from tabbiomed.analysis.paths import CONFIG, PACKAGE, RESULTS, PROCESSED
from tabbiomed.analysis.pretraining_overlap import pretraining_overlaps
from tabbiomed.analysis.ranks import normalized_scores_with_failures
CLASSIFICATION = {"binary", "classification"}
ATLAS_SOURCE_LABELS = {
"bioresponse": "Kaggle Predicting a Biological Response via OpenML 4134",
"Cardiovascular-Disease-dataset": "Kaggle via TALENT / OpenML 45547",
"CDC_Diabetes_Health_Indicators": "CDC BRFSS 2015, Alex Teboul’s Kaggle release via TALENT",
"vep_pathogenic_coding": (
"InstaDeepAI Genomics Long-Range Benchmark, "
"Coding Variant Pathogenicity Dataset (June 17, 2024 release)"
),
}
ATLAS_SOURCE_URLS = {
"vep_pathogenic_coding": (
"https://huggingface.co/datasets/InstaDeepAI/genomics-long-range-benchmark"
"/blob/84cfc69322404818d28fb80921b54d8edc39d39e"
"/variant_effect_pathogenic/vep_pathogenic_coding.csv"
),
}
REPO = Path(__file__).resolve().parents[1]
def finite(value):
return float(value) if pd.notna(value) and np.isfinite(value) else None
def ratings_table(frame, scoped=False):
if scoped:
frame = frame.rename(
columns={
"model": "competitor",
"elo_calibrated": "elo",
"elo_cal_lo_95": "elo_lo_95",
"elo_cal_hi_95": "elo_hi_95",
}
)
table = {}
for row in frame.itertuples():
table[row.competitor] = (
float(row.elo),
finite(row.elo_lo_95),
finite(row.elo_hi_95),
)
if len(table) != len(frame):
raise ValueError("Duplicate competitor in Elo table")
return table
def match_table(frame, saved, reference, count, observation_groups):
if saved is None:
if observation_groups is not None:
bootstrap = 200 if len(set(observation_groups.values())) >= MIN_BOOTSTRAP_GROUPS else 0
fitted = cluster_rate(frame, observation_groups, bootstrap=bootstrap, seed=42)
return ratings_table(fitted, scoped=True)
bootstrap = 200 if count >= 5 else 0
fitted, _, _ = rate(frame, bootstrap=bootstrap, seed=42, reference=reference)
if not bootstrap:
fitted[["elo_lo_95", "elo_hi_95"]] = np.nan
return ratings_table(fitted)
if set(frame.competitor) != set(saved):
raise ValueError(f"Result roster differs from paper summary for {reference}")
return saved
def mode_result(raw, timings, models, selected, saved, observation_groups=None):
reference = "RandomForest" if selected else "RandomForest::default"
frame, _ = apply_missing_policy(raw, "rf_imputed", reference=reference)
if frame.empty:
return None
if selected and frame.imputed.any():
# Selected Random Forest anchors Elo, but only its Default score fills gaps.
defaults = raw[raw.model.eq("RandomForest")].set_index("dataset").default_metric
imputed = frame.loc[frame.imputed, "dataset"].map(defaults)
if not np.isfinite(imputed).all():
raise ValueError("Random Forest Default is missing for an imputed task")
frame.loc[frame.imputed, "metric_value"] = imputed
count = frame.dataset.nunique()
elo = match_table(frame, saved, reference, count, observation_groups)
scores = raw[raw.dataset.isin(frame.dataset) & raw.competitor.isin(elo)]
tasks = scores.groupby("dataset").task.first()
matrix = scores.pivot(index="dataset", columns="competitor", values="metric_value")
normalized = normalized_scores_with_failures(matrix, tasks)
ranks = pd.DataFrame(
rankdata(-normalized, axis=1), index=matrix.index, columns=matrix.columns
)
oriented = matrix.mul(np.where(tasks.eq("binary"), 1, -1), axis=0)
values = oriented.to_numpy()
