update ogb

ogb_examples/graphproppred/main_pgl.py

+# Copyright (c) 2020 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.
+"""test ogb
+"""
+import argparse
+
+import pgl
+import numpy as np
+import paddle.fluid as fluid
+from pgl.contrib.ogb.graphproppred.dataset_pgl import PglGraphPropPredDataset
+from pgl.utils import paddle_helper
+from ogb.graphproppred import Evaluator
+from pgl.contrib.ogb.graphproppred.mol_encoder import AtomEncoder, BondEncoder
+
+
+def train(exe, batch_size, graph_wrapper, train_program, splitted_idx, dataset,
+          evaluator, fetch_loss, fetch_pred):
+    """Train"""
+    graphs, labels = dataset[splitted_idx["train"]]
+    perm = np.arange(0, len(graphs))
+    np.random.shuffle(perm)
+    start_batch = 0
+    batch_no = 0
+    pred_output = np.zeros_like(labels, dtype="float32")
+    while start_batch < len(perm):
+        batch_index = perm[start_batch:start_batch + batch_size]
+        start_batch += batch_size
+        batch_graph = pgl.graph.MultiGraph(graphs[batch_index])
+        batch_label = labels[batch_index]
+        batch_valid = (batch_label == batch_label).astype("float32")
+        batch_label = np.nan_to_num(batch_label).astype("float32")
+        feed_dict = graph_wrapper.to_feed(batch_graph)
+        feed_dict["label"] = batch_label
+        feed_dict["weight"] = batch_valid
+        loss, pred = exe.run(train_program,
+                             feed=feed_dict,
+                             fetch_list=[fetch_loss, fetch_pred])
+        pred_output[batch_index] = pred
+        batch_no += 1
+    print("train", evaluator.eval({"y_true": labels, "y_pred": pred_output}))
+
+
+def evaluate(exe, batch_size, graph_wrapper, val_program, splitted_idx,
+             dataset, mode, evaluator, fetch_pred):
+    """Eval"""
+    graphs, labels = dataset[splitted_idx[mode]]
+    perm = np.arange(0, len(graphs))
+    start_batch = 0
+    batch_no = 0
+    pred_output = np.zeros_like(labels, dtype="float32")
+    while start_batch < len(perm):
+        batch_index = perm[start_batch:start_batch + batch_size]
+        start_batch += batch_size
+        batch_graph = pgl.graph.MultiGraph(graphs[batch_index])
+        feed_dict = graph_wrapper.to_feed(batch_graph)
+        pred = exe.run(val_program, feed=feed_dict, fetch_list=[fetch_pred])
+        pred_output[batch_index] = pred[0]
+        batch_no += 1
+    print(mode, evaluator.eval({"y_true": labels, "y_pred": pred_output}))
+
+
+def send_func(src_feat, dst_feat, edge_feat):
+    """Send"""
+    return src_feat["h"] + edge_feat["h"]
+
+
+class GNNModel(object):
+    """GNNModel"""
+
+    def __init__(self, name, emb_dim, num_task, num_layers):
+        self.num_task = num_task
+        self.emb_dim = emb_dim
+        self.num_layers = num_layers
+        self.name = name
+        self.atom_encoder = AtomEncoder(name=name, emb_dim=emb_dim)
+        self.bond_encoder = BondEncoder(name=name, emb_dim=emb_dim)
+
+    def forward(self, graph):
+        """foward"""
+        h_node = self.atom_encoder(graph.node_feat['feat'])
+        h_edge = self.bond_encoder(graph.edge_feat['feat'])
+        for layer in range(self.num_layers):
+            msg = graph.send(
+                send_func,
+                nfeat_list=[("h", h_node)],
+                efeat_list=[("h", h_edge)])
+            h_node = graph.recv(msg, 'sum') + h_node
+            h_node = fluid.layers.fc(h_node,
+                                     size=self.emb_dim,
+                                     name=self.name + '_%s' % layer,
+                                     act="relu")
+        graph_nodes = pgl.layers.graph_pooling(graph, h_node, "average")
+        graph_pred = fluid.layers.fc(graph_nodes, self.num_task, name="final")
+        return graph_pred
+
+
+def main():
+    """main
+    """
+    # Training settings
+    parser = argparse.ArgumentParser(description='Graph Dataset')
+    parser.add_argument(
+        '--epochs',
+        type=int,
+        default=100,
+        help='number of epochs to train (default: 100)')
+    parser.add_argument(
+        '--dataset',
+        type=str,
+        default="ogbg-mol-tox21",
+        help='dataset name (default: proteinfunc)')
+    args = parser.parse_args()
+
+    place = fluid.CPUPlace()  # Dataset too big to use GPU
+
+    ### automatic dataloading and splitting
+    dataset = PglGraphPropPredDataset(name=args.dataset)
+    splitted_idx = dataset.get_idx_split()
+
+    ### automatic evaluator. takes dataset name as input
+    evaluator = Evaluator(args.dataset)
+
+    graph_data, label = dataset[:2]
+    batch_graph = pgl.graph.MultiGraph(graph_data)
+    graph_data = batch_graph
+
+    train_program = fluid.Program()
+    startup_program = fluid.Program()
+    test_program = fluid.Program()
+    # degree normalize
+    graph_data.edge_feat["feat"] = graph_data.edge_feat["feat"].astype("int64")
+    graph_data.node_feat["feat"] = graph_data.node_feat["feat"].astype("int64")
+
+    model = GNNModel(
+        name="gnn", num_task=dataset.num_tasks, emb_dim=64, num_layers=2)
+
+    with fluid.program_guard(train_program, startup_program):
+        gw = pgl.graph_wrapper.GraphWrapper(
+            "graph",
+            place=place,
+            node_feat=graph_data.node_feat_info(),
+            edge_feat=graph_data.edge_feat_info())
+        pred = model.forward(gw)
+        sigmoid_pred = fluid.layers.sigmoid(pred)
+
+    val_program = train_program.clone(for_test=True)
+
+    initializer = []
+    with fluid.program_guard(train_program, startup_program):
+        train_label = fluid.layers.data(
+            name="label", dtype="float32", shape=[None, dataset.num_tasks])
+        train_weight = fluid.layers.data(
+            name="weight", dtype="float32", shape=[None, dataset.num_tasks])
+        train_loss_t = fluid.layers.sigmoid_cross_entropy_with_logits(
+            x=pred, label=train_label) * train_weight
+        train_loss_t = fluid.layers.reduce_sum(train_loss_t)
+
+        adam = fluid.optimizer.Adam(
+            learning_rate=1e-2,
+            regularization=fluid.regularizer.L2DecayRegularizer(
+                regularization_coeff=0.0005))
+        adam.minimize(train_loss_t)
+
+    exe = fluid.Executor(place)
+    exe.run(startup_program)
+
+    for epoch in range(1, args.epochs + 1):
+        print("Epoch", epoch)
+        train(exe, 128, gw, train_program, splitted_idx, dataset, evaluator,
+              train_loss_t, sigmoid_pred)
+        evaluate(exe, 128, gw, val_program, splitted_idx, dataset, "valid",
+                 evaluator, sigmoid_pred)
+        evaluate(exe, 128, gw, val_program, splitted_idx, dataset, "test",
+                 evaluator, sigmoid_pred)
+
+
+if __name__ == "__main__":
+    main()

