Merge branch 'PaddlePaddle-main'

examples/SAGPool/main.py

@@ -124,8 +124,7 @@ def main(args, train_dataset, val_dataset, test_dataset):
             break
 
     correct = 0.
-    new_test_program = fluid.Program()
-    fluid.load(new_test_program, "./save/%s/%s" \
+    fluid.load(test_program, "./save/%s/%s" \
                % (args.dataset_name, args.save_model), exe)
     for feed_dict in test_loader:
         correct_ = exe.run(test_program,

examples/citation_benchmark/README.md

+# Easy Paper Reproduction for Citation Network (Cora/Pubmed/Citeseer)
+
+
+
+This page tries to reproduce all the **Graph Neural Network** paper for Citation Network (Cora/Pubmed/Citeseer), which is the **Hello world**  dataset (**small** and **fast**) for graph neural networks. But it's very hard to achieve very high performance.
+
+
+
+All datasets are runned with public split of  **semi-supervised** settings. And we report the averarge accuracy by running 10 times.
+
+
+
+# Experiment Results
+
+| Model                                                        | Cora         | Pubmed       | Citeseer     | Remarks                                                   |
+| ------------------------------------------------------------ | ------------ | ------------ | ------------ | --------------------------------------------------------- |
+| [Vanilla GCN (Kipf 2017)](https://openreview.net/pdf?id=SJU4ayYgl ) | 0.807(0.010) | 0.794(0.003) | 0.710(0.007) |                                                           |
+| [GAT (Veličković 2017)](https://arxiv.org/pdf/1710.10903.pdf) | 0.834(0.004) | 0.772(0.004) | 0.700(0.006) |                                                           |
+| [SGC(Wu 2019)](https://arxiv.org/pdf/1902.07153.pdf)         | 0.818(0.000) | 0.782(0.000) | 0.708(0.000) |                                                           |
+| [APPNP (Johannes 2018)](https://arxiv.org/abs/1810.05997)    | 0.846(0.003) | 0.803(0.002) | 0.719(0.003) | Almost the same with  the results reported in Appendix E. |
+| [GCNII (64 Layers, 1500 Epochs, Chen 2020)](https://arxiv.org/pdf/2007.02133.pdf) | 0.846(0.003) | 0.798(0.003) | 0.724(0.006) |                                                           |
+
+
+
+
+
+How to run the experiments?
+
+
+
+```shell
+# Device choose
+export CUDA_VISIBLE_DEVICES=0
+# GCN
+python train.py --conf config/gcn.yaml --use_cuda --dataset cora
+python train.py --conf config/gcn.yaml --use_cuda --dataset pubmed
+python train.py --conf config/gcn.yaml --use_cuda --dataset citeseer
+
+
+# GAT
+python train.py --conf config/gat.yaml --use_cuda --dataset cora
+python train.py --conf config/gat.yaml --use_cuda --dataset pubmed
+python train.py --conf config/gat.yaml --use_cuda --dataset citeseer
+
+
+# SGC (Slow version)
+python train.py --conf config/sgc.yaml --use_cuda --dataset cora
+python train.py --conf config/sgc.yaml --use_cuda --dataset pubmed
+python train.py --conf config/sgc.yaml --use_cuda --dataset citeseer
+
+# APPNP
+python train.py --conf config/appnp.yaml --use_cuda --dataset cora
+python train.py --conf config/appnp.yaml --use_cuda --dataset pubmed
+python train.py --conf config/appnp.yaml --use_cuda --dataset citeseer
+
+# GCNII (The original code use 1500 epochs.)
+python train.py --conf config/gcnii.yaml --use_cuda --dataset cora --epoch 1500
+python train.py --conf config/gcnii.yaml --use_cuda --dataset pubmed --epoch 1500
+python train.py --conf config/gcnii.yaml --use_cuda --dataset citeseer --epoch 1500
+```
+
+

examples/citation_benchmark/build_model.py

+import pgl
+import model
+from pgl import data_loader
+import paddle.fluid as fluid
+import numpy as np
+import time
+
+def build_model(dataset, config, phase, main_prog):
+    gw = pgl.graph_wrapper.GraphWrapper(
+            name="graph",
+            node_feat=dataset.graph.node_feat_info())
+
+    GraphModel = getattr(model, config.model_name)
+    m = GraphModel(config=config, num_class=dataset.num_classes) 
+    logits = m.forward(gw, gw.node_feat["words"], phase)
+
+    # Take the last
+    node_index = fluid.layers.data(
+            "node_index",
+            shape=[None, 1],
+            dtype="int64",
+            append_batch_size=False)
+    node_label = fluid.layers.data(
+            "node_label",
+            shape=[None, 1],
+            dtype="int64",
+            append_batch_size=False)
+
+    pred = fluid.layers.gather(logits, node_index)
+    loss, pred = fluid.layers.softmax_with_cross_entropy(
+        logits=pred, label=node_label, return_softmax=True)
+    acc = fluid.layers.accuracy(input=pred, label=node_label, k=1)
+    loss = fluid.layers.mean(loss)
+
+    if phase == "train":
+        adam = fluid.optimizer.Adam(
+            learning_rate=config.learning_rate,
+            regularization=fluid.regularizer.L2DecayRegularizer(
+                regularization_coeff=config.weight_decay))
+        adam.minimize(loss)
+    return gw, loss, acc
+
+

examples/citation_benchmark/config/appnp.yaml

+model_name: APPNP 
+k_hop: 10
+alpha: 0.1
+num_layer: 1
+learning_rate: 0.01
+dropout: 0.5
+hidden_size: 64
+weight_decay: 0.0005 
+edge_dropout: 0.0

examples/citation_benchmark/config/gat.yaml

+model_name: GAT 
+learning_rate: 0.005
+weight_decay: 0.0005 
+num_layers: 1
+feat_drop: 0.6
+attn_drop: 0.6
+num_heads: 8
+hidden_size: 8
+edge_dropout: 0.0

examples/citation_benchmark/config/gcn.yaml

+model_name: GCN 
+num_layers: 1
+dropout: 0.5
+hidden_size: 16 
+learning_rate: 0.01
+weight_decay: 0.0005 
+edge_dropout: 0.0

examples/citation_benchmark/config/gcnii.yaml

+model_name: GCNII 
+k_hop: 64 
+alpha: 0.1
+num_layer: 1
+learning_rate: 0.01
+dropout: 0.6
+hidden_size: 64
+weight_decay: 0.0005 
+edge_dropout: 0.0

