[NN] Add low memory support for nn.RelGraphConv (#1631)

* Add low memory support for relgraphconv
y

* lint

* Add tf support

* Fix lint

* Fix mxnet

* Fix mxnet

* Fix

* minor fix

* Fix
Co-authored-by: Ubuntu <ubuntu@ip-172-31-51-214.ec2.internal>

python/dgl/nn/mxnet/conv/relgraphconv.py

@@ -56,6 +56,9 @@ class RelGraphConv(gluon.Block):
         Activation function. Default: None
     self_loop : bool, optional
         True to include self loop message. Default: False
+    low_mem : bool, optional
+        Use low-memory implementation. MXNet currently does not support this.
+        Default: False.
     dropout : float, optional
         Dropout rate. Default: 0.0
     """
@@ -68,6 +71,7 @@ class RelGraphConv(gluon.Block):
                  bias=True,
                  activation=None,
                  self_loop=False,
+                 low_mem=False,
                  dropout=0.0):
         super(RelGraphConv, self).__init__()
         self.in_feat = in_feat
@@ -81,6 +85,8 @@ class RelGraphConv(gluon.Block):
         self.activation = activation
         self.self_loop = self_loop
 
+        assert low_mem is False, 'MXNet currently does not support low-memory implementation.'
+
         if regularizer == "basis":
             # add basis weights
             self.weight = self.params.get(

python/dgl/nn/pytorch/conv/relgraphconv.py

@@ -53,6 +53,10 @@ class RelGraphConv(nn.Module):
         Activation function. Default: None
     self_loop : bool, optional
         True to include self loop message. Default: False
+    low_mem : bool, optional
+        True to use low memory implementation of relation message passing function. Default: False
+        This option trade speed with memory consumption, and will slowdown the forward/backward.
+        Turn it on when you encounter OOM problem during training or evaluation.
     dropout : float, optional
         Dropout rate. Default: 0.0
     """
@@ -65,6 +69,7 @@ class RelGraphConv(nn.Module):
                  bias=True,
                  activation=None,
                  self_loop=False,
+                 low_mem=False,
                  dropout=0.0):
         super(RelGraphConv, self).__init__()
         self.in_feat = in_feat
@@ -77,6 +82,7 @@ class RelGraphConv(nn.Module):
         self.bias = bias
         self.activation = activation
         self.self_loop = self_loop
+        self.low_mem = low_mem
 
         if regularizer == "basis":
             # add basis weights
@@ -133,7 +139,22 @@ class RelGraphConv(nn.Module):
         else:
             weight = self.weight
 
+        # calculate msg @ W_r before put msg into edge
+        # if src is th.int64 we expect it is an index select
+        if edges.src['h'].dtype != th.int64 and self.low_mem:
+            etypes = th.unique(edges.data['type'])
+            msg = th.empty((edges.src['h'].shape[0], self.out_feat),
+                           device=edges.src['h'].device)
+            for etype in etypes:
+                loc = edges.data['type'] == etype
+                w = weight[etype]
+                src = edges.src['h'][loc]
+                sub_msg = th.matmul(src, w)
+                msg[loc] = sub_msg
+        else:
+            # put W_r into edges then do msg @ W_r
             msg = utils.bmm_maybe_select(edges.src['h'], weight, edges.data['type'])
+
         if 'norm' in edges.data:
             msg = msg * edges.data['norm']
         return {'msg': msg}
@@ -142,6 +163,20 @@ class RelGraphConv(nn.Module):
         """Message function for block-diagonal-decomposition regularizer"""
         if edges.src['h'].dtype == th.int64 and len(edges.src['h'].shape) == 1:
             raise TypeError('Block decomposition does not allow integer ID feature.')
+
+        # calculate msg @ W_r before put msg into edge
+        if self.low_mem:
+            etypes = th.unique(edges.data['type'])
+            msg = th.empty((edges.src['h'].shape[0], self.out_feat),
+                           device=edges.src['h'].device)
+            for etype in etypes:
+                loc = edges.data['type'] == etype
+                w = self.weight[etype].view(self.num_bases, self.submat_in, self.submat_out)
+                src = edges.src['h'][loc].view(-1, self.num_bases, self.submat_in)
+                sub_msg = th.einsum('abc,bcd->abd', src, w)
+                sub_msg = sub_msg.reshape(-1, self.out_feat)
+                msg[loc] = sub_msg
+        else:
             weight = self.weight.index_select(0, edges.data['type']).view(
                 -1, self.submat_in, self.submat_out)
             node = edges.src['h'].view(-1, 1, self.submat_in)

