add appnp, gat, gin

examples/gnn_depoly/models/__init__.py

@@ -12,6 +12,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from .appnp import APPNP
 from .gcn import GCN
 from .gat import GAT
+from .gin import GIN
 from .graphsage import GraphSage
+from .sgc import SGC

examples/gnn_depoly/models/appnp.py

@@ -0,0 +1,107 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import paddle
+import paddle.nn as nn
+import paddle.nn.functional as F
+
+from .base_conv import BaseConv
+
+
+def gcn_norm(edge_index, num_nodes):
+    _, col = edge_index[:, 0], edge_index[:, 1]
+    degree = paddle.zeros(shape=[num_nodes], dtype="int64")
+    degree = paddle.scatter(
+        x=degree,
+        index=col,
+        updates=paddle.ones_like(
+            col, dtype="int64"),
+        overwrite=False)
+    norm = paddle.cast(degree, dtype=paddle.get_default_dtype())
+    norm = paddle.clip(norm, min=1.0)
+    norm = paddle.pow(norm, -0.5)
+    norm = paddle.reshape(norm, [-1, 1])
+    return norm
+
+
+class APPNPConv(BaseConv):
+    def __init__(self, alpha=0.2, k_hop=10, self_loop=False):
+        super(APPNPConv, self).__init__()
+        self.alpha = alpha
+        self.k_hop = k_hop
+        self.self_loop = self_loop
+
+    def forward(self, edge_index, num_nodes, feature, norm=None):
+        if self.self_loop:
+            index = paddle.arange(start=0, end=num_nodes, dtype="int64")
+            self_loop_edges = paddle.transpose(paddle.stack((index, index)), [1, 0])
+
+            mask = edge_index[:, 0] != edge_index[:, 1]
+            mask_index = paddle.masked_select(paddle.arange(end=edge_index.shape[0]), mask)
+            edges = paddle.gather(edge_index, mask_index)  # remove self loop
+
+            edge_index = paddle.concat((self_loop_edges, edges), axis=0)
+        if norm is None:
+            norm = gcn_norm(edge_index, num_nodes)
+
+        h0 = feature
+
+        for _ in range(self.k_hop):
+            feature = feature * norm
+            feature = self.send_recv(edge_index, feature, "sum")
+            feature = feature * norm
+            feature = self.alpha * h0 + (1 - self.alpha) * feature
+
+        return feature
+
+
+class APPNP(nn.Layer):
+    """Implement of APPNP"""
+
+    def __init__(self,
+                 input_size,
+                 num_class,
+                 num_layers=1,
+                 hidden_size=64,
+                 dropout=0.5,
+                 k_hop=10,
+                 alpha=0.1,
+                 **kwargs):
+        super(APPNP, self).__init__()
+        self.num_class = num_class
+        self.num_layers = num_layers
+        self.hidden_size = hidden_size
+        self.dropout = dropout
+        self.alpha = alpha
+        self.k_hop = k_hop
+
+        self.mlps = nn.LayerList()
+        self.mlps.append(nn.Linear(input_size, self.hidden_size))
+        self.drop_fn = nn.Dropout(self.dropout)
+        for _ in range(self.num_layers - 1):
+            self.mlps.append(nn.Linear(self.hidden_size, self.hidden_size))
+
+        self.output = nn.Linear(self.hidden_size, num_class)
+        self.appnp = APPNPConv(alpha=self.alpha, k_hop=self.k_hop)
+
+    def forward(self, edge_index, num_nodes, feature):
+        for m in self.mlps:
+            feature = self.drop_fn(feature)
+            feature = m(feature)
+            feature = F.relu(feature)
+        feature = self.drop_fn(feature)
+        feature = self.output(feature)
+        feature = self.appnp(edge_index, num_nodes, feature)
+        return feature 

