"""Side-by-side cloverleaf comparison: mean-field × 3 seeds vs NTK vs narrow."""
from __future__ import annotations
from pathlib import Path
import numpy as np, pandas as pd, seaborn as sns, matplotlib.pyplot as plt, torch

ROOT = Path("/workspace/mft")
RUNS = [
    ("run_seed_1", "Mean-field, seed 1\n(width 2048)"),
    ("run_seed_2", "Mean-field, seed 2"),
    ("run_seed_3", "Mean-field, seed 3"),
    ("run_ntk",    "NTK scaling\n(no cloverleaf)"),
    ("run_narrow", "Narrow (width 16)\n(too few neurons)"),
]
sns.set_theme(style="white", context="paper")
fig, axes = plt.subplots(1, 5, figsize=(20, 4.2), sharex=True, sharey=True)
for ax, (run, title) in zip(axes, RUNS):
    ckpt = torch.load(ROOT/run/"model.pt", map_location="cpu", weights_only=True)
    W = ckpt["W"].numpy(); a = ckpt["a"].numpy()
    sgn = np.where(a > 0, "+", "−")
    df = pd.DataFrame({"w1": W[:,0], "w2": W[:,1], "sign_a": sgn})
    sns.scatterplot(df, x="w1", y="w2", hue="sign_a",
                    palette={"+":"#1f77b4","−":"#d62728"}, s=8, alpha=0.6,
                    edgecolor=None, ax=ax, legend=(ax is axes[0]))
    ax.set_title(title)
    ax.set_xlabel("w₁"); ax.set_ylabel("w₂" if ax is axes[0] else "")
    ax.set_xlim(-35,35); ax.set_ylim(-35,35); ax.set_aspect("equal")
fig.suptitle("Hidden-layer incoming-weight distribution by training regime",
             fontsize=13, y=1.02)
fig.tight_layout()
fig.savefig(ROOT/"comparison.pdf", bbox_inches="tight")
print("Wrote", ROOT/"comparison.pdf")

# Angle histogram comparison
fig, ax = plt.subplots(figsize=(9, 4))
for run, title in RUNS:
    ckpt = torch.load(ROOT/run/"model.pt", map_location="cpu", weights_only=True)
    W = ckpt["W"].numpy()
    # filter out near-dead neurons
    norms = np.linalg.norm(W, axis=1)
    mask = norms > np.quantile(norms, 0.5)
    angles = np.degrees(np.arctan2(W[mask,1], W[mask,0]))
    sns.kdeplot(angles, ax=ax, label=title.split('\n')[0], bw_adjust=0.4)
for x0 in [-180, -90, 0, 90, 180]:
    ax.axvline(x0, color="grey", lw=0.5, ls="--")
ax.set_xlabel("Incoming-weight angle [deg]"); ax.set_ylabel("Density")
ax.set_title("Top-50% by ‖w‖ — angle density")
ax.legend(fontsize=8, loc="upper right")
fig.tight_layout()
fig.savefig(ROOT/"angle_compare.pdf", bbox_inches="tight")
print("Wrote", ROOT/"angle_compare.pdf")