examples/citation_benchmark/config/sgc.yaml

+model_name: SGC 
+num_layers: 2
+learning_rate: 0.2
+weight_decay: 0.000005 
+feature_pre_normalize: False

examples/citation_benchmark/model.py

+import pgl
+import paddle.fluid.layers as L
+import pgl.layers.conv as conv
+
+def get_norm(indegree):
+    float_degree = L.cast(indegree, dtype="float32")
+    float_degree = L.clamp(float_degree, min=1.0)
+    norm = L.pow(float_degree, factor=-0.5) 
+    return norm
+    
+
+class GCN(object):
+    """Implement of GCN
+    """
+    def __init__(self, config, num_class):
+        self.num_class = num_class
+        self.num_layers = config.get("num_layers", 1)
+        self.hidden_size = config.get("hidden_size", 64)
+        self.dropout = config.get("dropout", 0.5)
+        self.edge_dropout = config.get("edge_dropout", 0.0)
+
+    def forward(self, graph_wrapper, feature, phase):
+        
+        for i in range(self.num_layers):
+
+            if phase == "train":
+                ngw = pgl.sample.edge_drop(graph_wrapper, self.edge_dropout) 
+                norm = get_norm(ngw.indegree())
+            else:
+                ngw = graph_wrapper
+                norm = graph_wrapper.node_feat["norm"]
+
+
+            feature = pgl.layers.gcn(ngw,
+                feature,
+                self.hidden_size,
+                activation="relu",
+                norm=norm,
+                name="layer_%s" % i)
+
+            feature = L.dropout(
+                    feature,
+                    self.dropout,
+                    dropout_implementation='upscale_in_train')
+
+        if phase == "train": 
+            ngw = pgl.sample.edge_drop(graph_wrapper, self.edge_dropout) 
+            norm = get_norm(ngw.indegree())
+        else:
+            ngw = graph_wrapper
+            norm = graph_wrapper.node_feat["norm"]
+
+        feature = conv.gcn(ngw,
+                     feature,
+                     self.num_class,
+                     activation=None,
+                     norm=norm,
+                     name="output")
+
+        return feature
+
+class GAT(object):
+    """Implement of GAT"""
+    def __init__(self, config, num_class):
+        self.num_class = num_class 
+        self.num_layers = config.get("num_layers", 1)
+        self.num_heads = config.get("num_heads", 8)
+        self.hidden_size = config.get("hidden_size", 8)
+        self.feat_dropout = config.get("feat_drop", 0.6)
+        self.attn_dropout = config.get("attn_drop", 0.6)
+        self.edge_dropout = config.get("edge_dropout", 0.0)
+
+    def forward(self, graph_wrapper, feature, phase):
+        if phase == "train": 
+            edge_dropout = self.edge_dropout
+        else:
+            edge_dropout = 0
+
+        for i in range(self.num_layers):
+            ngw = pgl.sample.edge_drop(graph_wrapper, edge_dropout) 
+                
+            feature = conv.gat(ngw,
+                                feature,
+                                self.hidden_size,
+                                activation="elu",
+                                name="gat_layer_%s" % i,
+                                num_heads=self.num_heads,
+                                feat_drop=self.feat_dropout,
+                                attn_drop=self.attn_dropout)
+
+        ngw = pgl.sample.edge_drop(graph_wrapper, edge_dropout) 
+        feature = conv.gat(ngw,
+                     feature,
+                     self.num_class,
+                     num_heads=1,
+                     activation=None,
+                     feat_drop=self.feat_dropout,
+                     attn_drop=self.attn_dropout,
+                     name="output")
+        return feature
+
+   
+class APPNP(object):
+    """Implement of APPNP"""
+    def __init__(self, config, num_class):
+        self.num_class = num_class
+        self.num_layers = config.get("num_layers", 1)
+        self.hidden_size = config.get("hidden_size", 64)
+        self.dropout = config.get("dropout", 0.5)
+        self.alpha = config.get("alpha", 0.1)
+        self.k_hop = config.get("k_hop", 10)
+        self.edge_dropout = config.get("edge_dropout", 0.0)
+
+    def forward(self, graph_wrapper, feature, phase):
+        if phase == "train": 
+            edge_dropout = self.edge_dropout
+        else:
+            edge_dropout = 0
+
+        for i in range(self.num_layers):
+            feature = L.dropout(
+                feature,
+                self.dropout,
+                dropout_implementation='upscale_in_train')
+            feature = L.fc(feature, self.hidden_size, act="relu", name="lin%s" % i)
+
+        feature = L.dropout(
+            feature,
+            self.dropout,
+            dropout_implementation='upscale_in_train')
+
+        feature = L.fc(feature, self.num_class, act=None, name="output")
+
+        feature = conv.appnp(graph_wrapper,
+            feature=feature,
+            edge_dropout=edge_dropout,
+            alpha=self.alpha,
+            k_hop=self.k_hop)
+        return feature
+
+class SGC(object):
+    """Implement of SGC"""
+    def __init__(self, config, num_class):
+        self.num_class = num_class
+        self.num_layers = config.get("num_layers", 1)
+
+    def forward(self, graph_wrapper, feature, phase):
+        feature = conv.appnp(graph_wrapper,
+            feature=feature,
+            edge_dropout=0,
+            alpha=0,
+            k_hop=self.num_layers)
+        feature.stop_gradient=True
+        feature = L.fc(feature, self.num_class, act=None, bias_attr=False, name="output")
+        return feature
+
+ 
+class GCNII(object):
+    """Implement of GCNII"""
+    def __init__(self, config, num_class):
+        self.num_class = num_class
+        self.num_layers = config.get("num_layers", 1)
+        self.hidden_size = config.get("hidden_size", 64)
+        self.dropout = config.get("dropout", 0.6)
+        self.alpha = config.get("alpha", 0.1)
+        self.lambda_l = config.get("lambda_l", 0.5)
+        self.k_hop = config.get("k_hop", 64)
+        self.edge_dropout = config.get("edge_dropout", 0.0)
+
+    def forward(self, graph_wrapper, feature, phase):
+        if phase == "train": 
+            edge_dropout = self.edge_dropout
+        else:
+            edge_dropout = 0
+
+        for i in range(self.num_layers):
+            feature = L.fc(feature, self.hidden_size, act="relu", name="lin%s" % i)
+            feature = L.dropout(
+                feature,
+                self.dropout,
+                dropout_implementation='upscale_in_train')
+
+        feature = conv.gcnii(graph_wrapper,
+            feature=feature,
+            name="gcnii",
+            activation="relu",
+            lambda_l=self.lambda_l,
+            alpha=self.alpha,
+            dropout=self.dropout,
+            k_hop=self.k_hop)
+
+        feature = L.fc(feature, self.num_class, act=None, name="output")
+        return feature

