GNN4Colliders / scripts /plot_label_efficiency.py
ho22joshua's picture
feat: evaluation subsample, Slurm wrappers for the study and label-efficiency plots
e71731f
Raw History Blame
6.91 kB
"""Plot validation AUC against the number of labelled training events, and
write the same numbers as Markdown and LaTeX tables.
Reads the sweeps of ``scripts/label_efficiency.py`` (``<out>/<name>/n<N>/
<split>_metrics.json``, one series per ``<name>``) and the probes of
``scripts/pretrain.sh`` (``<run>/probe/<split>_n<N>.json``, two series per
run: k-NN and linear):
uv run --extra plots python scripts/plot_label_efficiency.py \\
--sweeps outputs/label_efficiency --probes outputs/delphes_jepa \\
--output plots/label_efficiency
writes ``plots/label_efficiency.{pdf,png,md,tex}``. ``--metric`` chooses
another metric (``accuracy``, ``f1``, ...).
"""
from __future__ import annotations
import argparse
import json
import re
from pathlib import Path
#: Categorical colours in fixed order (validated for colour-vision
#: deficiency on a light background) and marker shapes, so that series are
#: not told apart by colour alone.
COLOURS = (
"#2a78d6",
"#eb6834",
"#1baf7a",
"#eda100",
"#e87ba4",
"#008300",
"#6250d6",
"#e34948",
)
MARKERS = ("o", "s", "^", "D", "v", "P", "X", "*")
TEXT, MUTED, GRID = "#2b2b29", "#6b6a64", "#e4e3de"
def _sweeps(root: Path, split: str, metric: str) -> dict[str, dict[int, float]]:
series: dict[str, dict[int, float]] = {}
for path in sorted(root.glob(f"*/n*/{split}_metrics.json")):
events = int(path.parent.name.removeprefix("n"))
value = json.loads(path.read_text())["metrics"][metric]
series.setdefault(path.parent.parent.name, {})[events] = value
return series
def _probes(run: Path, split: str, metric: str) -> dict[str, dict[int, float]]:
series: dict[str, dict[int, float]] = {}
for path in sorted((run / "probe").glob(f"{split}_n*.json")):
match = re.fullmatch(rf"{split}_n(\d+)\.json", path.name)
if match is None:
continue
summary = json.loads(path.read_text())
for probe, label in (("knn", "k-NN"), ("linear", "linear")):
name = f"{run.name} ({label} probe)"
series.setdefault(name, {})[int(match[1])] = summary[probe][metric]
return series
def _tables(series: dict[str, dict[int, float]], metric: str) -> tuple[str, str]:
counts = sorted({count for values in series.values() for count in values})
def cell(values: dict[int, float], count: int) -> str:
return f"{100 * values[count]:.2f}" if count in values else "--"
header = ["model", *(f"N = {count}" for count in counts)]
markdown = [
"| " + " | ".join(header) + " |",
"|---" * len(header) + "|",
*(
"| " + " | ".join([name, *(cell(v, c) for c in counts)]) + " |"
for name, v in series.items()
),
]
latex = [
r"\begin{tabular}{l" + "r" * len(counts) + "}",
r"\hline",
" & ".join(["Model", *(f"$N={count}$" for count in counts)]) + r" \\",
r"\hline",
*(
" & ".join([name.replace("_", r"\_"), *(cell(v, c) for c in counts)])
+ r" \\"
for name, v in series.items()
),
r"\hline",
r"\end{tabular}",
f"% {metric} (%) against labelled training events N",
]
return "\n".join(markdown) + "\n", "\n".join(latex) + "\n"
def _plot(series: dict[str, dict[int, float]], metric: str, output: Path) -> None:
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
from matplotlib.ticker import FuncFormatter, LogLocator
if len(series) > len(COLOURS):
raise SystemExit(f"at most {len(COLOURS)} series; plot fewer per figure")
figure, axes = plt.subplots(figsize=(6.4, 4.2), layout="constrained")
figure.set_facecolor("#fcfcfb")
axes.set_facecolor("#fcfcfb")
for index, (name, values) in enumerate(series.items()):
counts = sorted(values)
percent = [100 * values[count] for count in counts]
axes.plot(
counts,
percent,
color=COLOURS[index],
marker=MARKERS[index],
markersize=7,
linewidth=2,
markeredgecolor="#fcfcfb",
markeredgewidth=1.5,
label=name,
)
# A direct label at the last point when there are few series.
if len(series) <= 4:
axes.annotate(
name,
(counts[-1], percent[-1]),
xytext=(6, 0),
textcoords="offset points",
va="center",
fontsize=8,
color=TEXT,
)
axes.set_xscale("log")
axes.xaxis.set_major_locator(LogLocator(base=10))
axes.xaxis.set_major_formatter(FuncFormatter(lambda x, _: f"{x:,.0f}"))
if len(series) <= 4: # room on the right for the direct labels
low, high = axes.get_xlim()
axes.set_xlim(low, high * 4)
axes.set_xlabel("labelled training events", color=TEXT)
label = "AUC" if metric == "auc" else metric
axes.set_ylabel(f"validation {label} (%)", color=TEXT)
axes.grid(True, which="major", color=GRID, linewidth=0.8)
axes.tick_params(colors=MUTED)
for spine in ("top", "right"):
axes.spines[spine].set_visible(False)
for spine in ("left", "bottom"):
axes.spines[spine].set_color(GRID)
if len(series) > 1:
axes.legend(frameon=False, fontsize=8, labelcolor=TEXT)
for suffix in (".pdf", ".png"):
figure.savefig(output.with_suffix(suffix), dpi=200)
plt.close(figure)
def main() -> None:
parser = argparse.ArgumentParser(
description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter
)
parser.add_argument("--sweeps", nargs="*", type=Path, default=[])
parser.add_argument("--probes", nargs="*", type=Path, default=[])
parser.add_argument("--split", default="validation")
parser.add_argument("--metric", default="auc")
parser.add_argument("--output", type=Path, default=Path("plots/label_efficiency"))
args = parser.parse_args()
series: dict[str, dict[int, float]] = {}
found = [_sweeps(root, args.split, args.metric) for root in args.sweeps]
found += [_probes(run, args.split, args.metric) for run in args.probes]
for part in found:
for name, values in part.items():
merged = series.setdefault(name, {})
if merged.keys() & values.keys():
raise SystemExit(f"two results for {name} at the same N; plot apart")
merged.update(values)
if not series:
raise SystemExit("no results found")
args.output.parent.mkdir(parents=True, exist_ok=True)
markdown, latex = _tables(series, args.metric)
args.output.with_suffix(".md").write_text(markdown)
args.output.with_suffix(".tex").write_text(latex)
_plot(series, args.metric, args.output)
print(markdown, end="")
if __name__ == "__main__":
main()