From 642a745d4d5444825aa28a54334e087add69af3b Mon Sep 17 00:00:00 2001
From: Cedric Conday <277679649+CedricConday@users.noreply.github.com>
Date: Thu, 9 Jul 2026 17:53:42 +0000
Subject: [PATCH] fix(deepmd/hdf5): skip empty optional frame arrays on dump
 (#996)

The deepmd/hdf5 writer reshaped and wrote optional frame arrays (e.g. forces
or virials that are empty when cal_force/cal_stress is disabled) as a
(nframes, 0) dataset, which the loader then failed to reshape back to the full
shape, raising 'cannot reshape array of size 0'. Skip empty optional arrays on
dump, matching the existing deepmd/raw and deepmd/npy writers. Non-empty arrays
are unaffected.
---
 dpdata/formats/deepmd/hdf5.py |  7 ++++++
 tests/test_deepmd_hdf5.py     | 47 +++++++++++++++++++++++++++++++++++
 2 files changed, 54 insertions(+)

diff --git a/dpdata/formats/deepmd/hdf5.py b/dpdata/formats/deepmd/hdf5.py
index 6349a960..5885bb19 100644
--- a/dpdata/formats/deepmd/hdf5.py
+++ b/dpdata/formats/deepmd/hdf5.py
@@ -201,6 +201,13 @@ def dump(
     for dt, prop in data_types.items():
         if dt in data:
             if prop["dump"]:
+                if nframes > 0 and np.asarray(data[dt]).size == 0:
+                    # An optional frame property (e.g. forces/virials when
+                    # cal_force/cal_stress is disabled) may be empty while the
+                    # system still has frames. Skip it instead of writing a
+                    # (nframes, 0) dataset that cannot be reshaped on load. This
+                    # matches the behaviour of the deepmd/raw and deepmd/npy writers.
+                    continue
                 ddata = np.reshape(data[dt], prop["shape"])
                 if np.issubdtype(ddata.dtype, np.floating):
                     ddata = ddata.astype(comp_prec)
diff --git a/tests/test_deepmd_hdf5.py b/tests/test_deepmd_hdf5.py
index 9d65f29a..a198bc3c 100644
--- a/tests/test_deepmd_hdf5.py
+++ b/tests/test_deepmd_hdf5.py
@@ -300,3 +300,50 @@ def test_regular_hdf5_groups_are_not_mixed(self):
         )
 
         self.assertEqual(len(systems), 0)
+
+
+class TestHDF5EmptyOptionalArray(unittest.TestCase):
+    """Regression test for #996.
+
+    An empty optional frame array (e.g. forces when they are not computed) must not
+    break the deepmd/hdf5 write/read round-trip. Previously the writer stored a
+    ``(nframes, 0)`` dataset that could not be reshaped on load.
+    """
+
+    def tearDown(self):
+        if os.path.isfile("tmp.deepmd.empty.hdf5"):
+            os.remove("tmp.deepmd.empty.hdf5")
+
+    @staticmethod
+    def _make(forces):
+        return dpdata.LabeledSystem(
+            data={
+                "atom_names": ["H"],
+                "atom_numbs": [1],
+                "atom_types": np.array([0]),
+                "cells": np.eye(3).reshape(1, 3, 3),
+                "coords": np.zeros((1, 1, 3)),
+                "energies": np.zeros(1),
+                "orig": np.zeros(3),
+                "forces": forces,
+            }
+        )
+
+    def test_empty_forces_round_trip(self):
+        system = self._make(np.zeros((1, 0, 3)))
+        system.to("deepmd/hdf5", "tmp.deepmd.empty.hdf5")
+        reloaded = dpdata.LabeledSystem("tmp.deepmd.empty.hdf5", fmt="deepmd/hdf5")
+        # The empty optional array is skipped rather than written as (nframes, 0).
+        self.assertNotIn("forces", reloaded.data)
+        np.testing.assert_allclose(reloaded["coords"], system["coords"])
+
+    def test_real_forces_preserved(self):
+        forces = np.arange(3, dtype=float).reshape(1, 1, 3)
+        system = self._make(forces)
+        system.to("deepmd/hdf5", "tmp.deepmd.empty.hdf5")
+        reloaded = dpdata.LabeledSystem("tmp.deepmd.empty.hdf5", fmt="deepmd/hdf5")
+        np.testing.assert_allclose(reloaded["forces"], forces)
+
+
+if __name__ == "__main__":
+    unittest.main()