Merge pull request #61 from PaddlePaddle/develop

pgl v1.1

README.md

 <img src="./docs/source/_static/logo.png" alt="The logo of Paddle Graph Learning (PGL)" width="320">
 
+[![PyPi Latest Release](https://img.shields.io/pypi/v/pgl.svg)](https://pypi.org/project/pgl/)
+[![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](./LICENSE)
+
 [DOC](https://pgl.readthedocs.io/en/latest/) | [Quick Start](https://pgl.readthedocs.io/en/latest/quick_start/instruction.html) | [中文](./README.zh.md)
 
+## Breaking News !!
+
+PGL v1.1 2020.4.29
+
+- You can find **ERNIESage**, a novel model for modeling text and graph structures, and its introduction [here](./examples/erniesage/).
+
+- PGL for [Open Graph Benchmark](https://github.com/snap-stanford/ogb) examples can be find [here](./ogb_examples/).
+
+- We add newly graph level operators like **GraphPooling** and [**GraphNormalization**](https://arxiv.org/abs/2003.00982) for graph level predictions.
+
+- We relase a PGL-KE toolkit [here](./examples/pgl-ke) including classical knowledge graph embedding t algorithms like TransE, TransR, RotatE.
+
+------
+
 Paddle Graph Learning (PGL) is an efficient and flexible graph learning framework based on [PaddlePaddle](https://github.com/PaddlePaddle/Paddle).
 
 
-<img src="./docs/source/_static/framework_of_pgl.png" alt="The Framework of Paddle Graph Learning (PGL)" width="800">
+<img src="./docs/source/_static/framework_of_pgl_en.png" alt="The Framework of Paddle Graph Learning (PGL)" width="800">
 
 
 The newly released PGL supports heterogeneous graph learning on both walk based paradigm and message-passing based paradigm by providing MetaPath sampling and Message Passing mechanism on heterogeneous graph. Furthermor, The newly released PGL also support distributed graph storage and some distributed training algorithms, such as distributed deep walk and distributed graphsage. Combined with the PaddlePaddle deep learning framework, we are able to support both graph representation learning models and graph neural networks, and thus our framework has a wide range of graph-based applications.
@@ -82,10 +99,11 @@ In most cases of large-scale graph learning, we need distributed graph storage a
 
 
 ## Model Zoo
-The following are 13 graph learning models that have been implemented in the framework. See the details [here](https://pgl.readthedocs.io/en/latest/introduction.html#highlight-tons-of-models)
+The following graph learning models have been implemented in the framework. You can find more [examples](./examples) and the [details](https://pgl.readthedocs.io/en/latest/introduction.html#highlight-tons-of-models)
 
 |Model | feature |
 |---|---|
+| [**ERNIESage**](./examples/erniesage/) | ERNIE SAmple aggreGatE for Text and Graph |
 | GCN | Graph Convolutional Neural Networks |
 | GAT | Graph Attention Network |
 | GraphSage |Large-scale graph convolution network based on neighborhood sampling|

README.zh.md

 <img src="./docs/source/_static/logo.png" alt="The logo of Paddle Graph Learning (PGL)" width="320">
 
+[![PyPi Latest Release](https://img.shields.io/pypi/v/pgl.svg)](https://pypi.org/project/pgl/)
+[![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](./LICENSE)
+
 [文档](https://pgl.readthedocs.io/en/latest/) | [快速开始](https://pgl.readthedocs.io/en/latest/quick_start/instruction.html) | [English](./README.md)
 
+## 最新消息
+
+PGL v1.1 2020.4.29
+
+- **ERNIESage**是PGL团队最新提出的模型，可以用于建模文本以及图结构信息。你可以在[这里](./examples/erniesage)看到详细的介绍。
+
+- PGL现在提供[Open Graph Benchmark](https://github.com/snap-stanford/ogb)的一些例子，你可以在[这里](./ogb_examples)找到。
+
+- 新增了图级别的算子包括**GraphPooling**以及[**GraphNormalization**](https://arxiv.org/abs/2003.00982)，这样你就能实现更多复杂的图级别分类模型。
+
+- 新增PGL-KE工具包，里面包含许多经典知识图谱图嵌入算法，包括TransE, TransR, RotatE，详情可见[这里](./examples/pgl-ke)
+
+------
+
 Paddle Graph Learning (PGL)是一个基于[PaddlePaddle](https://github.com/PaddlePaddle/Paddle)的高效易用的图学习框架
 
 <img src="./docs/source/_static/framework_of_pgl.png" alt="The Framework of Paddle Graph Learning (PGL)" width="800">
@@ -77,10 +94,11 @@ Paddle Graph Learning (PGL)是一个基于[PaddlePaddle](https://github.com/Padd
 
 ## 丰富性——覆盖业界大部分图学习网络
 
-下列是框架中已经自带实现的十三种图网络学习模型。详情请参考[这里](https://pgl.readthedocs.io/en/latest/introduction.html#highlight-tons-of-models)
+下列是框架中部分已经实现的图网络模型，更多的模型在[这里](./examples)可以找到。详情请参考[这里](https://pgl.readthedocs.io/en/latest/introduction.html#highlight-tons-of-models)
 
 | 模型 | 特点 |
 |---|---|
+| [**ERNIESage**](./examples/erniesage/) | 能同时建模文本以及图结构的ERNIE SAmple aggreGatE |
 | GCN | 图卷积网络 |
 | GAT | 基于Attention的图卷积网络 |
 | GraphSage | 基于邻居采样的大规模图卷积网络 |

