[Code scan] Pressure task reports success_rate with the wrong denominator

[njzjz]

This issue was found by a Codex global repository scan of tracked non-test files at commit 8c93925cb10b401b2b83c738bd9263fd74474468.

Relevant code

LAMBench/lambench/tasks/calculator/pressure/pressure.py

Lines 57 to 84 in 8c93925

	num_samples = 0
	num_fails = 0
	for pressure in tqdm(["025", "050", "075", "100", "125", "150"]):
	init_traj = read(test_data / f"P{pressure}.traj", ":")
	final_traj = read(test_data / f"P{pressure}.traj", ":")
	for i in tqdm(range(len(init_traj))):
	init = init_traj[i]
	final = final_traj[i]
	assert init.get_chemical_formula() == final.get_chemical_formula()
	try:
	dft, lam = test_one(init, final, int(pressure), calc, fmax, max_steps)
	except Exception as e:
	logging.error(
	f"Error during test_one at pressure {pressure}, index {i}: {e}"
	)
	dft, lam = None, None
	if dft is None or lam is None:
	num_fails += 1
	continue
	num_samples += 1
	all_labels.append(dft)
	all_preds.append(lam)
	return {
	"MAE": mean_absolute_error(all_labels, all_preds), # A3/atom
	"RMSE": root_mean_squared_error(all_labels, all_preds), # A3/atom
	"success_rate": (num_samples - num_fails) / num_samples,

Impact

num_samples is incremented only after a successful calculation, while num_fails is incremented for failed attempts. The final formula is (num_samples - num_fails) / num_samples, which can underreport success and can crash when every sample fails.

For example, with one successful sample and one failed sample, the current formula reports (1 - 1) / 1 = 0.0 instead of the expected 1 / 2 = 0.5.

Suggested fix

Track total attempts separately from successful samples. Compute success_rate = num_successes / total_attempts, and return controlled NaN or None metrics plus 0.0 success when there are no successful samples.