ogb_examples/linkproppred/main_pgl.py

@@ -14,10 +14,13 @@
 """test ogb
 """
 import argparse
-
-import pgl
+import time
+import logging
 import numpy as np
+
 import paddle.fluid as fluid
+
+import pgl
 from pgl.contrib.ogb.linkproppred.dataset_pgl import PglLinkPropPredDataset
 from pgl.utils import paddle_helper
 from ogb.linkproppred import Evaluator
@@ -44,12 +47,12 @@ class GNNModel(object):
 
         self.src_nodes = fluid.layers.data(
             name='src_nodes',
-            shape=[None, 1],
+            shape=[None],
             dtype='int64', )
 
         self.dst_nodes = fluid.layers.data(
             name='dst_nodes',
-            shape=[None, 1],
+            shape=[None],
             dtype='int64', )
 
         self.edge_label = fluid.layers.data(
@@ -63,7 +66,6 @@ class GNNModel(object):
             shape=[self.num_nodes, self.emb_dim],
             dtype="float32",
             name=self.name + "_embedding")
-        #  edge_attr = fluid.layers.fc(graph.edge_feat["feat"], size=self.emb_dim)
 
         for layer in range(self.num_layers):
             msg = graph.send(
@@ -83,8 +85,8 @@ class GNNModel(object):
                 name=self.name + '_bias_%s' % layer)
             h = fluid.layers.elementwise_add(h, bias, act="relu")
 
