make_plots.py

"""Plotly HTML figures from the job's JSON results."""
import json, os, argparse
import plotly.graph_objects as go

BLUE, RED, GREEN, ORANGE = "#2563eb", "#dc2626", "#16a34a", "#f59e0b"


def claim2_plots(outdir):
    d = json.load(open(os.path.join(outdir, "claim2_results.json")))
    kb = d.get("kernel_bench", {})
    # ms vs B2 curves
    fig = go.Figure()
    pal = {"T-stage3": BLUE, "T-stage4": ORANGE, "B-stage3": GREEN, "B-stage4": RED}
    for name, r in kb.items():
        b2s = sorted(int(k) for k in r["ms_per_call_by_B2"])
        ms = [r["ms_per_call_by_B2"][str(b)] if str(b) in r["ms_per_call_by_B2"]
              else r["ms_per_call_by_B2"][b] for b in b2s]
        fig.add_trace(go.Scatter(x=b2s, y=ms, mode="lines+markers", name=name,
                                 line=dict(color=pal.get(name), width=3)))
    fig.update_layout(title="SSM kernel time vs fold factor B2 (A100, B=128, S=8)",
                      xaxis_title="fold factor B2 (B1=B/B2)", xaxis_type="log",
                      yaxis_title="ms per selective_scan call",
                      template="plotly_white", width=640, height=440)
    fig.write_html(os.path.join(outdir, "claim2_curves.html"), include_plotlyjs="cdn")
    # speedup bars vs paper range
    names = list(kb)
    sp = [kb[n]["speedup_pct"] for n in names]
    fig = go.Figure(go.Bar(x=names, y=sp, marker_color=BLUE,
                           text=[f"{s:.0f}%" for s in sp], textposition="outside"))
    fig.add_hrect(y0=110, y1=180, fillcolor="rgba(22,163,74,0.15)", line_width=0,
                  annotation_text="paper: 110–180%")
    fig.update_layout(title="Claim 2: SSM kernel speedup from batch folding",
                      yaxis_title="speedup vs B2=1 (%)",
                      template="plotly_white", width=600, height=440)
    fig.write_html(os.path.join(outdir, "claim2_speedup.html"), include_plotlyjs="cdn")


def claim1_plot(outdir):
    d = json.load(open(os.path.join(outdir, "claim1_results.json")))
    fig = go.Figure()
    pal = {"uni": RED, "swap": BLUE, "bidir": GREEN}
    for mode, r in d["modes"].items():
        acc = r["per_position_acc"]
        fig.add_trace(go.Scatter(x=list(range(1, len(acc) + 1)), y=acc,
                                 mode="lines+markers", name=mode,
                                 line=dict(color=pal.get(mode), width=3)))
    fig.add_hline(y=d["chance"], line_dash="dot", line_color="gray",
                  annotation_text="chance")
    fig.update_layout(title="Claim 1: per-position accuracy (label = last token's class)",
                      xaxis_title="position (1..T; T is the only causally-legal one)",
                      yaxis_title="test accuracy",
                      template="plotly_white", width=640, height=440)
    fig.write_html(os.path.join(outdir, "claim1_positions.html"), include_plotlyjs="cdn")


def claim34_plot(outdir):
    d = json.load(open(os.path.join(outdir, "claim34_results.json")))
    ours = {k: v for k, v in d["throughput_img_s"].items() if isinstance(v, (int, float))}
    paper = d["paper_img_s"]
    names = [n for n in ours if n in paper]
    fig = go.Figure()
    fig.add_trace(go.Bar(x=names, y=[paper[n] for n in names], name="paper (A100)",
                         marker_color="#94a3b8"))
    fig.add_trace(go.Bar(x=names, y=[ours[n] for n in names],
                         name=f"ours ({d['gpu']}, fp16)", marker_color=BLUE))
    fig.update_layout(title="Claims 3/4: throughput, paper vs our A100 (batch 128)",
                      yaxis_title="images / s", barmode="group",
                      template="plotly_white", width=680, height=440)
    fig.write_html(os.path.join(outdir, "claim34_throughput.html"), include_plotlyjs="cdn")


if __name__ == "__main__":
    ap = argparse.ArgumentParser()
    ap.add_argument("--out", default="outputs")
    a = ap.parse_args()
    for fn in (claim2_plots, claim1_plot, claim34_plot):
        try:
            fn(a.out)
            print(fn.__name__, "ok")
        except Exception as e:
            print(fn.__name__, "SKIP:", e)