# A completed result beats a failure. Two failures count as a tie.
left = np.where(np.isnan(values), -np.inf, values)[:, :, None]
right = left.transpose(0, 2, 1)
wins = (left > right) + 0.5 * (left == right)
mean_win = pd.Series(
100 * (wins.sum(axis=(0, 2)) - count / 2) / (count * (len(elo) - 1)),
index=matrix.columns,
)
timed = frame.merge(
timings,
on=["model", "dataset", "timing_variant"],
how="left",
validate="many_to_one",
)
timed["timing_imputed"] = timed.imputed
# In the paper's Pareto tables, imputed scores use the corresponding RF timing.
rf_times = timings[
timings.model.eq("RandomForest") & timings.timing_variant.eq("default")
].set_index("dataset")
fields = ["train_per_1k", "inference_per_1k", "final_fit_per_1k"]
for field in fields:
timed.loc[timed.imputed, field] = timed.loc[timed.imputed, "dataset"].map(
rf_times[field]
)
elo_frame = pd.DataFrame(
[{"competitor": key, "elo": val[0]} for key, val in elo.items()]
)
train = summarize_costs(
timed, elo_frame, "train_per_1k", "x_median_train_per_1k"
).set_index("competitor")
inference = summarize_costs(
timed, elo_frame, "inference_per_1k", "x_median_inference_per_1k"
).set_index("competitor")
entries = []
ordered = sorted(elo, key=lambda key: (-elo[key][0], key.casefold(), key))
for index, key in enumerate(ordered, 1):
model, _, variant = key.partition("::")
variant = variant or "selected"
score, low, high = elo[key]
stats = timed[timed.competitor.eq(key)]
n_completed = int(matrix[key].notna().sum())
paired = not selected and all(
f"{model}::{v}" in elo for v in ("default", "tuned")
)
entries.append(
dict(
key=key,
modelId=model,
variant=variant,
family=models[model]["family"],
hpoEnabled=models[model]["hpoEnabled"],
classificationOnly=models[model]["classificationOnly"],
paretoColor=models[model]["paretoColor"],
rank=index,
elo=score,
displayName=model if selected else f"{model} ({variant.title()})",
eloLo=low,
eloHi=high,
tunedMinusDefaultElo=(
elo[f"{model}::tuned"][0] - elo[f"{model}::default"][0]
)
if paired
else None,
meanWinRate=float(mean_win[key]),
normScoreMean=float(normalized[key].mean()),
normScoreMedian=float(normalized[key].median()),
meanRank=float(ranks[key].mean()),
rankMin=float(ranks[key].min()),
rankMax=float(ranks[key].max()),
improvabilityPct=float(100 * (1 - normalized[key].mean())),
nCompleted=n_completed,
nImputed=int(stats.imputed.sum()),
nDefaultFallback=int(stats.metric_source.eq("default_fallback").sum()),
nDatasets=count,
completionPct=100 * n_completed / count,
medianTrainPer1k=finite(train.at[key, "x_median_train_per_1k"])
if key in train.index
else None,
medianInferPer1k=finite(inference.at[key, "x_median_inference_per_1k"])
if key in inference.index
else None,
medianFinalFitPer1k=finite(stats.final_fit_per_1k.median()),
isParetoTrain=bool(train.at[key, "is_pareto"])
if key in train.index
else False,
isParetoInfer=bool(inference.at[key, "is_pareto"])
if key in inference.index
else False,
)
)
result = {
"nDatasets": count,
"datasetIds": sorted(frame.dataset.unique()),
"entries": entries,
}
if selected:
participants = sorted(elo, key=str.casefold)
dataset_meta = scores[["dataset", "task"]].drop_duplicates()
win_matrix, _, pairs, _ = pairwise_matrix(scores, dataset_meta, participants)
result["winRate"] = {
"models": participants,
"matrix": [
[finite(value) for value in row] for row in win_matrix.to_numpy()
],
"counts": {