examples/gnn_depoly/models/base_conv.py

@@ -31,7 +31,7 @@ def __init__(self):
     def send_recv(self, edge_index, feature, pool_type="sum"):
         src, dst = edge_index[:, 0], edge_index[:, 1]
         return paddle.geometric.send_u_recv(
-            feature, src, dst, pool_type=pool_type)
+            feature, src, dst, reduce_op=pool_type)
 
     def send(
             self,

examples/gnn_depoly/models/gat.py

@@ -15,6 +15,7 @@
 import paddle
 import paddle.nn as nn
 import paddle.nn.functional as F
+import pgl
 
 from .base_conv import BaseConv
 

examples/gnn_depoly/models/gin.py

@@ -0,0 +1,102 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle
+import paddle.nn as nn
+import paddle.nn.functional as F
+from pgl.math import segment_pool
+
+from .base_conv import BaseConv
+
+
+class GINConv(BaseConv):
+    def __init__(
+        self, input_size, output_size, activation=None, init_eps=0.0, train_eps=False
+    ):
+        super(GINConv, self).__init__()
+        self.input_size = input_size
+        self.output_size = output_size
+        self.linear1 = nn.Linear(input_size, output_size, bias_attr=True)
+        self.linear2 = nn.Linear(output_size, output_size, bias_attr=True)
+        self.layer_norm = nn.LayerNorm(output_size)
+        if train_eps:
+            self.epsilon = self.create_parameter(
+                shape=[1, 1],
+                dtype="float32",
+                default_initializer=nn.initializer.Constant(value=init_eps),
+            )
+        else:
+            self.epsilon = init_eps
+
+        if isinstance(activation, str):
+            activation = getattr(F, activation)
+        self.activation = activation
+
+    def forward(self, edge_index, feature):
+        neigh_feature = self.send_recv(edge_index, feature, "sum")
+        output = neigh_feature + feature * (self.epsilon + 1.0)
+
+        output = self.linear1(output)
+        output = self.layer_norm(output)
+
+        if self.activation is not None:
+            output = self.activation(output)
+
+        output = self.linear2(output)
+
+        return output
+
+
+class GIN(nn.Layer):
+    def __init__(self, input_size, num_class, num_layers, hidden_size, pool_type="sum", dropout_prob=0.0):
+        super(GIN, self).__init__()
+        self.input_size = input_size
+        self.num_class = num_class
+        self.num_layers = num_layers
+        self.hidden_size = hidden_size
+        self.pool_type = pool_type
+        self.dropout_prob = dropout_prob
+
+        self.gin_convs = nn.LayerList()
+        self.norms = nn.LayerList()
+        self.linears = nn.LayerList()
+        self.linears.append(nn.Linear(self.input_size, self.num_class))
+
+        for i in range(self.num_layers):
+            if i == 0:
+                input_size = self.input_size
+            else:
+                input_size = self.hidden_size
+            gin = GINConv(input_size, self.hidden_size, "relu")
+            self.gin_convs.append(gin)
+            ln = paddle.nn.LayerNorm(self.hidden_size)
+            self.norms.append(ln)
+            self.linears.append(nn.Linear(self.hidden_size, self.num_class))
+        self.relu = nn.ReLU()
+
+
+    def forward(self, edge_index, feature):
+        feature_list = [feature]
+        for i in range(self.num_layers):
+            h = self.gin_convs[i](edge_index, feature_list[i])
+            h = self.norms[i](h)
+            h = self.relu(h)
+            feature_list.append(h)
+
+        output = 0
+        for i, h in enumerate(feature_list):
+            h = F.dropout(h, p=self.dropout_prob)
+            output += self.linears[i](h)
+
+        return output