-        src = fluid.layers.gather(h, self.src_nodes)
-        dst = fluid.layers.gather(h, self.dst_nodes)
+        src = fluid.layers.gather(h, self.src_nodes, overwrite=False)
+        dst = fluid.layers.gather(h, self.dst_nodes, overwrite=False)
         edge_embed = src * dst
         pred = fluid.layers.fc(input=edge_embed,
                                size=1,
@@ -107,17 +109,22 @@ def main():
     parser.add_argument(
         '--epochs',
         type=int,
-        default=100,
+        default=4,
         help='number of epochs to train (default: 100)')
     parser.add_argument(
         '--dataset',
         type=str,
         default="ogbl-ppa",
         help='dataset name (default: protein protein associations)')
+    parser.add_argument('--use_cuda', action='store_true')
+    parser.add_argument('--batch_size', type=int, default=5120)
+    parser.add_argument('--embed_dim', type=int, default=64)
+    parser.add_argument('--num_layers', type=int, default=2)
+    parser.add_argument('--lr', type=float, default=0.001)
     args = parser.parse_args()
+    print(args)
 
-    #place = fluid.CUDAPlace(0)
-    place = fluid.CPUPlace()  # Dataset too big to use GPU
+    place = fluid.CUDAPlace(0) if args.use_cuda else fluid.CPUPlace()
 
     ### automatic dataloading and splitting
     print("loadding dataset")
@@ -135,19 +142,20 @@ def main():
 
     train_program = fluid.Program()
     startup_program = fluid.Program()
-    test_program = fluid.Program()
+
     # degree normalize
     indegree = graph_data.indegree()
     norm = np.zeros_like(indegree, dtype="float32")
     norm[indegree > 0] = np.power(indegree[indegree > 0], -0.5)
     graph_data.node_feat["norm"] = np.expand_dims(norm, -1).astype("float32")
+    #  graph_data.node_feat["index"] = np.array([i for i in range(graph_data.num_nodes)], dtype=np.int64).reshape(-1,1)
 
     with fluid.program_guard(train_program, startup_program):
         model = GNNModel(
             name="gnn",
             num_nodes=graph_data.num_nodes,
-            emb_dim=64,
-            num_layers=2)
+            emb_dim=args.embed_dim,
+            num_layers=args.num_layers)
         gw = pgl.graph_wrapper.GraphWrapper(
             "graph",
             place,
@@ -158,50 +166,106 @@ def main():
     val_program = train_program.clone(for_test=True)
 
     with fluid.program_guard(train_program, startup_program):
+        global_steps = int(splitted_edge['train_edge'].shape[0] /
+                           args.batch_size * 2)
+        learning_rate = fluid.layers.polynomial_decay(args.lr, global_steps,
+                                                      0.00005)
+
         adam = fluid.optimizer.Adam(
-            learning_rate=1e-2,
+            learning_rate=learning_rate,
             regularization=fluid.regularizer.L2DecayRegularizer(
                 regularization_coeff=0.0005))
         adam.minimize(loss)
 