f"{r.model_a}__{r.model_b}": [
r.wins,
r.losses,
r.ties,
r.wins_on_score,
r.wins_by_completion,
]
for r in pairs.itertuples()
},
}
return result
def scope_result(job):
key, label, datasets, raw, paired, timings, models, saved = job
names = {d["id"] for d in datasets}
classification_only = (
key == "global_classification"
or key.startswith(("domain_cls:", "quadrant_cls:"))
or key in {"task:binary", "task:classification"}
)
roster = {
name: model
for name, model in models.items()
if classification_only or not model["classificationOnly"]
}
raw = raw[raw.dataset.isin(names) & raw.model.isin(roster)]
paired = paired[paired.dataset.isin(names) & paired.model.isin(roster)]
if raw.shape[0] != len(datasets) * len(roster):
raise ValueError(f"Incomplete selected-result grid for {key}")
block = dict(
key=key,
label=label,
nDatasets=len(datasets),
nModels=len(roster),
datasetIds=sorted(names),
samplesMin=min(d["numSamples"] for d in datasets),
samplesMax=max(d["numSamples"] for d in datasets),
featuresMin=min(d["numFeatures"] for d in datasets),
featuresMax=max(d["numFeatures"] for d in datasets),
)
for view, results in (("selected", raw), ("paired", paired)):
block[view] = mode_result(
results, timings, roster, view == "selected", saved.get(view)
)
retained = names - set(pretraining_overlaps()["all"])
block["nDatasetsNoOverlap"] = len(retained)
if retained and retained != names:
cluster_scope = (
key.startswith("panel:")
or (key.startswith("domain:") and key.count(":") == 1)
)
observation_groups = (
{d["id"]: d["observationGroup"] for d in datasets if d["id"] in retained}
if cluster_scope else None
)
for view, results in (("selected", raw), ("paired", paired)):
block[f"{view}NoOverlap"] = mode_result(
results[results.dataset.isin(retained)],
timings,
roster,
view == "selected",
None,
observation_groups=observation_groups if view == "selected" else None,
)
return key, block
def scopes_for(datasets, domains, classification_quadrants):
def entries(key, label, predicate):
return key, label, [d for d in datasets if predicate(d)]
is_classification = lambda d: d["task"] in CLASSIFICATION
has_high_missingness = lambda d: d["missingCellFraction"] >= 0.01
yield entries(
"global_all_tasks", "All Biomedical Tasks (21 Models)", lambda d: True
)
yield entries(
"global_classification",
"Classification Tasks (All 23 Models)",
is_classification,
)
for domain in domains:
slug, label = domain["slug"], domain["fullName"]
yield entries(f"domain:{slug}", label, lambda d: d["domainSlug"] == slug)
yield entries(
f"domain_cls:{slug}",
f"{label} (Classification)",
lambda d: d["domainSlug"] == slug and is_classification(d),
)
for flag, name in ((True, "with_missing"), (False, "no_missing")):
yield entries(
f"domain:{slug}:{name}",
f"{label} · {name.replace('_', ' ')} values",
lambda d: d["domainSlug"] == slug and d["hasMissing"] == flag,
)
yield entries(
f"domain:{slug}:high_missing",
f"{label} · At Least 1% Missing Cells",
lambda d: d["domainSlug"] == slug and has_high_missingness(d),
)
for suffix, label in (("up_to_2000", "≤ 2,000"), ("over_2000", "> 2,000")):
key = f"ehr_{suffix}"
yield entries(
f"panel:{key}",
f"Clinical & Healthcare ({label} Samples)",
lambda d: d["clinicalSubpanel"] == key,
)
for key, label, predicate in (
(
"missing:with_missing",
"Datasets with Missing Values",
lambda d: d["hasMissing"],
),
(
"missing:no_missing",
"Complete Datasets (0% Missing)",