examples/citation_benchmark/train.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.
+import pgl
+import model# import LabelGraphGCN
+from pgl import data_loader
+from pgl.utils.logger import log
+import paddle.fluid as fluid
+import numpy as np
+import time
+import argparse
+from build_model import build_model
+import yaml
+from easydict import EasyDict as edict
+import tqdm
+
+def normalize(feat):
+    return feat / np.maximum(np.sum(feat, -1, keepdims=True), 1)
+
+
+def load(name, normalized_feature=True):
+    if name == 'cora':
+        dataset = data_loader.CoraDataset()
+    elif name == "pubmed":
+        dataset = data_loader.CitationDataset("pubmed", symmetry_edges=True)
+    elif name == "citeseer":
+        dataset = data_loader.CitationDataset("citeseer", symmetry_edges=True)
+    else:
+        raise ValueError(name + " dataset doesn't exists")
+
+    indegree = dataset.graph.indegree()
+    norm = np.maximum(indegree.astype("float32"), 1)
+    norm = np.power(norm, -0.5)
+    dataset.graph.node_feat["norm"] = np.expand_dims(norm, -1)
+    dataset.graph.node_feat["words"] =  normalize(dataset.graph.node_feat["words"])
+    return dataset
+
+
+def main(args, config):
+    dataset = load(args.dataset, args.feature_pre_normalize)
+    place = fluid.CUDAPlace(0) if args.use_cuda else fluid.CPUPlace()
+    train_program = fluid.default_main_program()
+    startup_program = fluid.default_startup_program()
+    with fluid.program_guard(train_program, startup_program):
+        with fluid.unique_name.guard():
+            gw, loss, acc = build_model(dataset,
+                                config=config,
+                                phase="train",
+                                main_prog=train_program)
+
+    test_program = fluid.Program()
+    with fluid.program_guard(test_program, startup_program):
+        with fluid.unique_name.guard():
+            _gw, v_loss, v_acc = build_model(dataset,
+                config=config,
+                phase="test",
+                main_prog=test_program)
+
+    test_program = test_program.clone(for_test=True)
+
+    exe = fluid.Executor(place)
+
+
+    train_index = dataset.train_index
+    train_label = np.expand_dims(dataset.y[train_index], -1)
+    train_index = np.expand_dims(train_index, -1)
+
+    val_index = dataset.val_index
+    val_label = np.expand_dims(dataset.y[val_index], -1)
+    val_index = np.expand_dims(val_index, -1)
+
+    test_index = dataset.test_index
+    test_label = np.expand_dims(dataset.y[test_index], -1)
+    test_index = np.expand_dims(test_index, -1)
+
+    dur = []
+
+    # Feed data
+    feed_dict = gw.to_feed(dataset.graph)
+
+
+    best_test = []
+ 
+    for run in range(args.runs):
+        exe.run(startup_program)
+        cal_val_acc = []
+        cal_test_acc = []
+        cal_val_loss = []
+        cal_test_loss = []
+        for epoch in tqdm.tqdm(range(args.epoch)):
+            feed_dict["node_index"] = np.array(train_index, dtype="int64")
+            feed_dict["node_label"] = np.array(train_label, dtype="int64")
+            train_loss, train_acc = exe.run(train_program,
+                                        feed=feed_dict,
+                                        fetch_list=[loss, acc],
+                                        return_numpy=True)
+
+
+            feed_dict["node_index"] = np.array(val_index, dtype="int64")
+            feed_dict["node_label"] = np.array(val_label, dtype="int64")
+            val_loss, val_acc = exe.run(test_program,
+                                    feed=feed_dict,
+                                    fetch_list=[v_loss, v_acc],
+                                    return_numpy=True)
+
+            cal_val_acc.append(val_acc[0])
+            cal_val_loss.append(val_loss[0])
+
+            feed_dict["node_index"] = np.array(test_index, dtype="int64")
+            feed_dict["node_label"] = np.array(test_label, dtype="int64")
+            test_loss, test_acc = exe.run(test_program,
+                                  feed=feed_dict,
+                                  fetch_list=[v_loss, v_acc],
+                                  return_numpy=True)
+
+            cal_test_acc.append(test_acc[0])
+            cal_test_loss.append(test_loss[0])
+
+     
+        log.info("Runs %s: Model: %s Best Test Accuracy: %f" % (run, config.model_name,
+                  cal_test_acc[np.argmin(cal_val_loss)]))
+
+        best_test.append(cal_test_acc[np.argmin(cal_val_loss)])
+    log.info("Dataset: %s Best Test Accuracy: %f ( stddev: %f )" % (args.dataset, np.mean(best_test), np.std(best_test)))
+    print("Dataset: %s Best Test Accuracy: %f ( stddev: %f )" % (args.dataset, np.mean(best_test), np.std(best_test)))
+    
+    
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='Benchmarking Citation Network')
+    parser.add_argument(
+        "--dataset", type=str, default="cora", help="dataset (cora, pubmed)")
+    parser.add_argument("--use_cuda", action='store_true', help="use_cuda")
+    parser.add_argument("--conf", type=str, help="config file for models")
+    parser.add_argument("--epoch", type=int, default=200, help="Epoch")
+    parser.add_argument("--runs", type=int, default=10, help="runs")
+    parser.add_argument("--feature_pre_normalize", type=bool, default=True, help="pre_normalize feature")
+    args = parser.parse_args()
+    config = edict(yaml.load(open(args.conf), Loader=yaml.FullLoader))
+    log.info(args)
+    main(args, config)

examples/erniesage/README.en.md

@@ -49,6 +49,8 @@ sh local_run.sh config/enriesage_v1_gpu.yaml
 sh local_run.sh config/enriesage_v1_cpu.yaml
 ```
 
+**NOTE**: To help users better understand the ERNIESage Model, we provide a running example in Baidu AIStudio. Please visit here: https://aistudio.baidu.com/aistudio/projectdetail/667443. 
+
 ## Hyperparamters
 
 - learner_type: `gpu` or `cpu`; gpu use fleet Collective mode, cpu use fleet Transpiler mode.

examples/erniesage/README.md

@@ -50,6 +50,8 @@ sh local_run.sh config/erniesage_v2_gpu.yaml
 sh local_run.sh config/erniesage_v2_cpu.yaml
 ```
 
