[Feature] Added floating-point conversion functions to dgl.transforms.functional (#3890)

* * Added half_(), float_(), and double_() functions to DGLHeteroGraph, HeteroNodeDataView, and HeteroEdgeDataView, for converting floating-point tensor data to float16, float32, or float64 precision

* * Extracted out private functions for floating-point type conversion, to reduce code duplication

* * Added test for floating-point data conversion functions, half_(), float_(), and double_()

* * Moved half_(), float_(), and double_() functions from HeteroNodeDataView and HeteroEdgeDataView to Frame class

* * Updated test_float_cast() to use dgl.heterograph instead of dgl.graph

* Added to CONTRIBUTORS.md

* * Changed data type conversion to be deferred until the data is accessed, to avoid redundant conversions of data that isn't used.

* * Addressed issues flagged by linter

* * Worked around a bug in the old version of mxnet that's currently used for DGL testing

* * Only defer Column data type co...

CONTRIBUTORS.md

@@ -62,3 +62,4 @@ Contributors
 * [Abdurrahman Yasar](https://github.com/ayasar70) from Nvidia
 * [Shaked Brody](https://github.com/shakedbr) from Technion
 * [Jiahui Liu](https://github.com/paoxiaode) from Nvidia
+* [Neil Dickson](https://github.com/ndickson-nvidia) from Nvidia

python/dgl/frame.py

@@ -182,11 +182,12 @@ class Column(TensorStorage):
     index : Tensor
         Index tensor
     """
-    def __init__(self, storage, scheme=None, index=None, device=None):
+    def __init__(self, storage, scheme=None, index=None, device=None, deferred_dtype=None):
         super().__init__(storage)
         self.scheme = scheme if scheme else infer_scheme(storage)
         self.index = index
         self.device = device
+        self.deferred_dtype = deferred_dtype
         self.pinned_by_dgl = False
 
     def __len__(self):
@@ -230,6 +231,11 @@ class Column(TensorStorage):
         if self.device is not None:
             self.storage = F.copy_to(self.storage, self.device[0], **self.device[1])
             self.device = None
+
+        # convert data to the right type
+        if self.deferred_dtype is not None:
+            self.storage = F.astype(self.storage, self.deferred_dtype)
+            self.deferred_dtype = None
         return self.storage
 
     @data.setter
@@ -258,6 +264,49 @@ class Column(TensorStorage):
         col.device = (device, kwargs)
         return col
 
+    @property
+    def dtype(self):
+        """ Return the effective data type of this Column """
+        if self.deferred_dtype is not None:
+            return self.deferred_dtype
+        return self.storage.dtype
+
+    def astype(self, new_dtype):
+        """ Return a new column such that when its data is requested,
+        it will be converted to new_dtype.
+
+        Parameters
+        ----------
+        new_dtype : Framework-specific type object
+            The type to convert the data to.
+
+        Returns
+        -------
+        Column
+            A new column
+        """
+        col = self.clone()
+        if col.dtype != new_dtype:
+            # If there is already a pending conversion, ensure that the pending
+            # conversion and transfer/sampling are done before this new conversion.
+            if col.deferred_dtype is not None:
+                _ = col.data
+
+            if (col.device is None) and (col.index is None):
+                # Do the conversion immediately if no device transfer or index
+                # sampling is pending.  The assumption is that this is most
+                # likely to be the desired behaviour, such as converting an
+                # entire graph's feature data to float16 (half) before transfer
+                # to device when training, or converting back to float32 (float)
+                # after fetching the data to a device.
+                col.storage = F.astype(col.storage, new_dtype)
+            else:
+                # Defer the conversion if there is a pending transfer or sampling.
+                # This is so that feature data that never gets accessed on the
+                # device never needs to be transferred or sampled or converted.
+                col.deferred_dtype = new_dtype
+        return col
+
     def __getitem__(self, rowids):
         """Return the feature data given the rowids.
 