docs/source/quick_start/md/quick_start.md

@@ -53,7 +53,6 @@ place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
 
 # use GraphWrapper as a container for graph data to construct a graph neural network
 gw = pgl.graph_wrapper.GraphWrapper(name='graph',
-                        place = place,
                         node_feat=g.node_feat_info())
 ```
 

docs/source/quick_start/md/quick_start_for_heterGraph.md

@@ -77,7 +77,6 @@ place = fluid.CPUPlace()
 
 # create a GraphWrapper as a container for graph data
 gw = heter_graph_wrapper.HeterGraphWrapper(name='heter_graph', 
-                                    place = place, 
                                     edge_types = g.edge_types_info(),
                                     node_feat=g.node_feat_info(),
                                     edge_feat=g.edge_feat_info())

examples/GATNE/model.py

@@ -53,7 +53,6 @@ class GATNE(object):
 
         self.gw = heter_graph_wrapper.HeterGraphWrapper(
             name="heter_graph",
-            place=place,
             edge_types=self.graph.edge_types_info(),
             node_feat=self.graph.node_feat_info(),
             edge_feat=self.graph.edge_feat_info())

examples/dgi/train.py

@@ -65,7 +65,6 @@ def main(args):
     with fluid.program_guard(train_program, startup_program):
         gw = pgl.graph_wrapper.GraphWrapper(
             name="graph",
-            place=place,
             node_feat=dataset.graph.node_feat_info())
 
         output = pgl.layers.gcn(gw,

examples/distribute_graphsage/train.py

@@ -170,7 +170,7 @@ def main(args):
 
     with fluid.program_guard(train_program, startup_program):
         graph_wrapper = pgl.graph_wrapper.GraphWrapper(
-            "sub_graph", fluid.CPUPlace(), node_feat=[('feats', [None, 602], np.dtype('float32'))])
+            "sub_graph", node_feat=[('feats', [None, 602], np.dtype('float32'))])
         model_loss, model_acc = build_graph_model(
             graph_wrapper,
             num_class=data["num_class"],

examples/erniesage/README.en.md

+# ERNIESage in PGL
+
+[中文版 README](./README.md)
+
+
+## Introduction
+In many industrial applications, there is often a special graph shown below: Text Graph. As the name implies, the node attributes of such graph consist of text, and the edges provide structural information. Take the search scenario for example, nodes can be expressed by search query, web page titles, and web page content, while the edges are constructed by user feedback or hyperlink information.
+
+<img src="./docs/source/_static/text_graph.png" alt="Text Graph" width="800">
+
+**ERNIESage** (abbreviation of ERNIE SAmple aggreGatE), a model proposed by the PGL team, effectively improves the performance on text graph by simultaneously modeling text semantics and graph structure information. It's worth mentioning that [**ERNIE**](https://github.com/PaddlePaddle/ERNIE) in **ERNIESage** is a continual pre-training framework for language understanding launched by Baidu.
+
+**ERNIESage** is an aggregation of ERNIE and GraphSAGE. Its structure is shown in the figure below. The main idea is to use ERNIE as an aggregation function (Aggregators) to model the semantic and structural relationship between its own nodes and neighbor nodes. In addition, for the position-independent characteristics of neighbor nodes, attention mask and independent position embedding mechanism for neighbor blindness are designed.
+
+<img src="./docs/source/_static/ernie_aggregator.png" alt="ERNIESage" width="800">
+
+GraphSAGE with ID feature can only model the graph structure information, while ERNIE can only deal with the text. With the help of PGL, the proposed **ERNIESage** model can combine the advantages of both models. Take the following recommendation example of text graph, we can see that **ERNIESage** achieves the best performance when compared to single ERNIE model or GraphSAGE model.
+
+<img src="./docs/source/_static/ERNIESage_result.png" alt="ERNIESage_result" width="800">
+
+Thanks to the flexibility and usability of PGL, **ERNIESage** can be quickly implemented under PGL's Message Passing paradigm. Acutally, there are four PGL version of ERNIESage:
+
+- **ERNIESage v1**: ERNIE is applied to the NODE of the text graph;
+- **ERNIESage v2**: ERNIE is applied to the EDGE of the text graph;
+- **ERNIESage v3**: ERNIE is applied to the first order neighbors and center node;
+- **ERNIESage v4**: ERNIE is applied to the N-order neighbors and center node.
+
+<img src="./docs/source/_static/ERNIESage_v1_4.png" alt="ERNIESage_v1_4" width="800">
+
+## Dependencies
+- paddlepaddle>=1.7
+- pgl>=1.1
+
+## Dataformat
+
+## How to run
+
+We adopt [PaddlePaddle Fleet](https://github.com/PaddlePaddle/Fleet) as our distributed training frameworks ```config/*.yaml``` are some example config files for hyperparameters.
+
+```sh
+# train ERNIESage in distributed gpu mode.
+sh local_run.sh config/enriesage_v1_gpu.yaml
+
+# train ERNIESage in distributed cpu mode.
+sh local_run.sh config/enriesage_v1_cpu.yaml
+```
+
+## Hyperparamters
+
+- learner_type: `gpu` or `cpu`; gpu use fleet Collective mode, cpu use fleet Transpiler mode.
+
+## Citation
+```
+@misc{ERNIESage,
+  author = {PGL Team},
+  title = {ERNIESage: ERNIE SAmple aggreGatE},
+  year = {2020},
+  publisher = {GitHub},
+  journal = {GitHub repository},
+  howpublished = {\url{https://github.com/PaddlePaddle/PGL/tree/master/examples/erniesage},
+}
+```

examples/erniesage/README.md

+# 使用PGL实现ERNIESage
+
+[ENG Readme](./README.en.md)
+
+## 背景介绍 
+
+在很多工业应用中，往往出现如下图所示的一种特殊的图：Text Graph。顾名思义，图的节点属性由文本构成，而边的构建提供了结构信息。如搜索场景下的Text Graph，节点可由搜索词、网页标题、网页正文来表达，用户反馈和超链信息则可构成边关系。
+
+<img src="./docs/source/_static/text_graph.png" alt="Text Graph" width="800">
+
+**ERNIESage** 由PGL团队提出，是ERNIE SAmple aggreGatE的简称，该模型可以同时建模文本语义与图结构信息，有效提升 Text Graph 的应用效果。其中 [**ERNIE**](https://github.com/PaddlePaddle/ERNIE) 是百度推出的基于知识增强的持续学习语义理解框架，
+
+**ERNIESage** 是 ERNIE 与 GraphSAGE 碰撞的结果，是 ERNIE SAmple aggreGatE 的简称，它的结构如下图所示，主要思想是通过 ERNIE 作为聚合函数（Aggregators），建模自身节点和邻居节点的语义与结构关系。ERNIESage 对于文本的建模是构建在邻居聚合的阶段，中心节点文本会与所有邻居节点文本进行拼接；然后通过预训练的 ERNIE 模型进行消息汇聚，捕捉中心节点以及邻居节点之间的相互关系；最后使用 ERNIESage 搭配独特的邻居互相看不见的 Attention Mask 和独立的 Position Embedding 体系，就可以轻松构建 TextGraph 中句子之间以及词之间的关系。
+
+<img src="./docs/source/_static/ernie_aggregator.png" alt="ERNIESage" width="800">
+
+使用ID特征的GraphSAGE只能够建模图的结构信息，而单独的ERNIE只能处理文本信息。通过PGL搭建的图与文本的桥梁，**ERNIESage**能够很简单的把GraphSAGE以及ERNIE的优点结合一起。以下面TextGraph的场景，**ERNIESage**的效果能够比单独的ERNIE以及GraphSAGE模型都要好。
+
+<img src="./docs/source/_static/ERNIESage_result.png" alt="ERNIESage_result" width="800">
+
+**ERNIESage**可以很轻松地在PGL中的消息传递范式中进行实现，目前PGL提供了4个版本的ERNIESage模型：
+
+- **ERNIESage v1**: ERNIE 作用于text graph节点上;
+- **ERNIESage v2**: ERNIE 作用在text graph的边上;
+- **ERNIESage v3**: ERNIE 作用于一阶邻居及起边上;
+- **ERNIESage v4**: ERNIE 作用于N阶邻居及边上;
+
+<img src="./docs/source/_static/ERNIESage_v1_4.png" alt="ERNIESage_v1_4" width="800">
+
+## 环境依赖
+- paddlepaddle>=1.7
+- pgl>=1.1
+
+## Dataformat
+
+## How to run
+
+我们采用了[PaddlePaddle Fleet](https://github.com/PaddlePaddle/Fleet)作为我们的分布式训练框架，在```config/*.yaml```中，由部分用于训练ERNIESage的配置。
+
+```sh
+# 分布式GPU模式或单机模式ERNIESage
+sh local_run.sh config/enriesage_v1_gpu.yaml
+
+# 分布式CPU模式训练ERNIESage
+sh local_run.sh config/enriesage_v1_cpu.yaml
+```
+
+## Hyperparamters
+
+- learner_type: `gpu` or `cpu`; gpu 使用fleet Collective 模式, cpu 使用fleet Transpiler 模式.
+
+## Citation
+```
+@misc{ERNIESage,
+  author = {PGL Team},
+  title = {ERNIESage: ERNIE SAmple aggreGatE},
+  year = {2020},
+  publisher = {GitHub},
+  journal = {GitHub repository},
+  howpublished = {\url{https://github.com/PaddlePaddle/PGL/tree/master/examples/erniesage},
+}
+```

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_v1_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: "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: 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: "ErnieSageModelV2"
+
+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_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: "ErnieSageModelV2"
+
+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_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/base_dataset.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.
+"""Base DataLoader 
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+from __future__ import unicode_literals
+from __future__ import absolute_import
+
+import os
+import sys
+import six
+from io import open
+from collections import namedtuple
+import numpy as np
+import tqdm
+import paddle
+from pgl.utils import mp_reader
+import collections
+import time
+from pgl.utils.logger import log
+import traceback
+
+
+if six.PY3:
+    import io
+    sys.stdout = io.TextIOWrapper(sys.stdout.buffer, encoding='utf-8')
+    sys.stderr = io.TextIOWrapper(sys.stderr.buffer, encoding='utf-8')
+
+
+def batch_iter(data, perm, batch_size, fid, num_workers):
+    """node_batch_iter
+    """
+    size = len(data)
+    start = 0
+    cc = 0
+    while start < size:
+        index = perm[start:start + batch_size]
+        start += batch_size
+        cc += 1
+        if cc % num_workers != fid:
+            continue
+        yield data[index]
+
+
+def scan_batch_iter(data, batch_size, fid, num_workers):
+    """node_batch_iter
+    """
+    batch = []
+    cc = 0
+    for line_example in data.scan(): 
+        cc += 1
+        if cc % num_workers != fid:
+            continue
+        batch.append(line_example)
+        if len(batch) == batch_size:
+            yield batch 
+            batch = []
+
+    if len(batch) > 0:
+        yield batch 
+
+
+class BaseDataGenerator(object):
+    """Base Data Geneartor"""
+
+    def __init__(self, buf_size, batch_size, num_workers, shuffle=True):
+        self.num_workers = num_workers
+        self.batch_size = batch_size
+        self.line_examples = []
+        self.buf_size = buf_size
+        self.shuffle = shuffle
+
+    def batch_fn(self, batch_examples):
+        """ batch_fn batch producer"""
+        raise NotImplementedError("No defined Batch Fn")
+
+    def batch_iter(self, fid, perm):
+        """ batch iterator"""
+        if self.shuffle:
+            for batch in batch_iter(self, perm, self.batch_size, fid, self.num_workers):
+                yield batch
+        else:
+            for batch in scan_batch_iter(self, self.batch_size, fid, self.num_workers):
+                yield batch
+
+    def __len__(self):
+        return len(self.line_examples)
+
+    def __getitem__(self, idx):
+        if isinstance(idx, collections.Iterable):
+            return [self[bidx] for bidx in idx]
+        else:
+            return self.line_examples[idx]
+
+    def generator(self):
+        """batch dict generator"""
+
+        def worker(filter_id, perm):