python/dgl/nn/tensorflow/conv/relgraphconv.py

@@ -53,6 +53,10 @@ class RelGraphConv(layers.Layer):
         Activation function. Default: None
     self_loop : bool, optional
         True to include self loop message. Default: False
+    low_mem : bool, optional
+        True to use low memory implementation of relation message passing function. Default: False
+        This option trade speed with memory consumption, and will slowdown the forward/backward.
+        Turn it on when you encounter OOM problem during training or evaluation.
     dropout : float, optional
         Dropout rate. Default: 0.0
     """
@@ -66,6 +70,7 @@ class RelGraphConv(layers.Layer):
                  bias=True,
                  activation=None,
                  self_loop=False,
+                 low_mem=False,
                  dropout=0.0):
         super(RelGraphConv, self).__init__()
         self.in_feat = in_feat
@@ -78,6 +83,7 @@ class RelGraphConv(layers.Layer):
         self.bias = bias
         self.activation = activation
         self.self_loop = self_loop
+        self.low_mem = low_mem
 
         xinit = tf.keras.initializers.glorot_uniform()
         zeroinit = tf.keras.initializers.zeros()
@@ -134,6 +140,20 @@ class RelGraphConv(layers.Layer):
         else:
             weight = self.weight
 
+        # calculate msg @ W_r before put msg into edge
+        # if src is th.int64 we expect it is an index select
+        if edges.src['h'].dtype != tf.int64 and self.low_mem:
+            etypes, _ = tf.unique(edges.data['type'])
+            msg = tf.zeros([edges.src['h'].shape[0], self.out_feat])
+            idx = tf.range(edges.src['h'].shape[0])
+            for etype in etypes:
+                loc = (edges.data['type'] == etype)
+                w = weight[etype]
+                src = tf.boolean_mask(edges.src['h'], loc)
+                sub_msg = tf.matmul(src, w)
+                indices = tf.reshape(tf.boolean_mask(idx, loc), (-1, 1))
+                msg = tf.tensor_scatter_nd_update(msg, indices, sub_msg)
+        else:
             msg = utils.bmm_maybe_select(
                 edges.src['h'], weight, edges.data['type'])
         if 'norm' in edges.data:
@@ -146,6 +166,24 @@ class RelGraphConv(layers.Layer):
                 len(edges.src['h'].shape) == 1):
             raise TypeError(
                 'Block decomposition does not allow integer ID feature.')
+
+        # calculate msg @ W_r before put msg into edge
+        # if src is th.int64 we expect it is an index select
+        if self.low_mem:
+            etypes, _ = tf.unique(edges.data['type'])
+            msg = tf.zeros([edges.src['h'].shape[0], self.out_feat])
+            idx = tf.range(edges.src['h'].shape[0])
+            for etype in etypes:
+                loc = (edges.data['type'] == etype)
+                w = tf.reshape(self.weight[etype],
+                               (self.num_bases, self.submat_in, self.submat_out))
+                src = tf.reshape(tf.boolean_mask(edges.src['h'], loc),
+                                 (-1, self.num_bases, self.submat_in))
+                sub_msg = tf.einsum('abc,bcd->abd', src, w)
+                sub_msg = tf.reshape(sub_msg, (-1, self.out_feat))
+                indices = tf.reshape(tf.boolean_mask(idx, loc), (-1, 1))
+                msg = tf.tensor_scatter_nd_update(msg, indices, sub_msg)
+        else:
             weight = tf.reshape(tf.gather(
                 self.weight, edges.data['type']), (-1, self.submat_in, self.submat_out))
             node = tf.reshape(edges.src['h'], (-1, 1, self.submat_in))