examples/gnn_depoly/models/sgc.py

@@ -0,0 +1,92 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import paddle
+import paddle.nn as nn
+import paddle.nn.functional as F
+
+from .base_conv import BaseConv
+
+
+def gcn_norm(edge_index, num_nodes):
+    _, col = edge_index[:, 0], edge_index[:, 1]
+    degree = paddle.zeros(shape=[num_nodes], dtype="int64")
+    degree = paddle.scatter(
+        x=degree,
+        index=col,
+        updates=paddle.ones_like(
+            col, dtype="int64"),
+        overwrite=False)
+    norm = paddle.cast(degree, dtype=paddle.get_default_dtype())
+    norm = paddle.clip(norm, min=1.0)
+    norm = paddle.pow(norm, -0.5)
+    norm = paddle.reshape(norm, [-1, 1])
+    return norm
+
+
+class SGCConv(BaseConv):
+    def __init__(self, input_size, output_size, k_hop=2, cached=True, activation=None, bias=False):
+        super(SGCConv, self).__init__()
+        self.input_size = input_size
+        self.output_size = output_size
+        self.k_hop = k_hop
+        self.linear = nn.Linear(input_size, output_size, bias_attr=False)
+        if bias:
+            self.bias = self.create_parameter(shape=[output_size], is_bias=True)
+
+        self.cached = cached
+        self.cached_output = None
+        if isinstance(activation, str):
+            activation = getattr(F, activation)
+        self.activation = activation
+
+    def forward(self, edge_index, num_nodes, feature):
+        if self.cached:
+            if self.cached_output is None:
+                norm = gcn_norm(edge_index, num_nodes)
+                for hop in range(self.k_hop):
+                    feature = feature * norm
+                    feature = self.send_recv(edge_index, feature, "sum")
+                    feature = feature * norm
+                self.cached_output = feature
+            else:
+                feature = self.cached_output
+        else:
+            norm = gcn_norm(edge_index, num_nodes)
+            for hop in range(self.k_hop):
+                feature = feature * norm
+                feature = self.send_recv(edge_index, feature, "sum")
+                feature = feature * norm
+
+        output = self.linear(feature)
+        if hasattr(self, "bias"):
+            output = output + self.bias
+
+        if self.activation is not None:
+            output = self.activation(output)
+        return output
+
+
+class SGC(nn.Layer):
+    def __init__(self, input_size, num_class, num_layers=1, **kwargs):
+        super(SGC, self).__init__()
+        self.num_class = num_class
+        self.num_layers = num_layers
+        self.sgc_layer = SGCConv(
+            input_size=input_size, output_size=num_class, k_hop=num_layers)
+
+    def forward(self, edge_index, num_nodes, feature):
+        feature = self.sgc_layer(edge_index, num_nodes, feature)
+        return feature

examples/gnn_depoly/node_classification_with_full_graph/export_model.py

@@ -21,7 +21,7 @@
 import pgl
 
 sys.path.insert(0, os.path.abspath(".."))
-from models import GCN, GAT, GraphSage
+from models import GCN, GAT, GraphSage, GIN, SGC, APPNP 
 
 
 def save_static_model(args):
@@ -35,6 +35,21 @@ def save_static_model(args):
                     attn_drop=0.6,
                     num_heads=8,
                     hidden_size=8)
+    elif args.model == "APPNP":
+        model = APPNP(input_size=dataset.graph.node_feat["words"].shape[1],
+                      num_class=dataset.num_classes,
+                      num_layers=2,
+                      alpha=0.1,
+                      k_hop=10)
+    elif args.model == "SGC":
+        model = SGC(input_size=dataset.graph.node_feat["words"].shape[1],
+                    num_class=dataset.num_classes,
+                    num_layers=2)
+    elif args.model == "GIN":
+        model = GIN(input_size=dataset.graph.node_feat["words"].shape[1],
+                    num_class=dataset.num_classes,
+                    num_layers=2,
+                    hidden_size=16) 
     elif args.model == "GCN":
         model = GCN(input_size=dataset.graph.node_feat["words"].shape[1],
                     num_class=dataset.num_classes,
@@ -54,7 +69,7 @@ def save_static_model(args):
     model.eval()
 
     # Convert to static graph with specific input description
-    if args.model == "GraphSage":
+    if args.model == "GraphSage" or args.model == "GIN":
         model = paddle.jit.to_static(
             model,
             input_spec=[