+            """ multiprocess worker"""
+
+            def func_run():
+                """ func_run """
+                pid = os.getpid()
+                np.random.seed(pid + int(time.time()))
+                for batch_examples in self.batch_iter(filter_id, perm):
+                    try:
+                        batch_dict = self.batch_fn(batch_examples)
+                    except Exception as e:
+                       traceback.print_exc()
+                       log.info(traceback.format_exc())
+                       log.info(str(e))
+                       continue
+
+                    if batch_dict is None:
+                        continue
+                    yield batch_dict
+
+
+
+            return func_run
+
+        # consume a seed
+        np.random.rand()
+
+        if self.shuffle:
+            perm = np.arange(0, len(self)) 
+            np.random.shuffle(perm)
+        else:
+            perm = None
+
+        if self.num_workers == 1:
+            r = paddle.reader.buffered(worker(0, perm), self.buf_size)
+        else:
+            worker_pool = [worker(wid, perm) for wid in range(self.num_workers)]
+            worker = mp_reader.multiprocess_reader(
+                worker_pool, use_pipe=True, queue_size=1000)
+            r = paddle.reader.buffered(worker, self.buf_size)
+
+        for batch in r():
+            yield batch
+
+    def scan(self): 
+        for line_example in self.line_examples:
+            yield line_example

examples/erniesage/dataset/graph_reader.py

+"""Graph Dataset
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+from __future__ import unicode_literals
+from __future__ import absolute_import
+
+import os
+import pgl
+import sys
+
+import numpy as np
+
+from pgl.utils.logger import log
+from dataset.base_dataset import BaseDataGenerator
+from pgl.sample import alias_sample
+from pgl.sample import pinsage_sample
+from pgl.sample import graphsage_sample 
+from pgl.sample import edge_hash
+
+
+class GraphGenerator(BaseDataGenerator):
+    def __init__(self, graph_wrappers, data, batch_size, samples,
+        num_workers, feed_name_list, use_pyreader,
+        phase, graph_data_path, shuffle=True, buf_size=1000):
+
+        super(GraphGenerator, self).__init__(
+            buf_size=buf_size,
+            num_workers=num_workers,
+            batch_size=batch_size, shuffle=shuffle)
+        # For iteration
+        self.line_examples = data
+
+        self.graph_wrappers = graph_wrappers
+        self.samples = samples
+        self.feed_name_list = feed_name_list
+        self.use_pyreader = use_pyreader
+        self.phase = phase
+        self.load_graph(graph_data_path)
+        self.num_layers = len(graph_wrappers)
+
+    def load_graph(self, graph_data_path):
+        self.graph = pgl.graph.MemmapGraph(graph_data_path)
+        self.alias = np.load(os.path.join(graph_data_path, "alias.npy"), mmap_mode="r")
+        self.events = np.load(os.path.join(graph_data_path, "events.npy"), mmap_mode="r")
+        self.term_ids = np.load(os.path.join(graph_data_path, "term_ids.npy"), mmap_mode="r")
+ 
+    def batch_fn(self, batch_ex):
+        # batch_ex = [
+        #     (src, dst, neg),
+        #     (src, dst, neg),
+        #     (src, dst, neg),
+        #     ]
+        #
+        batch_src = []
+        batch_dst = []
+        batch_neg = []
+        for batch in batch_ex:
+            batch_src.append(batch[0])
+            batch_dst.append(batch[1])
+            if len(batch) == 3: # default neg samples
+                batch_neg.append(batch[2])
+
+        if len(batch_src) != self.batch_size:
+            if self.phase == "train":
+                return None  #Skip
+
+        if len(batch_neg) > 0:
+            batch_neg = np.unique(np.concatenate(batch_neg))
+        batch_src = np.array(batch_src, dtype="int64")
+        batch_dst = np.array(batch_dst, dtype="int64")
+
+        sampled_batch_neg = alias_sample(batch_dst.shape, self.alias, self.events)
+    
+        if len(batch_neg) > 0:
+            batch_neg = np.concatenate([batch_neg, sampled_batch_neg], 0)
+        else:
+            batch_neg = sampled_batch_neg
+
+        if self.phase == "train":
+            ignore_edges = set()
+        else:
+            ignore_edges = set()
+
+        nodes = np.unique(np.concatenate([batch_src, batch_dst, batch_neg], 0))
+        subgraphs = graphsage_sample(self.graph, nodes, self.samples, ignore_edges=ignore_edges)
+        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_src_idx = subgraphs[0].reindex_from_parrent_nodes(batch_src)
+        sub_dst_idx = subgraphs[0].reindex_from_parrent_nodes(batch_dst)
+        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["neg_item_index"] = np.array(sub_neg_idx, dtype="int64")
+        feed_dict["term_ids"] = self.term_ids[subgraphs[0].node_feat["index"]]
+        return feed_dict
+
+    def __call__(self):
+        return self.generator()
+
+    def generator(self):
+        try:
+            for feed_dict in super(GraphGenerator, self).generator():
+                if self.use_pyreader:
+                    yield [feed_dict[name] for name in self.feed_name_list]
+                else:
+                    yield feed_dict
+
+        except Exception as e:
+            log.exception(e)
+ 
+
+    

examples/erniesage/infer.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.
+from __future__ import division
+from __future__ import absolute_import
+from __future__ import print_function
+from __future__ import unicode_literals
+import argparse
+import pickle
+import time
+import glob
+import os
+import io
+import traceback
+import pickle as pkl
+role = os.getenv("TRAINING_ROLE", "TRAINER")
+
+import numpy as np
+import yaml
+from easydict import EasyDict as edict
+import pgl
+from pgl.utils.logger import log
+from pgl.utils import paddle_helper
+import paddle
+import paddle.fluid as F
+
+from models.model_factory import Model
+from dataset.graph_reader import GraphGenerator 
+
+
+class PredictData(object):
+    def __init__(self, num_nodes):
+        trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0"))
+        trainer_count = int(os.getenv("PADDLE_TRAINERS_NUM", "1"))
+        train_usr = np.arange(trainer_id, num_nodes, trainer_count)
+        #self.data = (train_usr, train_usr)
+        self.data = train_usr
+
+    def __getitem__(self, index):
+        return [self.data[index], self.data[index]]
+
+def tostr(data_array):
+    return " ".join(["%.5lf" % d for d in  data_array])
+
+def run_predict(py_reader,
+              exe,
+              program,
+              model_dict,
+              log_per_step=1,
+              args=None):
+
+    if args.input_type == "text":
+        id2str = np.load(os.path.join(args.graph_path, "id2str.npy"), mmap_mode="r")
+
+    trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0"))
+    trainer_count = int(os.getenv("PADDLE_TRAINERS_NUM", "1"))
+    if not os.path.exists(args.output_path):
+        os.mkdir(args.output_path)
+
+    fout = io.open("%s/part-%s" % (args.output_path, trainer_id), "w", encoding="utf8")
+    batch = 0
+        
+    for batch_feed_dict in py_reader():
+        batch += 1
+        batch_usr_feat, batch_ad_feat, batch_src_real_index = exe.run(
+            program,
+            feed=batch_feed_dict,
+            fetch_list=model_dict.outputs)
+
+        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):
+            if args.input_type == "text":
+                sri = id2str[int(sri)]
+            line = "{}\t{}\n".format(sri, tostr(ufs))
+            fout.write(line)
+
+    fout.close()
+
+def _warmstart(exe, program, path='params'):
+    def _existed_persitables(var):
+        #if not isinstance(var, fluid.framework.Parameter):
+        #    return False
+        if not F.io.is_persistable(var):
+            return False
+        param_path = os.path.join(path, var.name)
+        log.info("Loading parameter: {} persistable: {} exists: {}".format(
+            param_path,
+            F.io.is_persistable(var),
+            os.path.exists(param_path),
+        ))
+        return os.path.exists(param_path)
+    F.io.load_vars(
+        exe,
+        path,
+        main_program=program,
+        predicate=_existed_persitables
+    )
+
+def main(config):
+    model = Model.factory(config)
+
+    if config.learner_type == "cpu":
+        place = F.CPUPlace()
+    elif config.learner_type == "gpu":
+        gpu_id = int(os.getenv("FLAGS_selected_gpus", "0"))
+        place = F.CUDAPlace(gpu_id)
+    else:
+        raise ValueError
+
+    exe = F.Executor(place)
+
+    val_program = F.default_main_program().clone(for_test=True)
+    exe.run(F.default_startup_program()) 
+    _warmstart(exe, F.default_startup_program(), path=config.infer_model)
+
+    num_threads = int(os.getenv("CPU_NUM", 1))
+    trainer_id = int(os.getenv("PADDLE_TRAINER_ID", 0))
+
+    exec_strategy = F.ExecutionStrategy()
+    exec_strategy.num_threads = num_threads
+    build_strategy = F.BuildStrategy()
+    build_strategy.enable_inplace = True
+    build_strategy.memory_optimize = True
+    build_strategy.remove_unnecessary_lock = False
+    build_strategy.memory_optimize = False
+
+    if num_threads > 1:
+        build_strategy.reduce_strategy = F.BuildStrategy.ReduceStrategy.Reduce
+
+    val_compiled_prog = F.compiler.CompiledProgram(
+        val_program).with_data_parallel(
+            build_strategy=build_strategy,
+            exec_strategy=exec_strategy)
+
+    num_nodes = int(np.load(os.path.join(config.graph_path, "num_nodes.npy")))
+
+    predict_data = PredictData(num_nodes)
+
+    predict_iter = GraphGenerator(
+        graph_wrappers=model.graph_wrappers,
+        batch_size=config.infer_batch_size,
+        data=predict_data,
+        samples=config.samples,
+        num_workers=config.sample_workers,
+        feed_name_list=[var.name for var in model.feed_list],
+        use_pyreader=config.use_pyreader,
+        phase="predict",
+        graph_data_path=config.graph_path,
+        shuffle=False)
+
+    if config.learner_type == "cpu":
+        model.data_loader.decorate_batch_generator(
+            predict_iter, places=F.cpu_places())
+    elif config.learner_type == "gpu":
+        gpu_id = int(os.getenv("FLAGS_selected_gpus", "0"))
+        place = F.CUDAPlace(gpu_id)
+        model.data_loader.decorate_batch_generator(
+            predict_iter, places=place)
+    else:
+        raise ValueError
+
+    run_predict(model.data_loader,
+                program=val_compiled_prog,
+                exe=exe,
+                model_dict=model,
+                args=config)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='main')
+    parser.add_argument("--conf", type=str, default="./config.yaml")