+**NOTE**：为了方便用户们学习使用ERNIESage，我们在百度AIStudio中提供了可以直接运行的ERNIESage实例，详情可见：https://aistudio.baidu.com/aistudio/projectdetail/667443.
+
 ## Hyperparamters
 
 - learner_type: `gpu` or `cpu`; gpu 使用fleet Collective 模式, cpu 使用fleet Transpiler 模式.

examples/erniesage/config/erniesage_v1_gpu.yaml → examples/erniesage/config/erniesage_link_predict.yaml

 # Global Enviroment Settings 
 #
 # trainer config ------
+task: "link_predict"
 learner_type: "gpu"
 optimizer_type: "adam"
 lr: 0.00005
 batch_size: 32
 CPU_NUM: 10
-epoch: 20
+epoch: 3
 log_per_step: 1
-save_per_step: 100
+save_per_step: 1000
 output_path: "./output"
 ckpt_path: "./ernie_base_ckpt"
 
 # data config ------
-input_data: "./data.txt"
-graph_path: "./workdir"
+train_data: "./example_data/link_predict/graph_data.txt"
+graph_data: "./example_data/link_predict/train_data.txt"
+
+graph_work_path: "./workdir"
 sample_workers: 1
 use_pyreader: true
 input_type: "text"
 
 # model config ------
 samples: [10]
-model_type: "ErnieSageModelV1"
+model_type: "ERNIESageV2"
 layer_type: "graphsage_sum"
 
 max_seqlen: 40
@@ -30,8 +33,9 @@ num_layers: 1
 hidden_size: 128
 final_fc: true
 final_l2_norm: true
-loss_type: "hinge"
-margin: 0.3
+loss_type: "global_hinge"
+margin: 0.1
+neg_type: "batch_neg"
 
 # infer config ------
 infer_model: "./output/last"

examples/erniesage/config/erniesage_v2_gpu.yaml → examples/erniesage/config/erniesage_node_classfication.yaml

 # Global Enviroment Settings 
 #
 # trainer config ------
+task: "node_classification"
 learner_type: "gpu"
 optimizer_type: "adam"
 lr: 0.00005
 batch_size: 32
 CPU_NUM: 10
 epoch: 3
-log_per_step: 10
+log_per_step: 1
 save_per_step: 1000
 output_path: "./output"
 ckpt_path: "./ernie_base_ckpt"
 
 # data config ------
-input_data: "./data.txt"
-graph_path: "./workdir"
+graph_data: "./example_data/node_classification/graph_data.txt"
+train_data: "./example_data/node_classification/train_data.txt"
+
+graph_work_path: "./workdir"
 sample_workers: 1
 use_pyreader: true
 input_type: "text"
 
 # model config ------
+num_label: 10
 samples: [10]
-model_type: "ErnieSageModelV2"
+model_type: "ERNIESageV2"
+layer_type: "graphsage_sum"
 
 max_seqlen: 40
 
@@ -29,8 +34,8 @@ num_layers: 1
 hidden_size: 128
 final_fc: true
 final_l2_norm: true
-loss_type: "hinge"
-margin: 0.3
+loss_type: "softmax_with_cross_entropy"
+margin: 0.1
 neg_type: "batch_neg"
 
 # infer config ------
@@ -49,7 +54,7 @@ ernie_config:
     max_position_embeddings: 513
     num_attention_heads: 12
     num_hidden_layers: 12
-    sent_type_vocab_size: 2
+    sent_type_vocab_size: 4
     task_type_vocab_size: 3
     vocab_size: 18000
     use_task_id: false

examples/erniesage/config/erniesage_v1_cpu.yaml

-# Global Enviroment Settings 
-#
-# trainer config ------
-learner_type: "cpu"
-optimizer_type: "adam"
-lr: 0.00005
-batch_size: 2
-CPU_NUM: 10
-epoch: 20
-log_per_step: 1
-save_per_step: 100
-output_path: "./output"
-ckpt_path: "./ernie_base_ckpt"
-
-# data config ------
-input_data: "./data.txt"
-graph_path: "./workdir"
-sample_workers: 1
-use_pyreader: true
-input_type: "text"
-
-# model config ------
-samples: [10]
-model_type: "ErnieSageModelV1"
-layer_type: "graphsage_sum"
-
-max_seqlen: 40
-
-num_layers: 1
-hidden_size: 128
-final_fc: true
-final_l2_norm: true
-loss_type: "hinge"
-margin: 0.3
-
-# infer config ------
-infer_model: "./output/last"
-infer_batch_size: 128
-
-# ernie config ------
-encoding: "utf8"
-ernie_vocab_file: "./vocab.txt"
-ernie_config:
-    attention_probs_dropout_prob: 0.1
-    hidden_act: "relu"
-    hidden_dropout_prob: 0.1
-    hidden_size: 768
-    initializer_range: 0.02
-    max_position_embeddings: 513
-    num_attention_heads: 12
-    num_hidden_layers: 12
-    sent_type_vocab_size: 4
-    task_type_vocab_size: 3
-    vocab_size: 18000
-    use_task_id: false
-    use_fp16: false

examples/erniesage/config/erniesage_v2_cpu.yaml

-# Global Enviroment Settings 
-#
-# trainer config ------
-learner_type: "cpu"
-optimizer_type: "adam"
-lr: 0.00005
-batch_size: 4
-CPU_NUM: 16
-epoch: 3
-log_per_step: 1
-save_per_step: 100
-output_path: "./output"
-ckpt_path: "./ernie_base_ckpt"
-
-# data config ------
-input_data: "./data.txt"
-graph_path: "./workdir"
-sample_workers: 1
-use_pyreader: true
-input_type: "text"
-
-# model config ------
-samples: [10]
-model_type: "ErnieSageModelV2"
-
-max_seqlen: 40
-
-num_layers: 1
-hidden_size: 128
-final_fc: true
-final_l2_norm: true
-loss_type: "hinge"
-margin: 0.3
-neg_type: "batch_neg"
-
-# infer config ------
-infer_model: "./output/last"
-infer_batch_size: 128
-
-# ernie config ------
-encoding: "utf8"
-ernie_vocab_file: "./vocab.txt"
-ernie_config:
-    attention_probs_dropout_prob: 0.1
-    hidden_act: "relu"
-    hidden_dropout_prob: 0.1
-    hidden_size: 768
-    initializer_range: 0.02
-    max_position_embeddings: 513
-    num_attention_heads: 12
-    num_hidden_layers: 12
-    sent_type_vocab_size: 2
-    task_type_vocab_size: 3
-    vocab_size: 18000
-    use_task_id: false
-    use_fp16: false