lambda d: not d["hasMissing"],
),
(
"missing:high_missing",
"At Least 1% Missing Cells",
has_high_missingness,
),
(
"wide:over_2000",
"Wide Datasets (> 2,000 Features)",
lambda d: d["numFeatures"] > 2000,
),
(
"cat:has_categorical",
"Datasets with Categorical Features",
lambda d: d["numCategorical"] > 0,
),
(
"cat:numeric_only",
"Purely Numerical Features",
lambda d: d["numCategorical"] == 0,
),
):
yield entries(key, label, predicate)
for task, label in (
("binary", "Binary Classification"),
("classification", "Multiclass Classification"),
("regression", "Single-Target Regression"),
("multiregression", "Multi-Output Regression"),
):
yield entries(f"task:{task}", label, lambda d: d["task"] == task)
for quad in ("Q1", "Q2", "Q3", "Q4"):
label = next(d["quadrantLabel"] for d in datasets if d["quadrant"] == quad)
yield entries(f"quadrant:{quad}", label, lambda d: d["quadrant"] == quad)
yield entries(
f"quadrant_cls:{quad}",
f"{label} (Classification)",
lambda d: classification_quadrants.get(d["id"]) == quad,
)
def dataset_metadata(read):
domains = read("elo_domains/datasets.csv")[
["dataset", "domain", "domain_slug", "observation_group"]
]
regimes = read("elo_regimes/all_tasks_datasets.csv")[
["dataset", "task", "num_samples", "num_features", "quadrant"]
]
return domains.merge(regimes, on="dataset", validate="one_to_one")
def missingness(x):
seen = np.zeros(x.shape[1], dtype=bool)
missing_cells = 0
step = max(1, 8_000_000 // x.shape[1])
for start in range(0, len(x), step):
chunk = x[start : start + step]
mask = (
np.isnan(chunk)
if np.issubdtype(x.dtype, np.number)
else (pd.isna(chunk) | (chunk == ""))
)
seen |= mask.any(axis=0)
missing_cells += int(mask.sum())
return missing_cells / x.size, int(seen.sum())
def dataset_records(metadata, catalog, known_missingness):
_, source_groups, _ = load_groups(set(metadata.dataset))
families = partition(source_groups, set(metadata.dataset))
records = []
for row in metadata.sort_values("dataset").itertuples():
md = catalog[row.dataset]["metadata"]
source_url = ATLAS_SOURCE_URLS.get(row.dataset)
if (row.num_samples, row.num_features, row.task) != (
md["num_samples"],
md["num_features"],
md["task"],
):
raise ValueError(
f"Dataset metadata does not match the release: {row.dataset}"
)
missing_fraction, missing_features = 0.0, 0
expected_missing = known_missingness.get(row.dataset)
if expected_missing is None or expected_missing:
x, _ = load_arrays(row.dataset, PROCESSED)
missing_fraction, missing_features = missingness(x)
if expected_missing is not None and not np.isclose(
expected_missing, missing_fraction
):
raise ValueError(
f"Missingness does not match released data: {row.dataset}"
)
num_categorical = len(md.get("cat_idx", []))
split_strategy = "cv" if md.get("split_strategy") == "cv" else "fixed"
records.append(
dict(
id=row.dataset,
domainSlug=row.domain_slug,
domain=row.domain,
clinicalSubpanel=(
"ehr_up_to_2000" if row.num_samples <= 2000 else "ehr_over_2000"
)
if row.domain_slug == "ehr"
else None,
observationGroup=row.observation_group,
sourceFamily=families[row.dataset],
task=row.task,
primaryMetric=PRIMARY_METRIC[row.task],
higherIsBetter=row.task == "binary",
numSamples=row.num_samples,
numFeatures=row.num_features,
numCategorical=num_categorical,
catFraction=num_categorical / row.num_features,
numClasses=md.get("num_classes"),
imbalanceRatio=md.get("imbalance_ratio"),