+    args = parser.parse_args()
+    config = edict(yaml.load(open(args.conf), Loader=yaml.FullLoader))
+    print(config)
+    main(config)

examples/erniesage/job.sh

+
+unset http_proxy https_proxy
+set -x
+mode=${1:-local}
+config=${2:-"./config.yaml"}
+
+function parse_yaml {
+   local prefix=$2
+   local s='[[:space:]]*' w='[a-zA-Z0-9_]*' fs=$(echo @|tr @ '\034')
+   sed -ne "s|^\($s\):|\1|" \
+        -e "s|^\($s\)\($w\)$s:$s[\"']\(.*\)[\"']$s\$|\1$fs\2$fs\3|p" \
+        -e "s|^\($s\)\($w\)$s:$s\(.*\)$s\$|\1$fs\2$fs\3|p"  $1 |
+   awk -F$fs '{
+      indent = length($1)/2;
+      vname[indent] = $2;
+      for (i in vname) {if (i > indent) {delete vname[i]}}
+      if (length($3) > 0) {
+         vn=""; for (i=0; i<indent; i++) {vn=(vn)(vname[i])("_")}
+         printf("%s%s%s=\"%s\"\n", "'$prefix'",vn, $2, $3);
+      }
+   }'
+}
+eval $(parse_yaml $config)
+
+export CPU_NUM=$CPU_NUM
+export FLAGS_rpc_deadline=3000000 
+export FLAGS_rpc_retry_times=1000
+
+if [[ $async_mode == "True" ]];then
+    echo "async_mode is True"
+else
+    export FLAGS_communicator_send_queue_size=1
+    export FLAGS_communicator_min_send_grad_num_before_recv=0
+    export FLAGS_communicator_max_merge_var_num=1 # important! 
+    export FLAGS_communicator_merge_sparse_grad=0
+fi
+
+export FLAGS_communicator_recv_wait_times=5000000
+
+mkdir -p output
+
+python ./train.py --conf $config
+if [[ $TRAINING_ROLE == "TRAINER" ]];then
+    python ./infer.py --conf $config
+fi

examples/erniesage/learner.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 time
+import os
+role = os.getenv("TRAINING_ROLE", "TRAINER")
+
+import numpy as np
+from pgl.utils.logger import log
+import paddle.fluid as F
+import paddle.fluid.layers as L
+from paddle.fluid.incubate.fleet.parameter_server.distribute_transpiler import StrategyFactory
+from paddle.fluid.incubate.fleet.collective import DistributedStrategy
+from paddle.fluid.transpiler.distribute_transpiler import DistributeTranspilerConfig
+from paddle.fluid.incubate.fleet.collective import fleet as cfleet
+from paddle.fluid.incubate.fleet.parameter_server.distribute_transpiler import fleet as tfleet
+import paddle.fluid.incubate.fleet.base.role_maker as role_maker
+from tensorboardX import SummaryWriter
+
+
+class Learner(object):
+    @classmethod
+    def factory(cls, name):
+        if name == "cpu":
+            return TranspilerLearner()
+        elif name == "gpu":
+            return CollectiveLearner()
+        else:
+            raise ValueError
+
+    def build(self, model, data_gen, config):
+        raise NotImplementedError
+
+    def warmstart(self, program, path='./checkpoints'):
+        def _existed_persitables(var):
+            #if not isinstance(var, fluid.framework.Parameter):
+            #    return False
+            if not F.io.is_persistable(var):
+                return False
+            param_path = os.path.join(path, var.name)
+            log.info("Loading parameter: {} persistable: {} exists: {}".format(
+                param_path,
+                F.io.is_persistable(var),
+                os.path.exists(param_path),
+            ))
+            return os.path.exists(param_path)
+        F.io.load_vars(
+            self.exe,
+            path,
+            main_program=program,
+            predicate=_existed_persitables
+        )
+
+    def start(self):
+        batch = 0
+        start = time.time()
+        trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0"))
+        if trainer_id == 0:
+            writer = SummaryWriter(os.path.join(self.config.output_path, "train_history"))
+
+        for epoch_idx in range(self.config.epoch):
+            for idx, batch_feed_dict in enumerate(self.model.data_loader()):
+                try:
+                    cpu_time = time.time()
+                    batch += 1
+                    batch_loss  = self.exe.run(
+                        self.program,
+                        feed=batch_feed_dict,
+                        fetch_list=[self.model.loss])
+                    end = time.time()
+                    if trainer_id == 0:
+                        writer.add_scalar("loss", np.mean(batch_loss), batch)
+                        if batch % self.config.log_per_step == 0:
+                            log.info(
+                                "Epoch %s Batch %s %s-Loss %s \t Speed(per batch) %.5lf/%.5lf sec"
+                                % (epoch_idx, batch, "train", np.mean(batch_loss), (end - start) /batch, (end - cpu_time)))
+                            writer.flush()
+                        if batch % self.config.save_per_step == 0:
+                            self.fleet.save_persistables(self.exe, os.path.join(self.config.output_path, str(batch)))
+                except Exception as e:
+                    log.info("Pyreader train error")
+                    log.exception(e)
+            log.info("epcoh %s done." % epoch_idx)
+
+    def stop(self):
+        self.fleet.save_persistables(self.exe, os.path.join(self.config.output_path, "last"))
+
+
+class TranspilerLearner(Learner):
+    def __init__(self):
+        training_role = os.getenv("TRAINING_ROLE", "TRAINER")
+        paddle_role = role_maker.Role.WORKER
+        place = F.CPUPlace()
+        if training_role == "PSERVER":
+            paddle_role = role_maker.Role.SERVER
+
+        # set the fleet runtime environment according to configure
+        port = os.getenv("PADDLE_PORT", "6174")
+        pserver_ips = os.getenv("PADDLE_PSERVERS")  # ip,ip...
+        eplist = []
+        for ip in pserver_ips.split(","):
+            eplist.append(':'.join([ip, port]))
+        pserver_endpoints = eplist  # ip:port,ip:port...
+        worker_num = int(os.getenv("PADDLE_TRAINERS_NUM", "0"))
+        trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0"))
+        role = role_maker.UserDefinedRoleMaker(
+            current_id=trainer_id,
+            role=paddle_role,
+            worker_num=worker_num,
+            server_endpoints=pserver_endpoints)
+        tfleet.init(role)
+        tfleet.save_on_pserver = True
+
+    def build(self, model, data_gen, config):
+        self.optimize(model.loss, config.optimizer_type, config.lr)
+        self.init_and_run_ps_worker(config.ckpt_path)
+        self.program = self.complie_program(model.loss)
+        self.fleet = tfleet
+        model.data_loader.decorate_batch_generator(
+            data_gen, places=F.cpu_places())
+        self.config = config
+        self.model = model
+
+    def optimize(self, loss, optimizer_type, lr):
+        strategy = DistributeTranspilerConfig()
+        strategy.sync_mode = False
+        log.info('learning rate:%f' % lr)
+        if optimizer_type == "sgd":
+            optimizer = F.optimizer.SGD(learning_rate=lr)
+        elif optimizer_type == "adam":
+            # Don't slice tensor ensure convergence 
+            optimizer = F.optimizer.Adam(learning_rate=lr, lazy_mode=True)
+        else:
+            raise ValueError("Unknown Optimizer %s" % optimizer_type)
+        #create the DistributeTranspiler configure
+        optimizer = tfleet.distributed_optimizer(optimizer, strategy)
+        optimizer.minimize(loss)
+
+    def init_and_run_ps_worker(self, ckpt_path):
+        # init and run server or worker
+        self.exe = F.Executor(F.CPUPlace())
+        if tfleet.is_server():
+            tfleet.init_server()
+            self.warmstart(tfleet.startup_program, path=ckpt_path)
+            tfleet.run_server()
+            exit()
+
+        if tfleet.is_worker():
+            log.info("start init worker done")
+            tfleet.init_worker()
+            self.exe.run(tfleet.startup_program)
+
+    def complie_program(self, loss):
+        num_threads = int(os.getenv("CPU_NUM", 1))
+        exec_strategy = F.ExecutionStrategy()
+        exec_strategy.num_threads = num_threads
+        exec_strategy.use_thread_barrier = False
+        build_strategy = F.BuildStrategy()
+        build_strategy.enable_inplace = True
+        build_strategy.memory_optimize = True
+        build_strategy.remove_unnecessary_lock = False
+        build_strategy.memory_optimize = False
+        build_strategy.async_mode = False
+
+        if num_threads > 1:
+            build_strategy.reduce_strategy = F.BuildStrategy.ReduceStrategy.Reduce
+
+        log.info("start build compile program...")
+        compiled_prog = F.compiler.CompiledProgram(tfleet.main_program
+            ).with_data_parallel(
+                loss_name=loss.name,
+                build_strategy=build_strategy,
+                exec_strategy=exec_strategy)
+
+        return compiled_prog
+
+
+class CollectiveLearner(Learner):
+    def __init__(self):
+        role = role_maker.PaddleCloudRoleMaker(is_collective=True)
+        cfleet.init(role)
+
+    def optimize(self, loss, optimizer_type, lr):
+        optimizer = F.optimizer.Adam(learning_rate=lr)
+        dist_strategy = DistributedStrategy()
+        optimizer = cfleet.distributed_optimizer(optimizer, strategy=dist_strategy)
+        _, param_grads = optimizer.minimize(loss, F.default_startup_program())
+    
+    def build(self, model, data_gen, config):
+        self.optimize(model.loss, config.optimizer_type, config.lr)
+        self.program = cfleet.main_program
+        gpu_id = int(os.getenv("FLAGS_selected_gpus", "0"))
+        place = F.CUDAPlace(gpu_id)
+        self.exe = F.Executor(place)
+        self.exe.run(F.default_startup_program())
+        self.warmstart(F.default_startup_program(), config.ckpt_path)
+        self.fleet = cfleet
+        model.data_loader.decorate_batch_generator(
+            data_gen, places=place)
+        self.config = config
+        self.model = model

examples/erniesage/local_run.sh

+#!/bin/bash 
+
+set -x
+config=${1:-"./config.yaml"}
+unset http_proxy https_proxy
+
+function parse_yaml {
+   local prefix=$2
+   local s='[[:space:]]*' w='[a-zA-Z0-9_]*' fs=$(echo @|tr @ '\034')
+   sed -ne "s|^\($s\):|\1|" \
+        -e "s|^\($s\)\($w\)$s:$s[\"']\(.*\)[\"']$s\$|\1$fs\2$fs\3|p" \
+        -e "s|^\($s\)\($w\)$s:$s\(.*\)$s\$|\1$fs\2$fs\3|p"  $1 |
+   awk -F$fs '{
+      indent = length($1)/2;
+      vname[indent] = $2;
+      for (i in vname) {if (i > indent) {delete vname[i]}}
+      if (length($3) > 0) {
+         vn=""; for (i=0; i<indent; i++) {vn=(vn)(vname[i])("_")}
+         printf("%s%s%s=\"%s\"\n", "'$prefix'",vn, $2, $3);
+      }
+   }'
+}
+
+transpiler_local_train(){
+    export PADDLE_TRAINERS_NUM=1
+    export PADDLE_PSERVERS_NUM=1
+    export PADDLE_PORT=6206
+    export PADDLE_PSERVERS="127.0.0.1"
+    export BASE="./local_dir"
+    echo `which python`
+    if [ -d ${BASE} ]; then
+        rm -rf ${BASE}
+    fi 
+    mkdir ${BASE}
+    rm job_id
