import re, matplotlib; matplotlib.use('Agg'); import matplotlib.pyplot as plt
def parse(p):
    s,c,e=[],[],[]
    for ln in open(p):
        m=re.search(r'step\s+(\d+)/\d+ \| val CE ([\d.]+) ema=([\d.]+)', ln)
        if m: s.append(int(m.group(1))); c.append(float(m.group(2))); e.append(float(m.group(3)))
    return s,c,e
fz=parse('ep_run/runs/ep_resreg_warm.log'); ad=parse('ep_run/runs/ep_rr_ajr.log')
fig,ax=plt.subplots(figsize=(11,5.5),dpi=140)
for (s,c,e),col,lab in [(fz,'#4363d8','frozen jr=0.1'),(ad,'#e6194B','adaptive jacreg (jr_max=16)')]:
    ax.plot(s,c,color=col,lw=0.7,alpha=0.25)
    ax.plot(s,e,color=col,lw=2.3,label=lab+f'  (last ema {e[-1]:.4f})')
ax.axhline(2.0,color='gray',ls='--',lw=1,alpha=0.7,label='val CE = 2.0')
ax.set_xlabel('training step'); ax.set_ylabel('val CE'); ax.set_ylim(1.92,2.6)
ax.set_title('Frozen jr vs adaptive jacreg — C512 EP, same s2000 warm-start (only jr differs)\nbold = ema, faint = raw per-log val CE', fontsize=10.5)
ax.legend(fontsize=9.5, loc='upper right'); ax.grid(alpha=0.2)
fig.tight_layout(); fig.savefig('frozen_vs_adaptive.png',dpi=140,bbox_inches='tight')
print(f"frozen: {len(fz[0])} pts, steps {fz[0][0]}-{fz[0][-1]}, best ema {min(fz[2]):.4f}")
print(f"adaptive: {len(ad[0])} pts, steps {ad[0][0]}-{ad[0][-1]}, best ema {min(ad[2]):.4f}")
print("saved")