tests/pytorch/test_nn.py

@@ -373,38 +373,66 @@ def test_rgcn():
     O = 8
 
     rgc_basis = nn.RelGraphConv(I, O, R, "basis", B).to(ctx)
+    rgc_basis_low = nn.RelGraphConv(I, O, R, "basis", B, low_mem=True).to(ctx)
+    rgc_basis_low.weight = rgc_basis.weight
+    rgc_basis_low.w_comp = rgc_basis.w_comp
     h = th.randn((100, I)).to(ctx)
     r = th.tensor(etype).to(ctx)
     h_new = rgc_basis(g, h, r)
+    h_new_low = rgc_basis_low(g, h, r)
     assert list(h_new.shape) == [100, O]
+    assert list(h_new_low.shape) == [100, O]
+    assert F.allclose(h_new, h_new_low)
 
     rgc_bdd = nn.RelGraphConv(I, O, R, "bdd", B).to(ctx)
+    rgc_bdd_low = nn.RelGraphConv(I, O, R, "bdd", B, low_mem=True).to(ctx)
+    rgc_bdd_low.weight = rgc_bdd.weight
     h = th.randn((100, I)).to(ctx)
     r = th.tensor(etype).to(ctx)
     h_new = rgc_bdd(g, h, r)
+    h_new_low = rgc_bdd_low(g, h, r)
     assert list(h_new.shape) == [100, O]
+    assert list(h_new_low.shape) == [100, O]
+    assert F.allclose(h_new, h_new_low)
 
     # with norm
     norm = th.zeros((g.number_of_edges(), 1)).to(ctx)
 
     rgc_basis = nn.RelGraphConv(I, O, R, "basis", B).to(ctx)
+    rgc_basis_low = nn.RelGraphConv(I, O, R, "basis", B, low_mem=True).to(ctx)
+    rgc_basis_low.weight = rgc_basis.weight
+    rgc_basis_low.w_comp = rgc_basis.w_comp
     h = th.randn((100, I)).to(ctx)
     r = th.tensor(etype).to(ctx)
     h_new = rgc_basis(g, h, r, norm)
+    h_new_low = rgc_basis_low(g, h, r, norm)
     assert list(h_new.shape) == [100, O]
+    assert list(h_new_low.shape) == [100, O]
+    assert F.allclose(h_new, h_new_low)
 
     rgc_bdd = nn.RelGraphConv(I, O, R, "bdd", B).to(ctx)
+    rgc_bdd_low = nn.RelGraphConv(I, O, R, "bdd", B, low_mem=True).to(ctx)
+    rgc_bdd_low.weight = rgc_bdd.weight
     h = th.randn((100, I)).to(ctx)
     r = th.tensor(etype).to(ctx)
     h_new = rgc_bdd(g, h, r, norm)
+    h_new_low = rgc_bdd_low(g, h, r, norm)
     assert list(h_new.shape) == [100, O]
+    assert list(h_new_low.shape) == [100, O]
+    assert F.allclose(h_new, h_new_low)
 