examples/erniesage/config/erniesage_v3_cpu.yaml

-# Global Enviroment Settings 
-#
-# trainer config ------
-learner_type: "cpu"
-optimizer_type: "adam"
-lr: 0.00005
-batch_size: 2
-CPU_NUM: 10
-epoch: 20
-log_per_step: 1
-save_per_step: 100
-output_path: "./output"
-ckpt_path: "./ernie_base_ckpt"
-
-# data config ------
-input_data: "./data.txt"
-graph_path: "./workdir"
-sample_workers: 1
-use_pyreader: true
-input_type: "text"
-
-# model config ------
-samples: [10]
-model_type: "ErnieSageModelV3"
-
-max_seqlen: 40
-
-num_layers: 1
-hidden_size: 128
-final_fc: true
-final_l2_norm: true
-loss_type: "hinge"
-margin: 0.3
-
-# infer config ------
-infer_model: "./output/last"
-infer_batch_size: 128
-
-# ernie config ------
-encoding: "utf8"
-ernie_vocab_file: "./vocab.txt"
-ernie_config:
-    attention_probs_dropout_prob: 0.1
-    hidden_act: "relu"
-    hidden_dropout_prob: 0.1
-    hidden_size: 768
-    initializer_range: 0.02
-    max_position_embeddings: 513
-    num_attention_heads: 12
-    num_hidden_layers: 12
-    sent_type_vocab_size: 4
-    task_type_vocab_size: 3
-    vocab_size: 18000
-    use_task_id: false
-    use_fp16: false

examples/erniesage/config/erniesage_v3_gpu.yaml

-# Global Enviroment Settings 
-#
-# trainer config ------
-learner_type: "gpu"
-optimizer_type: "adam"
-lr: 0.00005
-batch_size: 32
-CPU_NUM: 10
-epoch: 20
-log_per_step: 1
-save_per_step: 100
-output_path: "./output"
-ckpt_path: "./ernie_base_ckpt"
-
-# data config ------
-input_data: "./data.txt"
-graph_path: "./workdir"
-sample_workers: 1
-use_pyreader: true
-input_type: "text"
-
-# model config ------
-samples: [10]
-model_type: "ErnieSageModelV3"
-
-max_seqlen: 40
-
-num_layers: 1
-hidden_size: 128
-final_fc: true
-final_l2_norm: true
-loss_type: "hinge"
-margin: 0.3
-
-# infer config ------
-infer_model: "./output/last"
-infer_batch_size: 128
-
-# ernie config ------
-encoding: "utf8"
-ernie_vocab_file: "./vocab.txt"
-ernie_config:
-    attention_probs_dropout_prob: 0.1
-    hidden_act: "relu"
-    hidden_dropout_prob: 0.1
-    hidden_size: 768
-    initializer_range: 0.02
-    max_position_embeddings: 513
-    num_attention_heads: 12
-    num_hidden_layers: 12
-    sent_type_vocab_size: 4
-    task_type_vocab_size: 3
-    vocab_size: 18000
-    use_task_id: false
-    use_fp16: false

examples/erniesage/dataset/graph_reader.py

@@ -74,17 +74,15 @@ class GraphGenerator(BaseDataGenerator):
         batch_dst = np.array(batch_dst, dtype="int64")
 
         if self.neg_type == "batch_neg":
-            neg_shape = [1]
+            batch_neg = batch_dst
         else:
+            # TODO user define shape of neg_sample
             neg_shape = batch_dst.shape
-        sampled_batch_neg = alias_sample(neg_shape, self.alias, self.events)
-    
-        if len(batch_neg) > 0:
+            sampled_batch_neg = alias_sample(neg_shape, self.alias, self.events)
             batch_neg = np.concatenate([batch_neg, sampled_batch_neg], 0)
-        else:
-            batch_neg = sampled_batch_neg
 
         if self.phase == "train":
+            # TODO user define ignore edges or not
             #ignore_edges = np.concatenate([np.stack([batch_src, batch_dst], 1), np.stack([batch_dst, batch_src], 1)], 0)
             ignore_edges = set()
         else:
@@ -92,7 +90,8 @@ class GraphGenerator(BaseDataGenerator):
 
         nodes = np.unique(np.concatenate([batch_src, batch_dst, batch_neg], 0))
         subgraphs = graphsage_sample(self.graph, nodes, self.samples, ignore_edges=ignore_edges)
-        #subgraphs[0].reindex_to_parrent_nodes(subgraphs[0].nodes)
+        subgraphs[0].node_feat["index"] = subgraphs[0].reindex_to_parrent_nodes(subgraphs[0].nodes).astype(np.int64)
+        subgraphs[0].node_feat["term_ids"] = self.term_ids[subgraphs[0].node_feat["index"]].astype(np.int64)
         feed_dict = {}
         for i in range(self.num_layers):
             feed_dict.update(self.graph_wrappers[i].to_feed(subgraphs[i]))
@@ -103,9 +102,11 @@ class GraphGenerator(BaseDataGenerator):
         sub_neg_idx = subgraphs[0].reindex_from_parrent_nodes(batch_neg)
 
         feed_dict["user_index"] = np.array(sub_src_idx, dtype="int64")
-        feed_dict["item_index"] = np.array(sub_dst_idx, dtype="int64")
+        feed_dict["pos_item_index"] = np.array(sub_dst_idx, dtype="int64")
         feed_dict["neg_item_index"] = np.array(sub_neg_idx, dtype="int64")
-        feed_dict["term_ids"] = self.term_ids[subgraphs[0].node_feat["index"]].astype(np.int64)
+
+        feed_dict["user_real_index"] = np.array(batch_src, dtype="int64")
+        feed_dict["pos_item_real_index"] = np.array(batch_dst, dtype="int64")
         return feed_dict
 
     def __call__(self):
@@ -124,3 +125,37 @@ class GraphGenerator(BaseDataGenerator):
  