nFolds=5 if split_strategy == "cv" else 1,
splitStrategy=split_strategy,
quadrant=row.quadrant.split()[0],
quadrantLabel=row.quadrant,
missingCellFraction=missing_fraction,
missingFeatureCount=missing_features,
hasMissing=bool(missing_features),
hasZeroCodedMissingNote=row.dataset == "pima_diabetes",
description=md.get("description", ""),
sourceShort=ATLAS_SOURCE_LABELS.get(
row.dataset, md.get("source", row.domain)
),
license=md.get("license"),
licenseUrl=md.get("license_url"),
**({"sourceUrl": source_url} if source_url else {}),
)
)
return records
def saved_ratings(read):
by_scope = {
"global_all_tasks": {
"paired": ratings_table(read("elo/rf_imputed/elo_all_entries.csv")),
},
}
paper_domains = read("elo_domains/summary.csv")
try:
clinical = read("elo_domains/clinical_all_summary.csv")
except FileNotFoundError:
metadata = read("elo_domains/datasets.csv")
clinical_datasets = metadata[metadata.domain_slug.eq("ehr")].assign(panel="ehr")
inputs = read("elo_domains/inputs.csv")
clinical, _ = summarize_domains(
inputs,
clinical_datasets,
sorted(inputs.model.unique()),
bootstrap=200,
seed=0,
)
domain = pd.concat([paper_domains, clinical])
for name, group in domain.groupby("panel"):
prefix = "panel" if name.startswith("ehr_") else "domain"
by_scope[f"{prefix}:{name}"] = {"selected": ratings_table(group, scoped=True)}
for case, prefix in (("all_tasks", "quadrant"), ("classification", "quadrant_cls")):
for name, group in read(f"elo_regimes/{case}_summary.csv").groupby("quadrant"):
by_scope[f"{prefix}:{name.split()[0]}"] = {
"selected": ratings_table(group, scoped=True)
}
for name in ("with_missing", "no_missing"):
for mode in ("paired", "selected"):
try:
frame = read(f"elo_missingness/{name}_{mode}.csv")
except FileNotFoundError:
continue
by_scope.setdefault(f"missing:{name}", {})[mode] = ratings_table(frame)
return by_scope
def metric_inputs(raw):
paired = pd.concat(
[
raw.assign(variant="default", metric_value=raw.default_metric),
raw[raw.hpo_enabled].assign(
variant="tuned", metric_value=raw.loc[raw.hpo_enabled, "tuned_metric"]
),
],
ignore_index=True,
)
paired["competitor"] = paired.model + "::" + paired.variant
paired["timing_variant"] = paired.variant
paired["metric_source"] = "variant"
paired["fold"] = 0
return paired
def timing_inputs(paired, raw, plan):
imputed, _ = apply_missing_policy(paired, "rf_imputed")
selections = raw[["model", "dataset", "report_sha256"]].to_dict("records")
frame, folds = collect_timings(imputed, selections, plan)
fitted = folds.assign(
final_fit_per_1k=1000 * folds.final_fit_seconds / folds.n_fitted
)
final_fit = (
fitted.groupby(["model", "dataset", "variant"])
.final_fit_per_1k.agg(
lambda values: values.mean() if np.isfinite(values).all() else np.nan
)
.reset_index()
)
fields = ["model", "dataset", "variant", "train_per_1k", "inference_per_1k"]
return (
frame.loc[~frame.imputed, fields]
.merge(final_fit, on=fields[:3])
.rename(columns={"variant": "timing_variant"})
)
def model_records(config, plan, raw, failures):
colors = json.loads((CONFIG / "pareto_colors.json").read_text())
models = {}
for name in sorted(plan["models"], key=str.casefold):
rows = raw[raw.model.eq(name)]
failed = failures[failures.model.eq(name) & ~failures.completed]
pairs = rows.dropna(subset=["default_metric", "tuned_metric"])
better = np.where(pairs.task.eq("binary"), 1, -1) * (
pairs.tuned_metric - pairs.default_metric
)
models[name] = dict(
