Handle RunDPTrain when expanded training systems are empty

[SchrodingersCattt]

Problem

RunDPTrain.execute() can still launch dp train with no training systems. This happens when a pretrained init_model is supplied, init_data is empty, and the latest iter_data entry exists but expands to zero systems. Because the raw iter_data list is non-empty, skip_training() does not apply.

In that state, the generated input contains:

{
  "systems": [],
  "auto_prob": "prob_sys_size"
}

and dp train is invoked. PR #368 fixes invalid empty probability ranges, but the fully empty case needs an explicit workflow policy rather than another probability fallback.

Expected behavior

Detect the absence of expanded training systems before invoking DeePMD-kit, then either:

• skip training and propagate the supplied init_model, consistent with the existing no-iter_data path; or
• fail early with a clear FatalError explaining that there is no data to train on.

The chosen behavior should be covered by an end-to-end regression test that verifies dp train is not launched with training_data.systems == [].

Reproduction context

• init_model: supplied pretrained model
• init_data: []
• iter_data: a non-empty list whose latest entry expands to zero systems
• init_model_policy: "yes"

Related review discussion: #368 (comment)

[dosubot]

@SchrodingersCattt the root cause is straightforward. skip_training() checks the raw iter_data list, but the actual expansion to individual systems happens earlier via _expand_all_multi_sys_to_sys [1]. When iter_data is non-empty but expands to zero systems, the guard at line 454 never fires, and dp train proceeds with systems: [].

The fix should go after iter_data_exp is computed (line 208) and before write_data_to_input_script is called. Two options:

Option A – Skip training and propagate init_model (consistent with the existing no-iter_data path):

# After computing iter_data_exp (line 208)
if len(iter_data_exp) == 0 and init_data is not None and len(init_data) == 0:
    if init_model is not None:
        with set_directory(work_dir):
            with open(train_script_name, "w") as fp:
                json.dump(train_dict, fp, indent=4)
            Path("train.log").write_text(
                "Expanded training systems are empty and no init_data. "
                "Training skipped; propagating init_model.\n"
            )
            Path("lcurve.out").touch()
        return OPIO({
            "script": work_dir / train_script_name,
            "model": init_model,
            "lcurve": work_dir / "lcurve.out",
            "log": work_dir / "train.log",
        })

Option B – Fail early with FatalError:

if len(iter_data_exp) == 0 and (init_data is None or len(init_data) == 0):
    raise FatalError(
        "No training data: iter_data expanded to zero systems and init_data is empty. "
        "Cannot invoke dp train."
    )

Alternatively, skip_training() could accept iter_data_exp instead of (or in addition to) raw iter_data, so both the raw-empty and expanded-empty cases are handled in one place.

This is distinct from the fix in PR #368 — that PR handles the case where iter_data_exp is non-zero but the probability range numb_old:numb_new becomes degenerate. This issue is the fully-zero case where there is literally nothing to train on and a probability fallback cannot help.

A regression test should mock _expand_all_multi_sys_to_sys to return [] for the last entry while keeping iter_data non-empty, then assert that either dp train is never called or FatalError is raised.

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.