     exe = fluid.Executor(place)
     exe.run(startup_program)
-
     feed = gw.to_feed(graph_data)
+
+    print("evaluate result before training: ")
+    result = test(exe, val_program, prob, evaluator, feed, splitted_edge)
+    print(result)
+
+    print("training")
+    cc = 0
     for epoch in range(1, args.epochs + 1):
-        feed['src_nodes'] = splitted_edge["train_edge"][:, 0].reshape(-1, 1)
-        feed['dst_nodes'] = splitted_edge["train_edge"][:, 1].reshape(-1, 1)
-        feed['edge_label'] = splitted_edge["train_edge_label"].astype(
-            "float32").reshape(-1, 1)
-        res_loss, y_pred = exe.run(train_program,
-                                   feed=feed,
-                                   fetch_list=[loss, prob])
-        print("Loss %s" % res_loss[0])
-
-        result = {}
-        print("Evaluating...")
-        feed['src_nodes'] = splitted_edge["valid_edge"][:, 0].reshape(-1, 1)
-        feed['dst_nodes'] = splitted_edge["valid_edge"][:, 1].reshape(-1, 1)
-        feed['edge_label'] = splitted_edge["valid_edge_label"].astype(
-            "float32").reshape(-1, 1)
-        y_pred = exe.run(val_program, feed=feed, fetch_list=[prob])[0]
-        input_dict = {
-            "y_true": splitted_edge["valid_edge_label"],
-            "y_pred": y_pred.reshape(-1, ),
-        }
-        result["valid"] = evaluator.eval(input_dict)
-
-        feed['src_nodes'] = splitted_edge["test_edge"][:, 0].reshape(-1, 1)
-        feed['dst_nodes'] = splitted_edge["test_edge"][:, 1].reshape(-1, 1)
-        feed['edge_label'] = splitted_edge["test_edge_label"].astype(
-            "float32").reshape(-1, 1)
-        y_pred = exe.run(val_program, feed=feed, fetch_list=[prob])[0]
-        input_dict = {
-            "y_true": splitted_edge["test_edge_label"],
-            "y_pred": y_pred.reshape(-1, ),
-        }
-        result["test"] = evaluator.eval(input_dict)
-        print(result)
+        for batch_data, batch_label in data_generator(
+                graph_data,
+                splitted_edge["train_edge"],
+                splitted_edge["train_edge_label"],
+                batch_size=args.batch_size):
+            feed['src_nodes'] = batch_data[:, 0].reshape(-1, 1)
+            feed['dst_nodes'] = batch_data[:, 1].reshape(-1, 1)
+            feed['edge_label'] = batch_label.astype("float32")
+
+            res_loss, y_pred, b_lr = exe.run(
+                train_program,
+                feed=feed,
+                fetch_list=[loss, prob, learning_rate])
+            if cc % 1 == 0:
+                print("epoch %d | step %d | lr %s | Loss %s" %
+                      (epoch, cc, b_lr[0], res_loss[0]))
+            cc += 1
+
+            if cc % 20 == 0:
+                print("Evaluating...")
+                result = test(exe, val_program, prob, evaluator, feed,
+                              splitted_edge)
+                print("epoch %d | step %d" % (epoch, cc))
+                print(result)
+
+
+def test(exe, val_program, prob, evaluator, feed, splitted_edge):
+    """Evaluation"""
+    result = {}
+    feed['src_nodes'] = splitted_edge["valid_edge"][:, 0].reshape(-1, 1)
+    feed['dst_nodes'] = splitted_edge["valid_edge"][:, 1].reshape(-1, 1)
+    feed['edge_label'] = splitted_edge["valid_edge_label"].astype(
+        "float32").reshape(-1, 1)
+    y_pred = exe.run(val_program, feed=feed, fetch_list=[prob])[0]
+    input_dict = {
+        "y_true": splitted_edge["valid_edge_label"],
+        "y_pred": y_pred.reshape(-1, ),
+    }
+    result["valid"] = evaluator.eval(input_dict)
+
+    feed['src_nodes'] = splitted_edge["test_edge"][:, 0].reshape(-1, 1)
+    feed['dst_nodes'] = splitted_edge["test_edge"][:, 1].reshape(-1, 1)
+    feed['edge_label'] = splitted_edge["test_edge_label"].astype(
+        "float32").reshape(-1, 1)
+    y_pred = exe.run(val_program, feed=feed, fetch_list=[prob])[0]
+    input_dict = {
+        "y_true": splitted_edge["test_edge_label"],
+        "y_pred": y_pred.reshape(-1, ),
+    }
+    result["test"] = evaluator.eval(input_dict)
+    return result
+
+
+def data_generator(graph, data, label_data, batch_size, shuffle=True):
+    """Data Generator"""
+    perm = np.arange(0, len(data))
+    if shuffle:
+        np.random.shuffle(perm)
+
+    offset = 0
+    while offset < len(perm):
+        batch_index = perm[offset:(offset + batch_size)]
+        offset += batch_size
+        pos_data = data[batch_index]
+        pos_label = label_data[batch_index]
+
+        neg_src_node = pos_data[:, 0]
+        neg_dst_node = np.random.choice(
+            pos_data.reshape(-1, ), size=len(neg_src_node))
+        neg_data = np.hstack(
+            [neg_src_node.reshape(-1, 1), neg_dst_node.reshape(-1, 1)])
+        exists = graph.has_edges_between(neg_src_node, neg_dst_node)
+        neg_data = neg_data[np.invert(exists)]
+        neg_label = np.zeros(shape=len(neg_data), dtype=np.int64)
+
+        batch_data = np.vstack([pos_data, neg_data])
+        label = np.vstack([pos_label.reshape(-1, 1), neg_label.reshape(-1, 1)])
+        yield batch_data, label
 
