1 files changed, 33 insertions, 0 deletions
diff --git a/ep_run/watch_hr.py b/ep_run/watch_hr.py
new file mode 100644
index 0000000..33a44bb
--- /dev/null
+++ b/ep_run/watch_hr.py
@@ -0,0 +1,33 @@
+"""Watch EP(hr=0.2) [the fix] + BPTT. Fire on: EP diverges (fix failed) / EP reaches good loss (fix worked) / exit."""
+import time, os, re
+RUNS = [
+    ("ep_hr02", "/home/yurenh2/ept/ep_run/runs/ep_hr02.log",   1684249),
+    ("bptt",    "/home/yurenh2/ept/ep_run/runs/bptt_clean.log", 1646261),
+]
+def alive(pid):
+    try: os.kill(pid, 0); return True
+    except Exception: return False
+def latest(log):
+    try: ls = [l for l in open(log) if l.startswith("step")]
+    except FileNotFoundError: return None
+    if not ls: return None
+    m = re.search(r"step (\d+)/.*val CE ([\d.eE+-]+).*res=([\d.eE+-]+)", ls[-1])
+    return (int(m.group(1)), float(m.group(2)), float(m.group(3)), ls[-1].strip()) if m else None
+def status():
+    return "\n".join(f"[{t}] {'ALIVE' if alive(p) else 'DEAD'} | {(latest(l) or [None,None,None,'no steps'])[3]}" for t, l, p in RUNS)
+t0 = time.time(); fired = None
+while fired is None and time.time() - t0 < 18 * 3600:
+    for t, l, p in RUNS:
+        d = latest(l)
+        if d:
+            step, val, res, _ = d
+            if t == "ep_hr02":
+                if res > 0.2 or val > 15: fired = f"EP(hr=0.2) STILL DIVERGED res={res:.2e} val={val:.2f} step {step} -> hr was not the (only) cause"; break
+                if val < 2.00: fired = f"EP(hr=0.2) reached GOOD loss val {val:.4f} (res {res:.2e}) step {step} -> the hr fix WORKED (past the old 2.09 wall region)"; break
+            if t == "bptt":
+                if val < 1.95: fired = f"bptt reached GOOD loss val {val:.4f} (res {res:.2e}) step {step}"; break
+        if not alive(p):
+            fired = f"{t} EXITED; last: {(latest(l) or [None,None,None,'no steps'])[3]}"; break
+    if fired: break
+    time.sleep(180)
+print("=== HR-FIX WATCHER FIRED ==="); print("trigger:", fired or "18h timeout"); print(status())