Merge pull request #562 from helmholtz-analytics/bug/squeeze-split-se…

…mantics

Bug/squeeze split semantics

CHANGELOG.md

@@ -2,8 +2,8 @@
 
 - Update documentation theme to "Read the Docs"
 - [#429](https://github.com/helmholtz-analytics/heat/pull/429) Create submodule for Linear Algebra functions
-- [#429](https://github.com/helmholtz-analytics/heat/pull/429) Implementated QR
-- [#429](https://github.com/helmholtz-analytics/heat/pull/429) Implementated a tiling class to create Square tiles along the diagonal of a 2D matrix
+- [#429](https://github.com/helmholtz-analytics/heat/pull/429) Implemented QR
+- [#429](https://github.com/helmholtz-analytics/heat/pull/429) Implemented a tiling class to create Square tiles along the diagonal of a 2D matrix
 - [#429](https://github.com/helmholtz-analytics/heat/pull/429) Added PyTorch Jitter to inner function of matmul for increased speed
 - [#483](https://github.com/helmholtz-analytics/heat/pull/483) Bugfix: Underlying torch tensor moves to the right device on array initialisation
 - [#483](https://github.com/helmholtz-analytics/heat/pull/483) Bugfix: DNDarray.cpu() changes heat device to cpu
@@ -29,6 +29,7 @@
 - [#536](https://github.com/helmholtz-analytics/heat/pull/536) Getting rid of the docs folder
 - [#558](https://github.com/helmholtz-analytics/heat/pull/558) `sanitize_memory_layout` assumes default memory layout of the input tensor
 - [#558](https://github.com/helmholtz-analytics/heat/pull/558) Support for PyTorch 1.5.0 added
+- [#562](https://github.com/helmholtz-analytics/heat/pull/562) Bugfix: split semantics of ht.squeeze()
 
 # v0.3.0
 

heat/core/manipulations.py

@@ -1196,9 +1196,10 @@ def squeeze(x, axis=None):
     --------
     squeezed : ht.DNDarray
                The input tensor, but with all or a subset of the dimensions of length 1 removed.
-
+               Split semantics: see note below.
 
     Examples:
+    ---------
     >>> import heat as ht
     >>> import torch
     >>> torch.manual_seed(1)
@@ -1226,6 +1227,19 @@ def squeeze(x, axis=None):
     Traceback (most recent call last):
     ...
     ValueError: Dimension along axis 1 is not 1 for shape (1, 3, 1, 5)
+
+    Note:
+    -----
+    Split semantics: a distributed tensor will keep its original split dimension after "squeezing",
+    which, depending on the squeeze axis, may result in a lower numerical 'split' value, as in:
+    >>> x.shape
+    (10, 1, 12, 13)
+    >>> x.split
+    2
+    >>> x.squeeze().shape
+    (10, 12, 13)
+    >>> x.squeeze().split
+    1
     """
 
     # Sanitize input
@@ -1236,58 +1250,31 @@ def squeeze(x, axis=None):
     if axis is not None:
         if isinstance(axis, int):
             dim_is_one = x.shape[axis] == 1
-        if isinstance(axis, tuple):
-            dim_is_one = bool(
-                factories.array(list(x.shape[dim] == 1 for dim in axis)).all()._DNDarray__array
-            )
+            axis = (axis,)
+        elif isinstance(axis, tuple):
+            dim_is_one = bool(torch.tensor(list(x.shape[dim] == 1 for dim in axis)).all())
         if not dim_is_one:
             raise ValueError("Dimension along axis {} is not 1 for shape {}".format(axis, x.shape))
 
-    # Local squeeze
     if axis is None:
         axis = tuple(i for i, dim in enumerate(x.shape) if dim == 1)
-    if isinstance(axis, int):
-        axis = (axis,)
-    out_lshape = tuple(x.lshape[dim] for dim in range(len(x.lshape)) if dim not in axis)
+
+    if x.split is not None and x.split in axis:
+        # split dimension is about to disappear, set split to None
+        x.resplit_(axis=None)
+
+    out_lshape = tuple(x.lshape[dim] for dim in range(x.numdims) if dim not in axis)
+    out_gshape = tuple(x.gshape[dim] for dim in range(x.numdims) if dim not in axis)
     x_lsqueezed = x._DNDarray__array.reshape(out_lshape)
 
-    # Calculate split axis according to squeezed shape
+    # Calculate new split axis according to squeezed shape
     if x.split is not None:
         split = x.split - len(list(dim for dim in axis if dim < x.split))
     else:
-        split = x.split
-
-    # Distributed squeeze
-    if x.split is not None:
-        if x.comm.is_distributed():
-            if x.split in axis:
-                raise ValueError(
-                    "Cannot split AND squeeze along same axis. Split is {}, axis is {} for shape {}".format(
-                        x.split, axis, x.shape
-                    )
-                )
-            out_gshape = tuple(x.gshape[dim] for dim in range(len(x.gshape)) if dim not in axis)
-            x_gsqueezed = factories.empty(out_gshape, dtype=x.dtype)
-            loffset = factories.zeros(1, dtype=types.int64)
-            loffset.__setitem__(0, x.comm.chunk(x.gshape, x.split)[0])
-            displs = factories.zeros(x.comm.size, dtype=types.int64)
-            x.comm.Allgather(loffset, displs)
-
-            # TODO: address uneven distribution of dimensions (Allgatherv). Issue #273, #233
-            x.comm.Allgather(
-                x_lsqueezed, x_gsqueezed
-            )  # works with evenly distributed dimensions only
-            return dndarray.DNDarray(
-                x_gsqueezed,
-                out_gshape,
-                x_lsqueezed.dtype,
-                split=split,
-                device=x.device,
-                comm=x.comm,
-            )
+        split = None
 