@@ -329,7 +378,7 @@ class Column(TensorStorage):
 
     def clone(self):
         """Return a shallow copy of this column."""
-        return Column(self.storage, self.scheme, self.index, self.device)
+        return Column(self.storage, self.scheme, self.index, self.device, self.deferred_dtype)
 
     def deepclone(self):
         """Return a deepcopy of this column.
@@ -358,13 +407,13 @@ class Column(TensorStorage):
             Sub-column
         """
         if self.index is None:
-            return Column(self.storage, self.scheme, rowids, self.device)
+            return Column(self.storage, self.scheme, rowids, self.device, self.deferred_dtype)
         else:
             index = self.index
             if not isinstance(index, _LazyIndex):
                 index = _LazyIndex(self.index)
             index = index.slice(rowids)
-            return Column(self.storage, self.scheme, index, self.device)
+            return Column(self.storage, self.scheme, index, self.device, self.deferred_dtype)
 
     @staticmethod
     def create(data):
@@ -792,3 +841,32 @@ class Frame(MutableMapping):
         if necessary."""
         for column in self._columns.values():
             column.unpin_memory_()
+
+    def _astype_float(self, new_type):
+        assert new_type in [F.float64, F.float32, F.float16], \
+            "'new_type' must be floating-point type: %s" % str(new_type)
+        newframe = self.clone()
+        new_columns = {}
+        for name, column in self._columns.items():
+            dtype = column.dtype
+            if dtype != new_type and dtype in [F.float64, F.float32, F.float16]:
+                new_columns[name] = column.astype(new_type)
+            else:
+                new_columns[name] = column
+        newframe._columns = new_columns
+        return newframe
+
+    def half(self):
+        """ Return a new frame with all floating-point columns converted
+        to half-precision (float16) """
+        return self._astype_float(F.float16)
+
+    def float(self):
+        """ Return a new frame with all floating-point columns converted
+        to single-precision (float32) """
+        return self._astype_float(F.float32)
+
+    def double(self):
+        """ Return a new frame with all floating-point columns converted
+        to double-precision (float64) """
+        return self._astype_float(F.float64)

python/dgl/transforms/functional.py

@@ -17,6 +17,7 @@
 
 from collections.abc import Iterable, Mapping
 from collections import defaultdict
+import copy
 import numpy as np
 import scipy.sparse as sparse
 import scipy.sparse.linalg
@@ -3530,4 +3531,55 @@ def laplacian_pe(g, k):
 
     return PE
 
+def to_half(g):
+    r"""Cast this graph to use float16 (half-precision) for any
+    floating-point edge and node feature data.
+
+    A shallow copy is returned so that the original graph is not modified.
+    Feature tensors that are not floating-point will not be modified.
+
+    Returns
+    -------
+    DGLHeteroGraph
+        Clone of graph with the feature data converted to float16.
+    """
+    ret = copy.copy(g)
+    ret._edge_frames = [frame.half() for frame in ret._edge_frames]
+    ret._node_frames = [frame.half() for frame in ret._node_frames]
+    return ret
+
+def to_float(g):
+    r"""Cast this graph to use float32 (single-precision) for any
+    floating-point edge and node feature data.
+
+    A shallow copy is returned so that the original graph is not modified.
+    Feature tensors that are not floating-point will not be modified.
+
+    Returns
+    -------
+    DGLHeteroGraph
+        Clone of graph with the feature data converted to float32.
+    """
+    ret = copy.copy(g)
+    ret._edge_frames = [frame.float() for frame in ret._edge_frames]
+    ret._node_frames = [frame.float() for frame in ret._node_frames]
+    return ret
+
+def to_double(g):
+    r"""Cast this graph to use float64 (double-precision) for any
+    floating-point edge and node feature data.
+
+    A shallow copy is returned so that the original graph is not modified.
+    Feature tensors that are not floating-point will not be modified.
+
+    Returns
+    -------
+    DGLHeteroGraph
+        Clone of graph with the feature data converted to float64.
+    """
+    ret = copy.copy(g)
+    ret._edge_frames = [frame.double() for frame in ret._edge_frames]
+    ret._node_frames = [frame.double() for frame in ret._node_frames]
+    return ret
+
 _init_api("dgl.transform", __name__)