 
 if __name__ == "__main__":

pgl/contrib/ogb/graphproppred/mol_encoder.py

+# Copyright (c) 2020 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.
+"""MolEncoder for ogb
+"""
+import paddle.fluid as fluid
+from ogb.utils.features import get_atom_feature_dims, get_bond_feature_dims
+
+
+class AtomEncoder(object):
+    """AtomEncoder for encoding node features"""
+
+    def __init__(self, name, emb_dim):
+        self.emb_dim = emb_dim
+        self.name = name
+
+    def __call__(self, x):
+        atom_feature = get_atom_feature_dims()
+        atom_input = fluid.layers.split(
+            x, num_or_sections=len(atom_feature), dim=-1)
+        outputs = None
+        count = 0
+        for _x, _atom_input_dim in zip(atom_input, atom_feature):
+            count += 1
+            emb = fluid.layers.embedding(
+                _x,
+                size=(_atom_input_dim, self.emb_dim),
+                param_attr=fluid.ParamAttr(
+                    name=self.name + '_atom_feat_%s' % count))
+            if outputs is None:
+                outputs = emb
+            else:
+                outputs = outputs + emb
+        return outputs
+
+
+class BondEncoder(object):
+    """Bond for encoding edge features"""
+
+    def __init__(self, name, emb_dim):
+        self.emb_dim = emb_dim
+        self.name = name
+
+    def __call__(self, x):
+        bond_feature = get_bond_feature_dims()
+        bond_input = fluid.layers.split(
+            x, num_or_sections=len(bond_feature), dim=-1)
+        outputs = None
+        count = 0
+        for _x, _bond_input_dim in zip(bond_input, bond_feature):
+            count += 1
+            emb = fluid.layers.embedding(
+                _x,
+                size=(_bond_input_dim, self.emb_dim),
+                param_attr=fluid.ParamAttr(
+                    name=self.name + '_bond_feat_%s' % count))
+            if outputs is None:
+                outputs = emb
+            else:
+                outputs = outputs + emb
+        return outputs

pgl/contrib/ogb/linkproppred/dataset_pgl.py

@@ -60,7 +60,7 @@ class PglLinkPropPredDataset(object):
         """pre_process downlaoding data
         """
         processed_dir = osp.join(self.root, 'processed')
-        pre_processed_file_path = osp.join(processed_dir, 'dgl_data_processed')
+        pre_processed_file_path = osp.join(processed_dir, 'pgl_data_processed')
 
         if osp.exists(pre_processed_file_path):
             #TODO: Reload Preprocess files

pgl/layers/__init__.py

@@ -18,7 +18,10 @@ from pgl.layers import conv
 from pgl.layers.conv import *
 from pgl.layers import set2set
 from pgl.layers.set2set import *
+from pgl.layers import graph_pool
+from pgl.layers.graph_pool import *
 
 __all__ = []
 __all__ += conv.__all__
 __all__ += set2set.__all__
+__all__ += graph_pool.__all__

pgl/layers/graph_pool.py

+# Copyright (c) 2019 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.
+"""This package implements common layers to help building
+graph neural networks.
+"""
+import paddle.fluid as fluid
+from pgl import graph_wrapper
+from pgl.utils import paddle_helper
+from pgl.utils import op
+
+__all__ = ['graph_pooling']
+
+
+def graph_pooling(gw, node_feat, pool_type):
+    """Implementation of graph pooling 
+
+    This is an implementation of graph pooling
+
+    Args:
+        gw: Graph wrapper object (:code:`StaticGraphWrapper` or :code:`GraphWrapper`)
+
+        node_feat: A tensor with shape (num_nodes, feature_size).
+
+        pool_type: The type of pooling ("sum", "average" , "min")
+
+    Return:
+        A tensor with shape (num_graph, hidden_size)
+    """
+    graph_feat = op.nested_lod_reset(node_feat, gw.graph_lod)
+    graph_feat = fluid.layers.sequence_pool(graph_feat, pool_type)
+    return graph_feat