+    for((i=0;i<${PADDLE_PSERVERS_NUM};i++))
+    do
+        echo "start ps server: ${i}"
+        TRAINING_ROLE="PSERVER" PADDLE_TRAINER_ID=${i} sh job.sh local $config \
+            &> $BASE/pserver.$i.log &
+        echo $! >> job_id
+    done
+    sleep 3s 
+    for((j=0;j<${PADDLE_TRAINERS_NUM};j++))
+    do
+        echo "start ps work: ${j}"
+        TRAINING_ROLE="TRAINER" PADDLE_TRAINER_ID=${j} sh job.sh local $config \
+        echo $! >> job_id
+    done
+}
+
+collective_local_train(){
+    export PATH=./python27-gcc482-gpu/bin/:$PATH
+    echo `which python`
+    python -m paddle.distributed.launch train.py --conf $config
+    python -m paddle.distributed.launch infer.py --conf $config
+}
+
+eval $(parse_yaml $config)
+unalias python
+
+python3 ./preprocessing/dump_graph.py -i $input_data -o $graph_path --encoding $encoding \
+    -l $max_seqlen --vocab_file $ernie_vocab_file
+
+if [[ $learner_type == "cpu" ]];then
+    transpiler_local_train
+fi
+if [[ $learner_type == "gpu" ]];then
+    collective_local_train
+fi

examples/erniesage/models/base.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 time
+import glob
+import os
+
+import numpy as np
+
+import pgl
+import paddle.fluid as F
+import paddle.fluid.layers as L
+
+from models import message_passing
+
+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)
+
+
+class BaseGraphWrapperBuilder(object):
+    def __init__(self, config):
+        self.config = config
+        self.node_feature_info = []
+        self.edge_feature_info = []
+
+    def __call__(self):
+        place = F.CPUPlace()
+        graph_wrappers = []
+        for i in range(self.config.num_layers):
+            # all graph have same node_feat_info
+            graph_wrappers.append(
+                pgl.graph_wrapper.GraphWrapper(
+                    "layer_%s" % i, place, node_feat=self.node_feature_info, edge_feat=self.edge_feature_info))
+        return graph_wrappers
+
+
+class GraphsageGraphWrapperBuilder(BaseGraphWrapperBuilder):
+    def __init__(self, config):
+        super(GraphsageGraphWrapperBuilder, self).__init__(config)
+        self.node_feature_info.append(('index', [None], np.dtype('int64')))
+
+
+class BaseGNNModel(object):
+    def __init__(self, config):
+        self.config = config
+        self.graph_wrapper_builder = self.gen_graph_wrapper_builder(config) 
+        self.net_fn = self.gen_net_fn(config)
+        self.feed_list_builder = self.gen_feed_list_builder(config)
+        self.data_loader_builder = self.gen_data_loader_builder(config)
+        self.loss_fn = self.gen_loss_fn(config)
+        self.build()
+
+
+    def gen_graph_wrapper_builder(self, config): 
+        return GraphsageGraphWrapperBuilder(config)
+
+    def gen_net_fn(self, config):
+        return BaseNet(config)
+
+    def gen_feed_list_builder(self, config):
+        return BaseFeedListBuilder(config) 
+
+    def gen_data_loader_builder(self, config):
+        return BaseDataLoaderBuilder(config)
+
+    def gen_loss_fn(self, config):
+        return BaseLoss(config)
+
+    def build(self):
+        self.graph_wrappers = self.graph_wrapper_builder()
+        self.inputs, self.outputs = self.net_fn(self.graph_wrappers)
+        self.feed_list = self.feed_list_builder(self.inputs, self.graph_wrappers)
+        self.data_loader = self.data_loader_builder(self.feed_list)
+        self.loss = self.loss_fn(self.outputs)
+
+class BaseFeedListBuilder(object):
+    def __init__(self, config):
+        self.config = config
+
+    def __call__(self, inputs, graph_wrappers):
+        feed_list = []
+        for i in range(len(graph_wrappers)):
+            feed_list.extend(graph_wrappers[i].holder_list)
+        feed_list.extend(inputs)
+        return feed_list
+
+
+class BaseDataLoaderBuilder(object):
+    def __init__(self, config):
+        self.config = config
+
+    def __call__(self, feed_list):
+        data_loader = F.io.PyReader(
+            feed_list=feed_list, capacity=20, use_double_buffer=True, iterable=True)
+        return data_loader
+
+
+
+class BaseNet(object):
+    def __init__(self, config):
+        self.config = config
+
+    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 build_inputs(self):
+        user_index = L.data(
+            "user_index", shape=[None], dtype="int64", append_batch_size=False)
+        item_index = L.data(
+            "item_index", shape=[None], dtype="int64", append_batch_size=False)
+        return [user_index, item_index]
+
+    def build_embedding(self, graph_wrappers, inputs=None):
+        num_embed = int(np.load(os.path.join(self.config.graph_path, "num_nodes.npy")))
+        is_sparse = self.config.trainer_type == "Transpiler"
+
+        embed = L.embedding(
+            input=L.reshape(graph_wrappers[0].node_feat['index'], [-1, 1]),
+            size=[num_embed, self.config.hidden_size],
+            is_sparse=is_sparse,
+            param_attr=F.ParamAttr(name="node_embedding", initializer=F.initializer.Uniform(
+                low=-1. / self.config.hidden_size,
+                high=1. / self.config.hidden_size)))
+        return embed
+
+    def gnn_layers(self, graph_wrappers, feature):
+        features = [feature]
+
+        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))
+            features.append(feature)
+        return features
+
+    def __call__(self, graph_wrappers):
+        inputs = self.build_inputs()
+        feature = self.build_embedding(graph_wrappers, inputs)
+        features = self.gnn_layers(graph_wrappers, feature)
+        outputs = [self.take_final_feature(features[-1], i, "final_fc") for i in inputs]
+        src_real_index = L.gather(graph_wrappers[0].node_feat['index'], inputs[0])
+        outputs.append(src_real_index)
+        return inputs, outputs
+
+class BaseLoss(object):
+    def __init__(self, config):
+        self.config = config
+
+    def __call__(self, outputs):
+        user_feat, item_feat = outputs[0], outputs[1]
+        loss_type = self.config.loss_type
+        # Calc Loss
+        if self.config.loss_type == "hinge":
+            pos = L.reduce_sum(user_feat * item_feat, -1, keep_dim=True) # [B, 1]
+            neg = L.matmul(user_feat, item_feat, transpose_y=True) # [B, B]
+            loss = L.reduce_mean(L.relu(neg - pos + self.config.margin))
+        elif self.config.loss_type == "softmax":
+            pass
+            # TODO
+            # pos = L.reduce_sum(user_feat * item_feat, -1, keep_dim=True) # [B, 1]
+            # neg = L.matmul(user_feat, neg_feat, transpose_y=True) # [B, B]
+            # logits = L.concat([pos, neg], -1) # [B, 1+B]
+            # labels = L.fill_constant_batch_size_like(logits, [-1, 1], "int64", 0)
+            # loss = L.reduce_mean(L.softmax_with_cross_entropy(logits, labels))
+        else:
+            raise ValueError
+        return loss

examples/erniesage/models/ernie.py

+"""Ernie
+"""
+from models.base  import BaseNet, BaseGNNModel 
+
+class Ernie(BaseNet):
+
+    def build_inputs(self):
+        inputs = super(Ernie, 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]
+
+    def build_embedding(self, graph_wrappers, 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="student_")
+        feature = ernie.get_pooled_output()
+        return feature
+
+    def __call__(self, graph_wrappers):
+        inputs = self.build_inputs()
+        feature = self.build_embedding(graph_wrappers, inputs[-1])
+        features = [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
+
+
+class ErnieModel(BaseGNNModel):
+    def gen_net_fn(self, config):
+        return Ernie(config)
+
+

examples/erniesage/models/ernie_model/ernie.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.
+"""Ernie model."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+from __future__ import unicode_literals
+
+import json
+import six
+import logging
+import paddle.fluid as fluid
+import paddle.fluid.layers as L
+
+from io import open
+
+from models.ernie_model.transformer_encoder import encoder, pre_process_layer
+from models.ernie_model.transformer_encoder import graph_encoder
+
+log = logging.getLogger(__name__)
+
+
+class ErnieConfig(object):
+    def __init__(self, config_path):
+        self._config_dict = self._parse(config_path)
+
+    def _parse(self, config_path):
+        try:
+            with open(config_path, 'r', encoding='utf8') as json_file:
+                config_dict = json.load(json_file)
+        except Exception:
+            raise IOError("Error in parsing Ernie model config file '%s'" %
+                          config_path)
+        else:
+            return config_dict
+
+    def __getitem__(self, key):
+        return self._config_dict.get(key, None)
+
+    def print_config(self):
+        for arg, value in sorted(six.iteritems(self._config_dict)):
+            log.info('%s: %s' % (arg, value))
+        log.info('------------------------------------------------')
+
+
+class ErnieModel(object):
+    def __init__(self,
+                 src_ids,
+                 sentence_ids,
+                 task_ids=None,
+                 config=None,
+                 weight_sharing=True,
+                 use_fp16=False,
+                 name=""):
+
+        self._set_config(config, name, weight_sharing)
+        input_mask = self._build_input_mask(src_ids)
+        position_ids = self._build_position_ids(src_ids)
+        self._build_model(src_ids, position_ids, sentence_ids, task_ids,
+                          input_mask)
+        self._debug_summary(input_mask)
+
+    def _debug_summary(self, input_mask):
+        #histogram
+        seqlen_before_pad = L.cast(
+            L.reduce_sum(
+                input_mask, dim=1), dtype='float32')
+        seqlen_after_pad = L.reduce_sum(
+            L.cast(
+                L.zeros_like(input_mask), dtype='float32') + 1.0, dim=1)
+        pad_num = seqlen_after_pad - seqlen_before_pad
+        pad_rate = pad_num / seqlen_after_pad
+
+    def _build_position_ids(self, src_ids):
+        d_shape = L.shape(src_ids)
+        d_seqlen = d_shape[1]
+        d_batch = d_shape[0]
+        position_ids = L.reshape(
+            L.range(
+                0, d_seqlen, 1, dtype='int32'), [1, d_seqlen, 1],
+            inplace=True)
+        position_ids = L.expand(position_ids, [d_batch, 1, 1])
+        position_ids = L.cast(position_ids, 'int64')
+        position_ids.stop_gradient = True
+        return position_ids
+
+    def _build_input_mask(self, src_ids):
+        zero = L.fill_constant([1], dtype='int64', value=0)
+        input_mask = L.logical_not(L.equal(src_ids,
+                                           zero))  # assume pad id == 0
+        input_mask = L.cast(input_mask, 'float')
+        input_mask.stop_gradient = True
+        return input_mask