     # id input
     rgc_basis = nn.RelGraphConv(I, O, R, "basis", B).to(ctx)
+    rgc_basis_low = nn.RelGraphConv(I, O, R, "basis", B, low_mem=True).to(ctx)
+    rgc_basis_low.weight = rgc_basis.weight
+    rgc_basis_low.w_comp = rgc_basis.w_comp
     h = th.randint(0, I, (100,)).to(ctx)
     r = th.tensor(etype).to(ctx)
     h_new = rgc_basis(g, h, r)
+    h_new_low = rgc_basis_low(g, h, r)
     assert list(h_new.shape) == [100, O]
+    assert list(h_new_low.shape) == [100, O]
+    assert F.allclose(h_new, h_new_low)
 
 def test_gat_conv():
     ctx = F.ctx()

tests/tensorflow/test_nn.py

@@ -272,38 +272,66 @@ def test_rgcn():
     O = 8
 
     rgc_basis = nn.RelGraphConv(I, O, R, "basis", B)
+    rgc_basis_low = nn.RelGraphConv(I, O, R, "basis", B, low_mem=True)
+    rgc_basis_low.weight = rgc_basis.weight
+    rgc_basis_low.w_comp = rgc_basis.w_comp
     h = tf.random.normal((100, I))
     r = tf.constant(etype)
     h_new = rgc_basis(g, h, r)
+    h_new_low = rgc_basis_low(g, h, r)
     assert list(h_new.shape) == [100, O]
+    assert list(h_new_low.shape) == [100, O]
+    assert F.allclose(h_new, h_new_low)
 
     rgc_bdd = nn.RelGraphConv(I, O, R, "bdd", B)
+    rgc_bdd_low = nn.RelGraphConv(I, O, R, "bdd", B, low_mem=True)
+    rgc_bdd_low.weight = rgc_bdd.weight
     h = tf.random.normal((100, I))
     r = tf.constant(etype)
     h_new = rgc_bdd(g, h, r)
+    h_new_low = rgc_bdd_low(g, h, r)
     assert list(h_new.shape) == [100, O]
+    assert list(h_new_low.shape) == [100, O]
+    assert F.allclose(h_new, h_new_low)
 
     # with norm
     norm = tf.zeros((g.number_of_edges(), 1))
 
     rgc_basis = nn.RelGraphConv(I, O, R, "basis", B)
+    rgc_basis_low = nn.RelGraphConv(I, O, R, "basis", B, low_mem=True)
+    rgc_basis_low.weight = rgc_basis.weight
+    rgc_basis_low.w_comp = rgc_basis.w_comp
     h = tf.random.normal((100, I))
     r = tf.constant(etype)
     h_new = rgc_basis(g, h, r, norm)
+    h_new_low = rgc_basis_low(g, h, r, norm)
     assert list(h_new.shape) == [100, O]
+    assert list(h_new_low.shape) == [100, O]
+    assert F.allclose(h_new, h_new_low)
 
     rgc_bdd = nn.RelGraphConv(I, O, R, "bdd", B)
+    rgc_bdd_low = nn.RelGraphConv(I, O, R, "bdd", B, low_mem=True)
+    rgc_bdd_low.weight = rgc_bdd.weight
     h = tf.random.normal((100, I))
     r = tf.constant(etype)
     h_new = rgc_bdd(g, h, r, norm)
+    h_new_low = rgc_bdd_low(g, h, r, norm)
     assert list(h_new.shape) == [100, O]
+    assert list(h_new_low.shape) == [100, O]
+    assert F.allclose(h_new, h_new_low)
 
     # id input
     rgc_basis = nn.RelGraphConv(I, O, R, "basis", B)
+    rgc_basis_low = nn.RelGraphConv(I, O, R, "basis", B, low_mem=True)
+    rgc_basis_low.weight = rgc_basis.weight
+    rgc_basis_low.w_comp = rgc_basis.w_comp
     h = tf.constant(np.random.randint(0, I, (100,)))
     r = tf.constant(etype)
     h_new = rgc_basis(g, h, r)
+    h_new_low = rgc_basis_low(g, h, r)
     assert list(h_new.shape) == [100, O]
+    assert list(h_new_low.shape) == [100, O]
+    assert F.allclose(h_new, h_new_low)
 
 def test_gat_conv():
     g = dgl.DGLGraph(sp.sparse.random(100, 100, density=0.1), readonly=True)