add svd feature reduction

Author: gyzhou2000

gammagl/transforms/__init__.py

@@ -6,6 +6,7 @@
 from .drop_edge import DropEdge
 from .random_link_split import RandomLinkSplit
 from .vgae_pre import mask_test_edges, sparse_to_tuple
+from .svd_feature_reduction import SVDFeatureReduction
 
 __all__ = [
     'BaseTransform',
@@ -16,7 +17,8 @@
     'DropEdge',
     'RandomLinkSplit',
     'mask_test_edges',
-    'sparse_to_tuple'
+    'sparse_to_tuple',
+    'SVDFeatureReduction'
 
 ]
 

gammagl/transforms/svd_feature_reduction.py

@@ -0,0 +1,37 @@
+from gammagl.transforms import BaseTransform
+
+from gammagl.data import Graph
+from typing import List
+import numpy as np
+import tensorlayerx as tlx
+
+
+class SVDFeatureReduction(BaseTransform):
+    r"""Dimensionality reduction of node features via Singular Value Decomposition (SVD)
+    (functional name: :obj:`normalize_features`).
+
+    Parameters
+    ----------
+    out_channels: int
+        The dimensionlity of node features after reduction.
+
+    """
+
+    def __init__(self, out_channels: int):
+        self.out_channels = out_channels
+
+    def __call__(self, graph: Graph):
+        assert graph.x is not None
+
+        if graph.x.shape[-1] > self.out_channels:
+            x = tlx.convert_to_numpy(graph.x)
+            U, S, _ = np.linalg.svd(x, full_matrices=False)
+            U_reduced = U[:, :self.out_channels]
+            S_reduced = np.diag(S[:self.out_channels])
+            x = np.dot(U_reduced, S_reduced)
+            graph.x = tlx.convert_to_tensor(x)
+
+        return graph
+
+    def __repr__(self) -> str:
+        return f'{self.__class__.__name__}()'

tests/transforms/test_svd_feature_reduction.py

@@ -0,0 +1,15 @@
+import tensorlayerx as tlx
+
+from gammagl.data import Graph, HeteroGraph
+from gammagl.transforms import SVDFeatureReduction
+
+
+def test_normalize_scale():
+    assert SVDFeatureReduction(2).__repr__() == 'SVDFeatureReduction()'
+
+    x = tlx.convert_to_tensor([[1, 2, 3, 4, 5], [5, 4, 3, 2, 1]], dtype=tlx.float32)
+
+    data = Graph(x=x)
+    data = SVDFeatureReduction(out_channels=2)(data)
+    assert len(data) == 1
+    assert tlx.get_tensor_shape(data.x) == [2, 2]