 
     
+class NodeClassificationGenerator(GraphGenerator):
+    def batch_fn(self, batch_ex):
+        # batch_ex = [
+        #     (node, label),
+        #     (node, label),
+        #     ]
+        #
+        batch_node = []
+        batch_label = []
+        for batch in batch_ex:
+            batch_node.append(batch[0])
+            batch_label.append(batch[1])
+
+        if len(batch_node) != self.batch_size:
+            if self.phase == "train":
+                return None  #Skip
+
+        batch_node = np.array(batch_node, dtype="int64")
+        batch_label = np.array(batch_label, dtype="int64")
+
+        subgraphs = graphsage_sample(self.graph, batch_node, self.samples)
+        subgraphs[0].node_feat["index"] = subgraphs[0].reindex_to_parrent_nodes(subgraphs[0].nodes).astype(np.int64)
+        subgraphs[0].node_feat["term_ids"] = self.term_ids[subgraphs[0].node_feat["index"]].astype(np.int64)
+        feed_dict = {}
+        for i in range(self.num_layers):
+            feed_dict.update(self.graph_wrappers[i].to_feed(subgraphs[i]))
+
+        # only reindex from first subgraph
+        sub_node_idx = subgraphs[0].reindex_from_parrent_nodes(batch_node)
+
+        feed_dict["node_index"] = np.array(sub_node_idx, dtype="int64")
+        feed_dict["node_real_index"] = np.array(batch_node, dtype="int64")
+        feed_dict["label"] = np.array(batch_label, dtype="int64")
+        return feed_dict

examples/erniesage/infer.py → examples/erniesage/link_predict_infer.py

@@ -34,7 +34,7 @@ from pgl.utils import paddle_helper
 import paddle
 import paddle.fluid as F
 
-from models.model_factory import Model
+from models.model import LinkPredictModel
 from dataset.graph_reader import GraphGenerator 
 
 
@@ -59,7 +59,7 @@ def run_predict(py_reader,
               log_per_step=1,
               args=None):
 
-    id2str = io.open(os.path.join(args.graph_path, "terms.txt"), encoding=args.encoding).readlines()
+    id2str = io.open(os.path.join(args.graph_work_path, "terms.txt"), encoding=args.encoding).readlines()
 
     trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0"))
     trainer_count = int(os.getenv("PADDLE_TRAINERS_NUM", "1"))
@@ -71,7 +71,7 @@ def run_predict(py_reader,
         
     for batch_feed_dict in py_reader():
         batch += 1
-        batch_usr_feat, batch_ad_feat, _, batch_src_real_index = exe.run(
+        batch_usr_feat, _, _, batch_src_real_index, _ = exe.run(
             program,
             feed=batch_feed_dict,
             fetch_list=model_dict.outputs)
@@ -79,7 +79,7 @@ def run_predict(py_reader,
         if batch % log_per_step == 0:
             log.info("Predict %s finished" % batch)
 
-        for ufs, _, sri in zip(batch_usr_feat, batch_ad_feat, batch_src_real_index):
+        for ufs, sri in zip(batch_usr_feat, batch_src_real_index):
             if args.input_type == "text":
                 sri = id2str[int(sri)].strip("\n")
             line = "{}\t{}\n".format(sri, tostr(ufs))
@@ -108,7 +108,7 @@ def _warmstart(exe, program, path='params'):
     )
 
 def main(config):
-    model = Model.factory(config)
+    model = LinkPredictModel(config)
 
     if config.learner_type == "cpu":
         place = F.CPUPlace()
@@ -143,7 +143,7 @@ def main(config):
             build_strategy=build_strategy,
             exec_strategy=exec_strategy)
 
-    num_nodes = int(np.load(os.path.join(config.graph_path, "num_nodes.npy")))
+    num_nodes = int(np.load(os.path.join(config.graph_work_path, "num_nodes.npy")))
 
     predict_data = PredictData(num_nodes)
 
@@ -156,7 +156,7 @@ def main(config):
         feed_name_list=[var.name for var in model.feed_list],
         use_pyreader=config.use_pyreader,
         phase="predict",
-        graph_data_path=config.graph_path,
+        graph_data_path=config.graph_work_path,
         shuffle=False)
 
     if config.learner_type == "cpu":

examples/erniesage/train.py → examples/erniesage/link_predict_train.py

@@ -22,27 +22,26 @@ from pgl.utils.logger import log
 from pgl.utils import paddle_helper
 
 from learner import Learner
-from models.model_factory import Model
+from models.model import LinkPredictModel
 from dataset.graph_reader import GraphGenerator 
 
 
 class TrainData(object):
-    def __init__(self, graph_path):
+    def __init__(self, graph_work_path):
         trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0"))
         trainer_count = int(os.getenv("PADDLE_TRAINERS_NUM", "1"))
         log.info("trainer_id: %s, trainer_count: %s." % (trainer_id, trainer_count))
 
-        bidirectional_edges = np.load(os.path.join(graph_path, "edges.npy"), allow_pickle=True)
+        edges = np.load(os.path.join(graph_work_path, "train_data.npy"), allow_pickle=True)
         # edges is bidirectional.
-        edges = bidirectional_edges[0::2]
         train_usr = edges[trainer_id::trainer_count, 0]
         train_ad = edges[trainer_id::trainer_count, 1]
         returns = {
             "train_data": [train_usr, train_ad]
         }
 
-        if os.path.exists(os.path.join(graph_path, "neg_samples.npy")):
-            neg_samples = np.load(os.path.join(graph_path, "neg_samples.npy"), allow_pickle=True)
+        if os.path.exists(os.path.join(graph_work_path, "neg_samples.npy")):
+            neg_samples = np.load(os.path.join(graph_work_path, "neg_samples.npy"), allow_pickle=True)
             if neg_samples.size != 0:
                 train_negs = neg_samples[trainer_id::trainer_count]
                 returns["train_data"].append(train_negs)
@@ -50,7 +49,7 @@ class TrainData(object):
         self.data = returns
 
     def __getitem__(self, index):
-        return [ data[index] for data in self.data["train_data"]]
+        return [data[index] for data in self.data["train_data"]]
 
     def __len__(self):
         return len(self.data["train_data"][0])
@@ -58,10 +57,10 @@ class TrainData(object):
 
 def main(config):
     # Select Model
-    model = Model.factory(config)
+    model = LinkPredictModel(config)
 