+
+    def _set_config(self, config, name, weight_sharing):
+        self._emb_size = config['hidden_size']
+        self._n_layer = config['num_hidden_layers']
+        self._n_head = config['num_attention_heads']
+        self._voc_size = config['vocab_size']
+        self._max_position_seq_len = config['max_position_embeddings']
+        if config.get('sent_type_vocab_size'):
+            self._sent_types = config['sent_type_vocab_size']
+        else:
+            self._sent_types = config['type_vocab_size']
+
+        self._use_task_id = config['use_task_id']
+        if self._use_task_id:
+            self._task_types = config['task_type_vocab_size']
+        self._hidden_act = config['hidden_act']
+        self._postprocess_cmd = config.get('postprocess_cmd', 'dan')
+        self._preprocess_cmd = config.get('preprocess_cmd', '')
+        self._prepostprocess_dropout = config['hidden_dropout_prob']
+        self._attention_dropout = config['attention_probs_dropout_prob']
+        self._weight_sharing = weight_sharing
+        self.name = name
+
+        self._word_emb_name = self.name + "word_embedding"
+        self._pos_emb_name = self.name + "pos_embedding"
+        self._sent_emb_name = self.name + "sent_embedding"
+        self._task_emb_name = self.name + "task_embedding"
+        self._dtype = "float16" if config['use_fp16'] else "float32"
+        self._emb_dtype = "float32"
+
+        # Initialize all weigths by truncated normal initializer, and all biases
+        # will be initialized by constant zero by default.
+        self._param_initializer = fluid.initializer.TruncatedNormal(
+            scale=config['initializer_range'])
+
+    def _build_model(self, src_ids, position_ids, sentence_ids, task_ids,
+                     input_mask):
+
+        emb_out = self._build_embedding(src_ids, position_ids, sentence_ids,
+                                        task_ids)
+        self.input_mask = input_mask
+        self._enc_out, self.all_hidden, self.all_attn, self.all_ffn = encoder(
+            enc_input=emb_out,
+            input_mask=input_mask,
+            n_layer=self._n_layer,
+            n_head=self._n_head,
+            d_key=self._emb_size // self._n_head,
+            d_value=self._emb_size // self._n_head,
+            d_model=self._emb_size,
+            d_inner_hid=self._emb_size * 4,
+            prepostprocess_dropout=self._prepostprocess_dropout,
+            attention_dropout=self._attention_dropout,
+            relu_dropout=0,
+            hidden_act=self._hidden_act,
+            preprocess_cmd=self._preprocess_cmd,
+            postprocess_cmd=self._postprocess_cmd,
+            param_initializer=self._param_initializer,
+            name=self.name + 'encoder')
+        if self._dtype == "float16":
+            self._enc_out = fluid.layers.cast(
+                x=self._enc_out, dtype=self._emb_dtype)
+
+    def _build_embedding(self, src_ids, position_ids, sentence_ids, task_ids):
+        # padding id in vocabulary must be set to 0
+        emb_out = fluid.layers.embedding(
+            input=src_ids,
+            size=[self._voc_size, self._emb_size],
+            dtype=self._emb_dtype,
+            param_attr=fluid.ParamAttr(
+                name=self._word_emb_name, initializer=self._param_initializer),
+            is_sparse=False)
+
+        position_emb_out = fluid.layers.embedding(
+            input=position_ids,
+            size=[self._max_position_seq_len, self._emb_size],
+            dtype=self._emb_dtype,
+            param_attr=fluid.ParamAttr(
+                name=self._pos_emb_name, initializer=self._param_initializer))
+
+        sent_emb_out = fluid.layers.embedding(
+            sentence_ids,
+            size=[self._sent_types, self._emb_size],
+            dtype=self._emb_dtype,
+            param_attr=fluid.ParamAttr(
+                name=self._sent_emb_name, initializer=self._param_initializer))
+
+        self.all_emb = [emb_out, position_emb_out, sent_emb_out]
+        emb_out = emb_out + position_emb_out
+        emb_out = emb_out + sent_emb_out
+
+        if self._use_task_id:
+            task_emb_out = fluid.layers.embedding(
+                task_ids,
+                size=[self._task_types, self._emb_size],
+                dtype=self._emb_dtype,
+                param_attr=fluid.ParamAttr(
+                    name=self._task_emb_name,
+                    initializer=self._param_initializer))
+
+            emb_out = emb_out + task_emb_out
+
+        emb_out = pre_process_layer(
+            emb_out,
+            'nd',
+            self._prepostprocess_dropout,
+            name=self.name + 'pre_encoder')
+
+        if self._dtype == "float16":
+            emb_out = fluid.layers.cast(x=emb_out, dtype=self._dtype)
+        return emb_out
+
+    def get_sequence_output(self):
+        return self._enc_out
+
+    def get_pooled_output(self):
+        """Get the first feature of each sequence for classification"""
+        next_sent_feat = self._enc_out[:, 0, :]
+        #next_sent_feat = fluid.layers.slice(input=self._enc_out, axes=[1], starts=[0], ends=[1])
+        next_sent_feat = fluid.layers.fc(
+            input=next_sent_feat,
+            size=self._emb_size,
+            act="tanh",
+            param_attr=fluid.ParamAttr(
+                name=self.name + "pooled_fc.w_0",
+                initializer=self._param_initializer),
+            bias_attr=self.name + "pooled_fc.b_0")
+        return next_sent_feat
+
+    def get_lm_output(self, mask_label, mask_pos):
+        """Get the loss & accuracy for pretraining"""
+
+        mask_pos = fluid.layers.cast(x=mask_pos, dtype='int32')
+
+        # extract the first token feature in each sentence
+        self.next_sent_feat = self.get_pooled_output()
+        reshaped_emb_out = fluid.layers.reshape(
+            x=self._enc_out, shape=[-1, self._emb_size])
+        # extract masked tokens' feature
+        mask_feat = fluid.layers.gather(input=reshaped_emb_out, index=mask_pos)
+
+        # transform: fc
+        mask_trans_feat = fluid.layers.fc(
+            input=mask_feat,
+            size=self._emb_size,
+            act=self._hidden_act,
+            param_attr=fluid.ParamAttr(
+                name=self.name + 'mask_lm_trans_fc.w_0',
+                initializer=self._param_initializer),
+            bias_attr=fluid.ParamAttr(name=self.name + 'mask_lm_trans_fc.b_0'))
+
+        # transform: layer norm 
+        mask_trans_feat = fluid.layers.layer_norm(
+            mask_trans_feat,
+            begin_norm_axis=len(mask_trans_feat.shape) - 1,
+            param_attr=fluid.ParamAttr(
+                name=self.name + 'mask_lm_trans_layer_norm_scale',
+                initializer=fluid.initializer.Constant(1.)),
+            bias_attr=fluid.ParamAttr(
+                name=self.name + 'mask_lm_trans_layer_norm_bias',
+                initializer=fluid.initializer.Constant(0.)))
+        # transform: layer norm 
+        #mask_trans_feat = pre_process_layer(
+        #    mask_trans_feat, 'n', name=self.name + 'mask_lm_trans')
+
+        mask_lm_out_bias_attr = fluid.ParamAttr(
+            name=self.name + "mask_lm_out_fc.b_0",
+            initializer=fluid.initializer.Constant(value=0.0))
+        if self._weight_sharing:
+            fc_out = fluid.layers.matmul(
+                x=mask_trans_feat,
+                y=fluid.default_main_program().global_block().var(
+                    self._word_emb_name),
+                transpose_y=True)
+            fc_out += fluid.layers.create_parameter(
+                shape=[self._voc_size],
+                dtype=self._emb_dtype,
+                attr=mask_lm_out_bias_attr,
+                is_bias=True)
+
+        else:
+            fc_out = fluid.layers.fc(input=mask_trans_feat,
+                                     size=self._voc_size,
+                                     param_attr=fluid.ParamAttr(
+                                         name=self.name + "mask_lm_out_fc.w_0",
+                                         initializer=self._param_initializer),
+                                     bias_attr=mask_lm_out_bias_attr)
+
+        mask_lm_loss = fluid.layers.softmax_with_cross_entropy(
+            logits=fc_out, label=mask_label)
+        return mask_lm_loss
+
+    def get_task_output(self, task, task_labels):
+        task_fc_out = fluid.layers.fc(
+            input=self.next_sent_feat,
+            size=task["num_labels"],
+            param_attr=fluid.ParamAttr(
+                name=self.name + task["task_name"] + "_fc.w_0",
+                initializer=self._param_initializer),
+            bias_attr=self.name + task["task_name"] + "_fc.b_0")
+        task_loss, task_softmax = fluid.layers.softmax_with_cross_entropy(
+            logits=task_fc_out, label=task_labels, return_softmax=True)
+        task_acc = fluid.layers.accuracy(input=task_softmax, label=task_labels)
+        return task_loss, task_acc
+
+
+class ErnieGraphModel(ErnieModel):
+    def __init__(self,
+                 src_ids,
+                 task_ids=None,
+                 config=None,
+                 weight_sharing=True,
+                 use_fp16=False,
+                 slot_seqlen=40,
+                 name=""):
+        self.slot_seqlen = slot_seqlen
+        self._set_config(config, name, weight_sharing)
+        input_mask = self._build_input_mask(src_ids)
+        position_ids = self._build_position_ids(src_ids)
+        sentence_ids = self._build_sentence_ids(src_ids)
+        self._build_model(src_ids, position_ids, sentence_ids, task_ids,
+                          input_mask)
+        self._debug_summary(input_mask)
+
+    def _build_position_ids(self, src_ids):
+        src_shape = L.shape(src_ids)
+        src_seqlen = src_shape[1]
+        src_batch = src_shape[0]
+
+        slot_seqlen = self.slot_seqlen
+
+        num_b = (src_seqlen / slot_seqlen) - 1
+        a_position_ids = L.reshape(
+            L.range(
+                0, slot_seqlen, 1, dtype='int32'), [1, slot_seqlen, 1],
+            inplace=True) # [1, slot_seqlen, 1]
+        a_position_ids = L.expand(a_position_ids, [src_batch, 1, 1]) # [B, slot_seqlen * num_b, 1]
+
+        zero = L.fill_constant([1], dtype='int64', value=0)
+        input_mask = L.cast(L.equal(src_ids[:, :slot_seqlen], zero), "int32")  # assume pad id == 0 [B, slot_seqlen, 1]
+        a_pad_len = L.reduce_sum(input_mask, 1) # [B, 1, 1]
+
+        b_position_ids = L.reshape(
+            L.range(
+                slot_seqlen, 2*slot_seqlen, 1, dtype='int32'), [1, slot_seqlen, 1],
+            inplace=True) # [1, slot_seqlen, 1]
+        b_position_ids = L.expand(b_position_ids, [src_batch, num_b, 1]) # [B, slot_seqlen * num_b, 1]
