mechanistic_interp/scripts/diff_gradient_ascent.py

"""
Per-cell difference of two gradient-ascent heatmaps:  Δ = A − B  (per alpha, per (l,l')).

A and B are the JSON outputs of gradient_ascent.py (each {alphas, delta:{alpha:matrix}}).
Use to compare two models/conditions, e.g. base induction (A) minus LoRA induction (B):
positive cell ⇒ A's bathroom→toilet influence is larger there (LoRA reduced it).

Plots with the shared plot_heatmaps (normal | meannorm), same as gradient_ascent.py.
"""
import argparse
import json

import numpy as np
from mechanistic_interp.gradient_ascent import plot_heatmaps


def load(path):
    d = json.load(open(path))
    mats = {float(a): np.array([[np.nan if v is None else v for v in row] for row in M])
            for a, M in d["delta"].items()}
    return mats, d


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--json_a", required=True, help="minuend JSON (A)")
    ap.add_argument("--json_b", required=True, help="subtrahend JSON (B)")
    ap.add_argument("--out", required=True)
    ap.add_argument("--plot_mode", default="meannorm", choices=["normal", "meannorm"])
    ap.add_argument("--title", default=None)
    args = ap.parse_args()

    A, da = load(args.json_a)
    B, db = load(args.json_b)
    alphas = sorted(set(A) & set(B))
    if not alphas:
        raise SystemExit(f"No common alphas: A={sorted(A)} B={sorted(B)}")
    diff = {a: A[a] - B[a] for a in alphas}     # NaN where either is NaN (below diagonal)
    nA, nB = da.get("n_images", "?"), db.get("n_images", "?")
    title = args.title or f"A − B  (A: n={nA}, B: n={nB})  per cell — {args.plot_mode}"
    # integer-looking alpha labels read better
    alphas_disp = [int(a) if float(a).is_integer() else a for a in alphas]
    plot_heatmaps({ad: diff[a] for a, ad in zip(alphas, alphas_disp)}, alphas_disp,
                  args.out, title=title, mode=args.plot_mode, cbar_label="Δ_A − Δ_B")


if __name__ == "__main__":
    main()