     # Build Train Edges
-    data = TrainData(config.graph_path)
+    data = TrainData(config.graph_work_path)
 
     # Build Train Data
     train_iter = GraphGenerator(
@@ -73,7 +72,7 @@ def main(config):
         feed_name_list=[var.name for var in model.feed_list],
         use_pyreader=config.use_pyreader,
         phase="train",
-        graph_data_path=config.graph_path,
+        graph_data_path=config.graph_work_path,
         shuffle=True,
         neg_type=config.neg_type)
 

examples/erniesage/models/erniesage_v2.py → examples/erniesage/models/encoder.py

+import numpy as np
 import pgl
 import paddle.fluid as F
 import paddle.fluid.layers as L
-from models.base import BaseNet, BaseGNNModel
 from models.ernie_model.ernie import ErnieModel
+from models.ernie_model.ernie import ErnieGraphModel
+from models.ernie_model.ernie import ErnieConfig
+from models import message_passing
+from models.message_passing import copy_send
 
 
-class ErnieSageV2(BaseNet):
+def get_layer(layer_type, gw, feature, hidden_size, act, initializer, learning_rate, name, is_test=False):
+    return getattr(message_passing, layer_type)(gw, feature, hidden_size, act, initializer, learning_rate, name)
 
-    def build_inputs(self):
-        inputs = super(ErnieSageV2, self).build_inputs()
-        term_ids = L.data(
-            "term_ids", shape=[None, self.config.max_seqlen], dtype="int64", append_batch_size=False)
-        return inputs + [term_ids]
+
+class Encoder(object):
+
+    def __init__(self, config):
+        self.config = config
+
+    @classmethod
+    def factory(cls, config):
+        model_type = config.model_type
+        if model_type == "ERNIESageV1":
+            return ERNIESageV1Encoder(config)
+        elif model_type == "ERNIESageV2":
+            return ERNIESageV2Encoder(config)
+        elif model_type == "ERNIESageV3":
+            return ERNIESageV3Encoder(config)
+        elif model_type == "ERNIESageV4":
+            return ERNIESageV4Encoder(config)
+        else:
+            raise ValueError
+
+    def __call__(self, graph_wrappers, inputs):
+        raise NotImplementedError
+
+
+class ERNIESageV1Encoder(Encoder):
+    def __call__(self, graph_wrappers, inputs):
+        feature = self.build_embedding(graph_wrappers[0].node_feat["term_ids"])
+
+        initializer = None
+        fc_lr = self.config.lr / 0.001
+        for i in range(self.config.num_layers):
+            if i == self.config.num_layers - 1:
+                act = None
+            else:
+                act = "leaky_relu"
+            feature = get_layer(
+                self.config.layer_type,
+                graph_wrappers[i],
+                feature,
+                self.config.hidden_size,
+                act,
+                initializer,
+                learning_rate=fc_lr,
+                name="%s_%s" % (self.config.layer_type, i))
+    
+        final_feats = [self.take_final_feature(feature, i, "final_fc") for i in inputs]
+        return final_feats
+
+    def build_embedding(self, term_ids):
+        term_ids = L.unsqueeze(term_ids, [-1])
+        ernie_config = self.config.ernie_config
+        ernie = ErnieModel(
+            src_ids=term_ids,
+            sentence_ids=L.zeros_like(term_ids),
+            task_ids=None,
+            config=ernie_config,
+            use_fp16=False,
+            name="")
+        feature = ernie.get_pooled_output()
+        return feature
+
+    def take_final_feature(self, feature, index, name):
+        """take final feature"""
+        feat = L.gather(feature, index, overwrite=False)
+
+        if self.config.final_fc:
+            feat = L.fc(feat,
+                           self.config.hidden_size,
+                           param_attr=F.ParamAttr(name=name + '_w'),
+                           bias_attr=F.ParamAttr(name=name + '_b'))
+
+        if self.config.final_l2_norm:
+            feat = L.l2_normalize(feat, axis=1)
+        return feat
+
+class ERNIESageV2Encoder(Encoder):
+    
+    def __call__(self, graph_wrappers, inputs):
+        feature = graph_wrappers[0].node_feat["term_ids"]
+        feature = self.gnn_layer(graph_wrappers[0], feature, self.config.hidden_size, 'leaky_relu', None, 1., "erniesage_v2_0")
+
+        initializer = None
+        fc_lr = self.config.lr / 0.001
+        for i in range(1, self.config.num_layers):
+            if i == self.config.num_layers - 1:
+                act = None
+            else:
+                act = "leaky_relu"
+            feature = get_layer(
+                self.config.layer_type,
+                graph_wrappers[i],
+                feature,
+                self.config.hidden_size,
+                act,
+                initializer,
+                learning_rate=fc_lr,
+                name="%s_%s" % (self.config.layer_type, i))
+    
+        final_feats = [self.take_final_feature(feature, i, "final_fc") for i in inputs]
+        return final_feats
+
+    def take_final_feature(self, feature, index, name):
+        """take final feature"""
+        feat = L.gather(feature, index, overwrite=False)
+
+        if self.config.final_fc:
+            feat = L.fc(feat,
+                           self.config.hidden_size,
+                           param_attr=F.ParamAttr(name=name + '_w'),
+                           bias_attr=F.ParamAttr(name=name + '_b'))
+
+        if self.config.final_l2_norm:
+            feat = L.l2_normalize(feat, axis=1)
+        return feat
 
     def gnn_layer(self, gw, feature, hidden_size, act, initializer, learning_rate, name):
         def build_position_ids(src_ids, dst_ids):
@@ -97,6 +211,36 @@ class ErnieSageV2(BaseNet):
             return output
         return erniesage_v2_aggregator(gw, feature, hidden_size, act, initializer, learning_rate, name)
 