id=name,
family=config.model_family[name],
hpoEnabled=plan["models"][name]["hpo_enabled"],
classificationOnly=name in config.regression_optional_missing_models,
paretoColor=colors.get(name, "#64748b"),
failureSummary=dict(
defaultFailed=int(failed.variant.eq("default").sum()),
tunedFailed=int(failed.variant.eq("tuned").sum()),
usable=int(rows.metric_value.notna().sum()),
totalApplicable=len(rows),
causes=dict(Counter(failed.cause.fillna("unknown"))),
),
hpoComparison=dict(
pairedCount=len(pairs),
tunedBetter=int((better > 0).sum()),
tie=int((better == 0).sum()),
defaultBetter=int((better < 0).sum()),
),
)
return models
def evaluation_records(raw, timings, failures, models, datasets):
clocks = timings.set_index(["model", "dataset", "timing_variant"])
causes = (
failures[~failures.completed].set_index(["model", "dataset", "variant"]).cause
)
model_indices = {name: i for i, name in enumerate(models)}
dataset_indices = {d["id"]: i for i, d in enumerate(datasets)}
records = []
for row in raw.itertuples():
def clock(variant, field):
key = (row.model, row.dataset, variant)
return finite(clocks.at[key, field]) if key in clocks.index else None
fail = [causes.get((row.model, row.dataset, v)) for v in ("default", "tuned")]
fail = {c for c in fail if pd.notna(c)}
records.append(
dict(
m=model_indices[row.model],
d=dataset_indices[row.dataset],
defScore=finite(row.default_metric),
tunScore=finite(row.tuned_metric),
selScore=finite(row.metric_value),
selVar=row.selected_variant,
src=row.metric_source,
defTrain1k=clock("default", "train_per_1k"),
defInfer1k=clock("default", "inference_per_1k"),
tunTrain1k=clock("tuned", "train_per_1k"),
tunInfer1k=clock("tuned", "inference_per_1k"),
selFit1k=clock(row.selected_variant, "final_fit_per_1k"),
failCause=("mixed" if len(fail) > 1 else next(iter(fail), None)),
)
)
return records
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument(
"--summaries",
type=Path,
required=True,
help="Generated paper CSV output directory",
)
parser.add_argument(
"--output", type=Path, default=REPO / "public/data/benchmarkData.json"
)
parser.add_argument("--workers", type=int, default=4)
args = parser.parse_args()
def read(name):
return pd.read_csv(args.summaries / name, float_precision="round_trip")
plan = load_benchmark()
config = load_benchmark_models_config()
catalog = json.loads((PACKAGE / "configs/datasets.json").read_text())["datasets"]
metadata = dataset_metadata(read)
try:
known_missingness = (
read("datasets_metadata.csv")
.set_index("dataset")
.missing_cell_fraction.to_dict()
)
except FileNotFoundError:
known_missingness = {}
if metadata.dataset.duplicated().any() or set(metadata.dataset) != set(
plan["datasets"]
):
raise ValueError("Dataset summary scope does not match the paper plan")
raw_main = read("model_summary/selected_results.csv").drop(
columns=["train_per_1k", "inference_per_1k"]
)
raw_extra = read("elo_regimes/classification_inputs.csv")
raw_extra = raw_extra[
raw_extra.model.isin(config.regression_optional_missing_models)
].copy()
raw_extra["metric_value"] = raw_extra.raw_metric
raw_extra.loc[raw_extra.imputed, "metric_source"] = "failed"
raw = pd.concat(
[raw_main, raw_extra.reindex(columns=raw_main.columns)], ignore_index=True
)
raw["variant"] = "selected"
raw["competitor"] = raw.model
raw["fold"] = 0
raw["timing_variant"] = raw.selected_variant
paired = metric_inputs(raw)