+        b_position_ids = b_position_ids - a_pad_len # [B, slot_seqlen * num_b, 1]
+
+        position_ids = L.concat([a_position_ids, b_position_ids], 1)
+        position_ids = L.cast(position_ids, 'int64')
+        position_ids.stop_gradient = True
+        return position_ids
+
+    def _build_sentence_ids(self, src_ids):
+        src_shape = L.shape(src_ids)
+        src_seqlen = src_shape[1]
+        src_batch = src_shape[0]
+
+        slot_seqlen = self.slot_seqlen
+
+        zeros = L.zeros([src_batch, slot_seqlen, 1], "int64")
+        ones = L.ones([src_batch, src_seqlen-slot_seqlen, 1], "int64")
+
+        sentence_ids = L.concat([zeros, ones], 1)
+        sentence_ids.stop_gradient = True
+        return sentence_ids
+
+    def _build_model(self, src_ids, position_ids, sentence_ids, task_ids,
+                     input_mask):
+
+        emb_out = self._build_embedding(src_ids, position_ids, sentence_ids,
+                                        task_ids)
+        self.input_mask = input_mask
+        self._enc_out, self.all_hidden, self.all_attn, self.all_ffn = graph_encoder(
+            enc_input=emb_out,
+            input_mask=input_mask,
+            n_layer=self._n_layer,
+            n_head=self._n_head,
+            d_key=self._emb_size // self._n_head,
+            d_value=self._emb_size // self._n_head,
+            d_model=self._emb_size,
+            d_inner_hid=self._emb_size * 4,
+            prepostprocess_dropout=self._prepostprocess_dropout,
+            attention_dropout=self._attention_dropout,
+            relu_dropout=0,
+            hidden_act=self._hidden_act,
+            preprocess_cmd=self._preprocess_cmd,
+            postprocess_cmd=self._postprocess_cmd,
+            param_initializer=self._param_initializer,
+            slot_seqlen=self.slot_seqlen,
+            name=self.name + 'encoder')
+        if self._dtype == "float16":
+            self._enc_out = fluid.layers.cast(
+                x=self._enc_out, dtype=self._emb_dtype)

examples/erniesage/models/ernie_model/transformer_encoder.py

+#   Copyright (c) 2018 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.
+
+from functools import partial
+import numpy as np
+from contextlib import contextmanager
+
+import paddle.fluid as fluid
+import paddle.fluid.layers as L
+import paddle.fluid.layers as layers
+#import propeller.paddle as propeller
+#from propeller import log
+
+#determin this at the begining
+to_3d = lambda a: a  # will change later
+to_2d = lambda a: a
+
+
+def multi_head_attention(queries,
+                         keys,
+                         values,
+                         attn_bias,
+                         d_key,
+                         d_value,
+                         d_model,
+                         n_head=1,
+                         dropout_rate=0.,
+                         cache=None,
+                         param_initializer=None,
+                         name='multi_head_att'):
+    """
+    Multi-Head Attention. Note that attn_bias is added to the logit before
+    computing softmax activiation to mask certain selected positions so that
+    they will not considered in attention weights.
+    """
+    keys = queries if keys is None else keys
+    values = keys if values is None else values
+
+    def __compute_qkv(queries, keys, values, n_head, d_key, d_value):
+        """
+        Add linear projection to queries, keys, and values.
+        """
+        q = layers.fc(input=queries,
+                      size=d_key * n_head,
+                      num_flatten_dims=len(queries.shape) - 1,
+                      param_attr=fluid.ParamAttr(
+                          name=name + '_query_fc.w_0',
+                          initializer=param_initializer),
+                      bias_attr=name + '_query_fc.b_0')
+        k = layers.fc(input=keys,
+                      size=d_key * n_head,
+                      num_flatten_dims=len(keys.shape) - 1,
+                      param_attr=fluid.ParamAttr(
+                          name=name + '_key_fc.w_0',
+                          initializer=param_initializer),
+                      bias_attr=name + '_key_fc.b_0')
+        v = layers.fc(input=values,
+                      size=d_value * n_head,
+                      num_flatten_dims=len(values.shape) - 1,
+                      param_attr=fluid.ParamAttr(
+                          name=name + '_value_fc.w_0',
+                          initializer=param_initializer),
+                      bias_attr=name + '_value_fc.b_0')
+        return q, k, v
+
+    def __split_heads(x, n_head):
+        """
+        Reshape the last dimension of inpunt tensor x so that it becomes two
+        dimensions and then transpose. Specifically, input a tensor with shape
+        [bs, max_sequence_length, n_head * hidden_dim] then output a tensor
+        with shape [bs, n_head, max_sequence_length, hidden_dim].
+        """
+        hidden_size = x.shape[-1]
+        # The value 0 in shape attr means copying the corresponding dimension
+        # size of the input as the output dimension size.
+        reshaped = layers.reshape(
+            x=x, shape=[0, 0, n_head, hidden_size // n_head], inplace=True)
+
+        # permuate the dimensions into:
+        # [batch_size, n_head, max_sequence_len, hidden_size_per_head]
+        return layers.transpose(x=reshaped, perm=[0, 2, 1, 3])
+
+    def __combine_heads(x):
+        """
+        Transpose and then reshape the last two dimensions of inpunt tensor x
+        so that it becomes one dimension, which is reverse to __split_heads.
+        """
+        if len(x.shape) == 3: return x
+        if len(x.shape) != 4:
+            raise ValueError("Input(x) should be a 4-D Tensor.")
+        trans_x = layers.transpose(x, perm=[0, 2, 1, 3])
+        # The value 0 in shape attr means copying the corresponding dimension
+        # size of the input as the output dimension size.
+        #trans_x.desc.set_shape((-1, 1, n_head, d_value))
+        return layers.reshape(x=trans_x, shape=[0, 0, d_model], inplace=True)
+
+    def scaled_dot_product_attention(q, k, v, attn_bias, d_key, dropout_rate):
+        """
+        Scaled Dot-Product Attention
+        """
+        scaled_q = layers.scale(x=q, scale=d_key**-0.5)
+        product = layers.matmul(x=scaled_q, y=k, transpose_y=True)
+        if attn_bias:
+            product += attn_bias
+        weights = layers.softmax(product)
+        if dropout_rate:
+            weights = layers.dropout(
+                weights,
+                dropout_prob=dropout_rate,
+                dropout_implementation="upscale_in_train",
+                is_test=False)
+        out = layers.matmul(weights, v)
+        #return out, product
+        return out, weights
+
+    q, k, v = __compute_qkv(queries, keys, values, n_head, d_key, d_value)
+    q = to_3d(q)
+    k = to_3d(k)
+    v = to_3d(v)
+
+    if cache is not None:  # use cache and concat time steps
+        # Since the inplace reshape in __split_heads changes the shape of k and
+        # v, which is the cache input for next time step, reshape the cache
+        # input from the previous time step first.
+        k = cache["k"] = layers.concat(
+            [layers.reshape(
+                cache["k"], shape=[0, 0, d_model]), k], axis=1)
+        v = cache["v"] = layers.concat(
+            [layers.reshape(
+                cache["v"], shape=[0, 0, d_model]), v], axis=1)
+
+    q = __split_heads(q, n_head)
+    k = __split_heads(k, n_head)
+    v = __split_heads(v, n_head)
+
+    ctx_multiheads, ctx_multiheads_attn = scaled_dot_product_attention(
+        q, k, v, attn_bias, d_key, dropout_rate)
+
+    out = __combine_heads(ctx_multiheads)
+
+    out = to_2d(out)
+
+    # Project back to the model size.
+    proj_out = layers.fc(input=out,
+                         size=d_model,
+                         num_flatten_dims=len(out.shape) - 1,
+                         param_attr=fluid.ParamAttr(
+                             name=name + '_output_fc.w_0',
+                             initializer=param_initializer),
+                         bias_attr=name + '_output_fc.b_0')
+    return proj_out, ctx_multiheads_attn
+
+
+def positionwise_feed_forward(x,
+                              d_inner_hid,
+                              d_hid,
+                              dropout_rate,
+                              hidden_act,
+                              param_initializer=None,
+                              name='ffn'):
+    """
+    Position-wise Feed-Forward Networks.
+    This module consists of two linear transformations with a ReLU activation
+    in between, which is applied to each position separately and identically.
+    """
+    hidden = layers.fc(input=x,
+                       size=d_inner_hid,
+                       num_flatten_dims=len(x.shape) - 1,
+                       act=hidden_act,
+                       param_attr=fluid.ParamAttr(
+                           name=name + '_fc_0.w_0',
+                           initializer=param_initializer),
+                       bias_attr=name + '_fc_0.b_0')
+    if dropout_rate:
+        hidden = layers.dropout(
+            hidden,
+            dropout_prob=dropout_rate,
+            dropout_implementation="upscale_in_train",
+            is_test=False)
+    out = layers.fc(input=hidden,
+                    size=d_hid,
+                    num_flatten_dims=len(hidden.shape) - 1,
+                    param_attr=fluid.ParamAttr(
+                        name=name + '_fc_1.w_0',
+                        initializer=param_initializer),
+                    bias_attr=name + '_fc_1.b_0')
+    return out
+
+
+def pre_post_process_layer(prev_out,
+                           out,
+                           process_cmd,
+                           dropout_rate=0.,
+                           name=''):
+    """
+    Add residual connection, layer normalization and droput to the out tensor
+    optionally according to the value of process_cmd.
+    This will be used before or after multi-head attention and position-wise
+    feed-forward networks.
+    """
+    for cmd in process_cmd:
+        if cmd == "a":  # add residual connection
+            out = out + prev_out if prev_out else out
+        elif cmd == "n":  # add layer normalization
+            out_dtype = out.dtype
+            if out_dtype == fluid.core.VarDesc.VarType.FP16:
+                out = layers.cast(x=out, dtype="float32")
+            out = layers.layer_norm(
+                out,
+                begin_norm_axis=len(out.shape) - 1,
+                param_attr=fluid.ParamAttr(
+                    name=name + '_layer_norm_scale',
+                    initializer=fluid.initializer.Constant(1.)),