     return dndarray.DNDarray(
-        x_lsqueezed, out_lshape, x.dtype, split=split, device=x.device, comm=x.comm
+        x_lsqueezed, out_gshape, x.dtype, split=split, device=x.device, comm=x.comm
     )
 
 

heat/core/tests/test_manipulations.py

@@ -1129,6 +1129,7 @@ def test_sort(self):
             ht.sort(data, axis="1")
 
         rank = ht.MPI_WORLD.rank
+        ht.random.seed(1)
         data = ht.random.randn(100, 1, split=0, device=ht_device)
         result, _ = ht.sort(data, axis=0)
         counts, _, _ = ht.get_comm().counts_displs_shape(data.gshape, axis=0)
@@ -1307,40 +1308,53 @@ def test_squeeze(self):
         self.assertTrue((result._DNDarray__array == data._DNDarray__array.squeeze()).all())
 
         # 4D split tensor, along the axis
-        # TODO: reinstate this test of uneven dimensions distribution
-        # after update to Allgatherv implementation (Issue  #273 depending on #233)
-        # data = ht.array(ht.random.randn(1, 4, 5, 1), split=1)
-        # result = ht.squeeze(data, axis=-1)
-        # self.assertIsInstance(result, ht.DNDarray)
-        # # TODO: the following works locally but not when distributed,
-        # #self.assertEqual(result.dtype, ht.float32)
-        # #self.assertEqual(result._DNDarray__array.dtype, torch.float32)
-        # self.assertEqual(result.shape, (1, 12, 5))
-        # self.assertEqual(result.lshape, (1, 12, 5))
-        # self.assertEqual(result.split, 1)
+        data = ht.array(ht.random.randn(1, 4, 5, 1), split=1)
+        result = ht.squeeze(data, axis=-1)
+        self.assertIsInstance(result, ht.DNDarray)
+        self.assertEqual(result.dtype, ht.float32)
+        self.assertEqual(result._DNDarray__array.dtype, torch.float32)
+        self.assertEqual(result.shape, (1, 4, 5))
+        self.assertEqual(result.split, 1)
+
+        # 4D split tensor, axis = split
+        data = ht.array(ht.random.randn(3, 1, 5, 6), split=1)
+        result = ht.squeeze(data, axis=1)
+        self.assertIsInstance(result, ht.DNDarray)
+        self.assertEqual(result.dtype, ht.float32)
+        self.assertEqual(result._DNDarray__array.dtype, torch.float32)
+        self.assertEqual(result.shape, (3, 5, 6))
+        self.assertEqual(result.split, None)
+
+        # 4D split tensor, axis = split = last dimension
+        data = ht.array(ht.random.randn(3, 6, 5, 1), split=-1)
+        result = ht.squeeze(data, axis=-1)
+        self.assertIsInstance(result, ht.DNDarray)
+        self.assertEqual(result.dtype, ht.float32)
+        self.assertEqual(result._DNDarray__array.dtype, torch.float32)
+        self.assertEqual(result.shape, (3, 6, 5))
+        self.assertEqual(result.split, None)
 
         # 3D split tensor, across the axis
-        size = ht.MPI_WORLD.size * 2
-        data = ht.triu(ht.ones((1, size, size), split=1, device=ht_device), k=1)
+        size = ht.MPI_WORLD.size
+        data = ht.triu(ht.ones((1, size * 2, size), split=1, device=ht_device), k=1)
 
         result = ht.squeeze(data, axis=0)
         self.assertIsInstance(result, ht.DNDarray)
-        # TODO: the following works locally but not when distributed,
-        # self.assertEqual(result.dtype, ht.float32)
-        # self.assertEqual(result._DNDarray__array.dtype, torch.float32)
-        self.assertEqual(result.shape, (size, size))
-        self.assertEqual(result.lshape, (size, size))
-        # self.assertEqual(result.split, None)
+        self.assertEqual(result.dtype, ht.float32)
+        self.assertEqual(result._DNDarray__array.dtype, torch.float32)
+        self.assertEqual(result.shape, (size * 2, size))
+        self.assertEqual(result.lshape, (2, size))
+        self.assertEqual(result.split, 0)
 
         # check exceptions
-        with self.assertRaises(ValueError):
-            data.squeeze(axis=(0, 1))
         with self.assertRaises(TypeError):
             data.squeeze(axis=1.1)
         with self.assertRaises(TypeError):
             data.squeeze(axis="y")
         with self.assertRaises(ValueError):
-            ht.argmin(data, axis=-4)
+            ht.squeeze(data, axis=-4)
+        with self.assertRaises(ValueError):
+            ht.squeeze(data, axis=1)
 
     def test_unique(self):
         size = ht.MPI_WORLD.size