tests/compute/test_heterograph.py

@@ -1924,6 +1924,65 @@ def test_dtype_cast(idtype):
         assert g_cast.idtype == F.int32
     test_utils.check_graph_equal(g, g_cast, check_idtype=False)
 
+def test_float_cast():
+    for t in [F.float16, F.float32, F.float64]:
+        idtype = F.int32
+        g = dgl.heterograph({
+            ('user', 'follows', 'user'): (F.tensor([0, 1, 1, 2, 2, 3], dtype=idtype),
+                                        F.tensor([0, 0, 1, 1, 2, 2], dtype=idtype)),
+            ('user', 'plays', 'game'): (F.tensor([0, 1, 1], dtype=idtype),
+                                        F.tensor([0, 0, 1], dtype=idtype))},
+            idtype=idtype, device=F.ctx())
+        uvalues = [1, 2, 3, 4]
+        gvalues = [5, 6]
+        fvalues = [7, 8, 9, 10, 11, 12]
+        pvalues = [13, 14, 15]
+        dataNamesTypes = [
+            ('a',F.float16),
+            ('b',F.float32),
+            ('c',F.float64),
+            ('d',F.int32),
+            ('e',F.int64)]
+        for name,type in dataNamesTypes:
+            g.nodes['user'].data[name] = F.copy_to(F.tensor(uvalues, dtype=type), ctx=F.ctx())
+        for name,type in dataNamesTypes:
+            g.nodes['game'].data[name] = F.copy_to(F.tensor(gvalues, dtype=type), ctx=F.ctx())
+        for name,type in dataNamesTypes:
+            g.edges['follows'].data[name] = F.copy_to(F.tensor(fvalues, dtype=type), ctx=F.ctx())
+        for name,type in dataNamesTypes:
+            g.edges['plays'].data[name] = F.copy_to(F.tensor(pvalues, dtype=type), ctx=F.ctx())
+
+        if t == F.float16:
+            g = dgl.transforms.functional.to_half(g)
+        if t == F.float32:
+            g = dgl.transforms.functional.to_float(g)
+        if t == F.float64:
+            g = dgl.transforms.functional.to_double(g)
+
+        for name,origType in dataNamesTypes:
+            # integer tensors shouldn't be converted
+            reqType = t if (origType in [F.float16,F.float32,F.float64]) else origType
+
+            values = g.nodes['user'].data[name]
+            assert values.dtype == reqType
+            assert len(values) == len(uvalues)
+            assert F.allclose(values, F.tensor(uvalues), 0, 0)
+
+            values = g.nodes['game'].data[name]
+            assert values.dtype == reqType
+            assert len(values) == len(gvalues)
+            assert F.allclose(values, F.tensor(gvalues), 0, 0)
+
+            values = g.edges['follows'].data[name]
+            assert values.dtype == reqType
+            assert len(values) == len(fvalues)
+            assert F.allclose(values, F.tensor(fvalues), 0, 0)
+
+            values = g.edges['plays'].data[name]
+            assert values.dtype == reqType
+            assert len(values) == len(pvalues)
+            assert F.allclose(values, F.tensor(pvalues), 0, 0)
+
 @parametrize_dtype
 def test_format(idtype):
     # single relation
@@ -2865,6 +2924,7 @@ if __name__ == '__main__':
     # test_isolated_ntype()
     # test_bipartite()
     # test_dtype_cast()
+    # test_float_cast()
     # test_reverse("int32")
     # test_format()
     #test_add_edges(F.int32)