+class ERNIESageV3Encoder(Encoder):
+
+    def __call__(self, graph_wrappers, inputs):
+        feature = graph_wrappers[0].node_feat["term_ids"]
+        feature = self.gnn_layer(graph_wrappers[0], feature, self.config.hidden_size, 'leaky_relu', None, 1., "erniesage_v3_0")
+
+        final_feats = [self.take_final_feature(feature, i, "final_fc") for i in inputs]
+        return final_feats
+
+    def gnn_layer(self, gw, feature, hidden_size, act, initializer, learning_rate, name):
+        def ernie_recv(feat):
+            """doc"""
+            num_neighbor = self.config.samples[0]
+            pad_value = L.zeros([1], "int64")
+            out, _ = L.sequence_pad(feat, pad_value=pad_value, maxlen=num_neighbor)
+            out = L.reshape(out, [0, self.config.max_seqlen*num_neighbor])
+            return out
+
+        def erniesage_v3_aggregator(gw, feature, hidden_size, act, initializer, learning_rate, name):
+            msg = gw.send(copy_send, nfeat_list=[("h", feature)])
+            neigh_feature = gw.recv(msg, ernie_recv)
+            neigh_feature = L.cast(L.unsqueeze(neigh_feature, [-1]), "int64")
+
+            feature = L.unsqueeze(feature, [-1])
+            cls = L.fill_constant_batch_size_like(feature, [-1, 1, 1], "int64", 1)
+            term_ids = L.concat([cls, feature[:, :-1], neigh_feature], 1)
+            term_ids.stop_gradient = True
+            return term_ids
+        return erniesage_v3_aggregator(gw, feature, hidden_size, act, initializer, learning_rate, name)
+
     def gnn_layers(self, graph_wrappers, feature):
         features = [feature]
 
@@ -116,20 +260,28 @@ class ErnieSageV2(BaseNet):
                 act,
                 initializer,
                 learning_rate=fc_lr,
-                name="%s_%s" % ("erniesage_v2", i))
+                name="%s_%s" % ("erniesage_v3", i))
             features.append(feature)
         return features
 
-    def __call__(self, graph_wrappers):
-        inputs = self.build_inputs()
-        feature = inputs[-1]
-        features = self.gnn_layers(graph_wrappers, feature)
-        outputs = [self.take_final_feature(features[-1], i, "final_fc") for i in inputs[:-1]]
-        src_real_index = L.gather(graph_wrappers[0].node_feat['index'], inputs[0])
-        outputs.append(src_real_index)
-        return inputs, outputs
+    def take_final_feature(self, feature, index, name):
+        """take final feature"""
+        feat = L.gather(feature, index, overwrite=False)
+
+        ernie_config = self.config.ernie_config
+        ernie = ErnieGraphModel(
+            src_ids=feat,
+            config=ernie_config,
+            slot_seqlen=self.config.max_seqlen)
+        feat = ernie.get_pooled_output()
+        fc_lr = self.config.lr / 0.001
 
+        if self.config.final_fc:
+            feat = L.fc(feat,
+                           self.config.hidden_size,
+                           param_attr=F.ParamAttr(name=name + '_w'),
+                           bias_attr=F.ParamAttr(name=name + '_b'))
 
-class ErnieSageModelV2(BaseGNNModel):
-    def gen_net_fn(self, config):
-        return ErnieSageV2(config)
+        if self.config.final_l2_norm:
+            feat = L.l2_normalize(feat, axis=1)
+        return feat

examples/stgcn/README.md

@@ -23,7 +23,7 @@ You can make your customized dataset by the following format:
 
 For examples, use gpu to train STGCN on your dataset.
 ```
-python main.py --use_cuda --input_file dataset/input_csv --label_file dataset/output.csv --adj_mat_file dataset/W.csv --city_file dataset/city.csv 
+python main.py --use_cuda --input_file dataset/input.csv --label_file dataset/output.csv --adj_mat_file dataset/W.csv --city_file dataset/city.csv 
 ```
 
 #### Hyperparameters

examples/stgcn/data_loader/data_utils.py

@@ -167,9 +167,6 @@ def data_gen_mydata(input_file, label_file, n, n_his, n_pred, n_config):
     x = x.drop(columns=['date'])
     y = y.drop(columns=['date'])
 
-    x = x.drop(columns=['武汉'])
-    y = y.drop(columns=['武汉'])
-
     # param
     n_val, n_test = n_config
     n_train = len(y) - n_val - n_test - 2

examples/stgcn/dataset/city.csv

+num,city
+0,A
+1,B
+2,C
+3,D
+4,E
+

pgl/__init__.py

@@ -22,3 +22,4 @@ from pgl import heter_graph
 from pgl import heter_graph_wrapper
 from pgl import contrib
 from pgl import message_passing
+from pgl import sample 

pgl/layers/__init__.py

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

pgl/tests/test_graph_saint_sample.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.
+"""graph saint sample test
+"""
+from __future__ import division
+from __future__ import absolute_import
+from __future__ import print_function
+from __future__ import unicode_literals
+import unittest
+import numpy as np
+
+import pgl
+import paddle.fluid as fluid
+from pgl.sample import graph_saint_random_walk_sample
+
+
+class GraphSaintSampleTest(unittest.TestCase):
+    """GraphSaintSampleTest"""
+
+    def test_randomwalk_sampler(self):
+        """test_randomwalk_sampler"""
+        g = pgl.graph.Graph(
+            num_nodes=8,
+            edges=[(1, 2), (2, 3), (0, 2), (0, 1), (6, 7), (4, 5), (6, 4),
+                   (7, 4), (3, 4)])
+        subgraph = graph_saint_random_walk_sample(g, [6, 7], 2)
+        print('reindex', subgraph._from_reindex)
+        print('subedges', subgraph.edges)
+        assert len(subgraph.nodes) == 4
+        assert len(subgraph.edges) == 4
+        true_edges = np.array([[0, 1], [2, 3], [2, 0], [3, 0]])
+        assert "{}".format(subgraph.edges) == "{}".format(true_edges)
+
+
+if __name__ == '__main__':
+    unittest.main()

pgl/utils/op.py

@@ -68,3 +68,18 @@ def read_rows(data, index):
         return new_data
     else:
         return paddle_helper.gather(data, index)
+
+
+class RowReader(object):
+    """Memory Efficient RowReader 
+    """
+    def __init__(self, nfeat, index):
+        self.nfeat = nfeat
+        self.loaded_nfeat = {}
+        self.index = index
+ 
+    def __getitem__(self, key):
+        if key not in self.loaded_nfeat:
+            self.loaded_nfeat[key] = read_rows(self.nfeat[key], self.index)
+        return self.loaded_nfeat[key]
+