failures = read("failures/variant_completion.csv")
models = model_records(config, plan, raw, failures)
print("Validating consolidated reports and timings", flush=True)
timings = timing_inputs(paired, raw, plan)
print("Reading release dataset metadata", flush=True)
datasets = dataset_records(metadata, catalog, known_missingness)
metadata["has_missing"] = metadata.dataset.map(
{d["id"]: d["hasMissing"] for d in datasets}
)
domains = []
domain_failures = []
for slug, subset in metadata.groupby("domain_slug", sort=True):
names = set(subset.dataset)
name = str(subset.domain.iloc[0])
domains.append(
dict(
slug=slug,
name="Drug Discovery" if slug == "drug" else name,
fullName=name,
nDatasets=len(subset),
nWithMissing=int(subset.has_missing.sum()),
samplesMin=int(subset.num_samples.min()),
samplesMax=int(subset.num_samples.max()),
featuresMin=int(subset.num_features.min()),
featuresMax=int(subset.num_features.max()),
taskCounts=dict(Counter(subset.task)),
)
)
applicable = failures[failures.dataset.isin(names)]
failed = applicable[~applicable.completed]
causes = Counter(failed.cause)
domain_failures.append(
dict(
domain=name,
domainSlug=slug,
nExpected=len(applicable),
nFailed=len(failed),
**{
k: causes[k]
for k in (
"gpu_oom",
"host_ram_oom",
"time_limit",
"segmentation_fault",
"mixed",
"unknown",
)
},
)
)
classification_quadrants = (
read("elo_regimes/classification_datasets.csv")
.set_index("dataset")
.quadrant.str.split()
.str[0]
.to_dict()
)
saved = saved_ratings(read)
jobs = [
(key, label, subset, raw, paired, timings, models, saved.get(key, {}))
for key, label, subset in scopes_for(
datasets, domains, classification_quadrants
)
if subset
]
scopes = {}
print("Building static benchmark scopes", flush=True)
with ProcessPoolExecutor(max_workers=args.workers) as pool:
for key, block in pool.map(scope_result, jobs):
scopes[key] = block
print(key, block["nDatasets"], flush=True)
paper_elo = json.loads((args.summaries / "elo/summary.json").read_text())
if (
scopes["global_all_tasks"]["paired"]["nDatasets"]
!= paper_elo["rf_imputed"]["n_datasets"]
):
raise ValueError("RF-imputed scope disagrees with the paper")
report_identity = hashlib.sha256(
"\n".join(sorted(raw.report_sha256)).encode()
).hexdigest()
revision = json.loads((PACKAGE / "configs/data_source.json").read_text())[
"revision"
]
output = dict(
meta=dict(
benchmarkName="TabBioMed",
subtitle="Benchmark for Biomedical Tabular Learning",
generatedAt=datetime.now(timezone.utc).isoformat(),
resultsRevision=f"sha256:{report_identity}",
dataRevision=revision,
nDatasets=len(datasets),
nDomains=len(domains),
nModelsAllTasks=len(config.expected_models_regression),
nModelsClassification=len(models),
medLog10Samples=float(np.median(np.log10(metadata.num_samples))),
medLog10Features=float(np.median(np.log10(metadata.num_features))),
),
domains=domains,
models=list(models.values()),
datasets=datasets,
evaluations=evaluation_records(raw, timings, failures, models, datasets),
scopes=scopes,
domainFailures=domain_failures,
)
text = json.dumps(
output, ensure_ascii=False, allow_nan=False, separators=(",", ":")
)
target = args.output
target.parent.mkdir(parents=True, exist_ok=True)
staged = target.with_suffix(".tmp")
staged.write_text(text + "\n", encoding="utf-8")
staged.replace(target)
print(
f"Wrote {target}: {len(scopes)} scopes, {len(output['evaluations'])} evaluations"
)
if __name__ == "__main__":
main()