+                bias_attr=fluid.ParamAttr(
+                    name=name + '_layer_norm_bias',
+                    initializer=fluid.initializer.Constant(0.)))
+            if out_dtype == fluid.core.VarDesc.VarType.FP16:
+                out = layers.cast(x=out, dtype="float16")
+        elif cmd == "d":  # add dropout
+            if dropout_rate:
+                out = layers.dropout(
+                    out,
+                    dropout_prob=dropout_rate,
+                    dropout_implementation="upscale_in_train",
+                    is_test=False)
+    return out
+
+
+pre_process_layer = partial(pre_post_process_layer, None)
+post_process_layer = pre_post_process_layer
+
+
+def encoder_layer(enc_input,
+                  attn_bias,
+                  n_head,
+                  d_key,
+                  d_value,
+                  d_model,
+                  d_inner_hid,
+                  prepostprocess_dropout,
+                  attention_dropout,
+                  relu_dropout,
+                  hidden_act,
+                  preprocess_cmd="n",
+                  postprocess_cmd="da",
+                  param_initializer=None,
+                  name=''):
+    """The encoder layers that can be stacked to form a deep encoder.
+    This module consits of a multi-head (self) attention followed by
+    position-wise feed-forward networks and both the two components companied
+    with the post_process_layer to add residual connection, layer normalization
+    and droput.
+    """
+    #L.Print(L.reduce_mean(enc_input), message='1')
+    attn_output, ctx_multiheads_attn = multi_head_attention(
+        pre_process_layer(
+            enc_input,
+            preprocess_cmd,
+            prepostprocess_dropout,
+            name=name + '_pre_att'),
+        None,
+        None,
+        attn_bias,
+        d_key,
+        d_value,
+        d_model,
+        n_head,
+        attention_dropout,
+        param_initializer=param_initializer,
+        name=name + '_multi_head_att')
+    #L.Print(L.reduce_mean(attn_output), message='1')
+    attn_output = post_process_layer(
+        enc_input,
+        attn_output,
+        postprocess_cmd,
+        prepostprocess_dropout,
+        name=name + '_post_att')
+
+    #L.Print(L.reduce_mean(attn_output), message='2')
+    ffd_output = positionwise_feed_forward(
+        pre_process_layer(
+            attn_output,
+            preprocess_cmd,
+            prepostprocess_dropout,
+            name=name + '_pre_ffn'),
+        d_inner_hid,
+        d_model,
+        relu_dropout,
+        hidden_act,
+        param_initializer=param_initializer,
+        name=name + '_ffn')
+    #L.Print(L.reduce_mean(ffd_output), message='3')
+    ret = post_process_layer(
+        attn_output,
+        ffd_output,
+        postprocess_cmd,
+        prepostprocess_dropout,
+        name=name + '_post_ffn')
+    #L.Print(L.reduce_mean(ret), message='4')
+    return ret, ctx_multiheads_attn, ffd_output
+
+
+def build_pad_idx(input_mask):
+    pad_idx = L.where(L.cast(L.squeeze(input_mask, [2]), 'bool'))
+    return pad_idx
+
+
+def build_attn_bias(input_mask, n_head, dtype):
+    attn_bias = L.matmul(
+        input_mask, input_mask, transpose_y=True)  # [batch, seq, seq]
+    attn_bias = (1. - attn_bias) * -10000.
+    attn_bias = L.stack([attn_bias] * n_head, 1) # [batch, n_head, seq, seq]
+    if attn_bias.dtype != dtype:
+        attn_bias = L.cast(attn_bias, dtype)
+    return attn_bias
+
+
+def build_graph_attn_bias(input_mask, n_head, dtype, slot_seqlen):
+
+    input_shape = L.shape(input_mask)
+    input_batch = input_shape[0]
+    input_seqlen = input_shape[1]
+    num_slot = input_seqlen / slot_seqlen
+    num_b = num_slot - 1
+    ones = L.ones([num_b], dtype="float32") # [num_b]
+    diag_ones = L.diag(ones) # [num_b, num_b]
+    diag_ones = L.unsqueeze(diag_ones, [1, -1]) # [num_b, 1, num_b, 1]
+    diag_ones = L.expand(diag_ones, [1, slot_seqlen, 1, slot_seqlen]) # [num_b, seqlen, num_b, seqlen]
+    diag_ones = L.reshape(diag_ones, [1, num_b*slot_seqlen, num_b*slot_seqlen]) # [1, num_b*seqlen, num_b*seqlen]
+    
+    graph_attn_bias = L.concat([L.ones([1, num_b*slot_seqlen, slot_seqlen], dtype="float32"), diag_ones], 2)
+    graph_attn_bias = L.concat([L.ones([1, slot_seqlen, num_slot*slot_seqlen], dtype="float32"), graph_attn_bias], 1) # [1, seq, seq]
+
+    pad_attn_bias = L.matmul(
+        input_mask, input_mask, transpose_y=True)  # [batch, seq, seq]
+    attn_bias = graph_attn_bias * pad_attn_bias
+
+    attn_bias = (1. - attn_bias) * -10000.
+    attn_bias = L.stack([attn_bias] * n_head, 1) # [batch, n_head, seq, seq]
+    if attn_bias.dtype != dtype:
+        attn_bias = L.cast(attn_bias, dtype)
+    return attn_bias
+
+
+def encoder(enc_input,
+            input_mask,
+            n_layer,
+            n_head,
+            d_key,
+            d_value,
+            d_model,
+            d_inner_hid,
+            prepostprocess_dropout,
+            attention_dropout,
+            relu_dropout,
+            hidden_act,
+            preprocess_cmd="n",
+            postprocess_cmd="da",
+            param_initializer=None,
+            name=''):
+    """
+    The encoder is composed of a stack of identical layers returned by calling
+    encoder_layer.
+    """
+
+    #global to_2d, to_3d  #, batch, seqlen, dynamic_dim
+    d_shape = L.shape(input_mask)
+    pad_idx = build_pad_idx(input_mask)
+    attn_bias = build_attn_bias(input_mask, n_head, enc_input.dtype)
+
+    # d_batch = d_shape[0]
+    # d_seqlen = d_shape[1]
+    # pad_idx = L.where(
+    # L.cast(L.reshape(input_mask, [d_batch, d_seqlen]), 'bool'))
+
+    # attn_bias = L.matmul(
+    # input_mask, input_mask, transpose_y=True)  # [batch, seq, seq]
+    # attn_bias = (1. - attn_bias) * -10000.
+    # attn_bias = L.stack([attn_bias] * n_head, 1)
+    # if attn_bias.dtype != enc_input.dtype:
+    # attn_bias = L.cast(attn_bias, enc_input.dtype)
+
+    # def to_2d(t_3d):
+        # t_2d = L.gather_nd(t_3d, pad_idx)
+        # return t_2d
+
+    # def to_3d(t_2d):
+        # t_3d = L.scatter_nd(
+        # pad_idx, t_2d, shape=[d_shape[0], d_shape[1], d_model])
+        # return t_3d
+
+    enc_input = to_2d(enc_input)
+    all_hidden = []
+    all_attn = []
+    all_ffn = []
+    for i in range(n_layer):
+        enc_output, ctx_multiheads_attn, ffn_output = encoder_layer(
+            enc_input,
+            attn_bias,
+            n_head,
+            d_key,
+            d_value,
+            d_model,
+            d_inner_hid,
+            prepostprocess_dropout,
+            attention_dropout,
+            relu_dropout,
+            hidden_act,
+            preprocess_cmd,
+            postprocess_cmd,
+            param_initializer=param_initializer,
+            name=name + '_layer_' + str(i))
+        all_hidden.append(enc_output)
+        all_attn.append(ctx_multiheads_attn)
+        all_ffn.append(ffn_output)
+        enc_input = enc_output
+    enc_output = pre_process_layer(
+        enc_output,
+        preprocess_cmd,
+        prepostprocess_dropout,
+        name="post_encoder")
+    enc_output = to_3d(enc_output)
+    #enc_output.desc.set_shape((-1, 1, final_dim))
+    return enc_output, all_hidden, all_attn, all_ffn
+
+def graph_encoder(enc_input,
+            input_mask,
+            n_layer,
+            n_head,
+            d_key,
+            d_value,
+            d_model,
+            d_inner_hid,
+            prepostprocess_dropout,
+            attention_dropout,
+            relu_dropout,
+            hidden_act,
+            preprocess_cmd="n",
+            postprocess_cmd="da",
+            param_initializer=None,
+            slot_seqlen=40,
+            name=''):
+    """
+    The encoder is composed of a stack of identical layers returned by calling
+    encoder_layer.
+    """
+
+    #global to_2d, to_3d  #, batch, seqlen, dynamic_dim
+    d_shape = L.shape(input_mask)
+    pad_idx = build_pad_idx(input_mask)
+    attn_bias = build_graph_attn_bias(input_mask, n_head, enc_input.dtype, slot_seqlen)
+    #attn_bias = build_attn_bias(input_mask, n_head, enc_input.dtype)
+
+    # d_batch = d_shape[0]
+    # d_seqlen = d_shape[1]
+    # pad_idx = L.where(
+    # L.cast(L.reshape(input_mask, [d_batch, d_seqlen]), 'bool'))
+
+    # attn_bias = L.matmul(
+    # input_mask, input_mask, transpose_y=True)  # [batch, seq, seq]
+    # attn_bias = (1. - attn_bias) * -10000.
+    # attn_bias = L.stack([attn_bias] * n_head, 1)
+    # if attn_bias.dtype != enc_input.dtype:
+    # attn_bias = L.cast(attn_bias, enc_input.dtype)
+
+    # def to_2d(t_3d):
+        # t_2d = L.gather_nd(t_3d, pad_idx)
+        # return t_2d
+
+    # def to_3d(t_2d):
+        # t_3d = L.scatter_nd(
+        # pad_idx, t_2d, shape=[d_shape[0], d_shape[1], d_model])
+        # return t_3d
+
+    enc_input = to_2d(enc_input)
+    all_hidden = []
+    all_attn = []
+    all_ffn = []
+    for i in range(n_layer):
+        enc_output, ctx_multiheads_attn, ffn_output = encoder_layer(
+            enc_input,
+            attn_bias,
+            n_head,
+            d_key,
+            d_value,
+            d_model,
+            d_inner_hid,
+            prepostprocess_dropout,
+            attention_dropout,
+            relu_dropout,
+            hidden_act,
+            preprocess_cmd,
+            postprocess_cmd,
+            param_initializer=param_initializer,
+            name=name + '_layer_' + str(i))
+        all_hidden.append(enc_output)
+        all_attn.append(ctx_multiheads_attn)
+        all_ffn.append(ffn_output)
+        enc_input = enc_output
+    enc_output = pre_process_layer(
+        enc_output,
+        preprocess_cmd,
+        prepostprocess_dropout,
+        name="post_encoder")
+    enc_output = to_3d(enc_output)
+    #enc_output.desc.set_shape((-1, 1, final_dim))
+    return enc_output, all_hidden, all_attn, all_ffn

examples/erniesage/models/erniesage_v1.py

+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
+
+class ErnieSageV1(BaseNet):
+
+    def build_inputs(self):
+        inputs = super(ErnieSageV1, 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]
+
+    def build_embedding(self, graph_wrappers, 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="student_")
+        feature = ernie.get_pooled_output()
+        return feature
+
+    def __call__(self, graph_wrappers):
+        inputs = self.build_inputs()
+        feature = self.build_embedding(graph_wrappers, 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
+
+
+class ErnieSageModelV1(BaseGNNModel):
+    def gen_net_fn(self, config):
+        return ErnieSageV1(config)

examples/erniesage/models/erniesage_v2.py

+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
+
+
+class ErnieSageV2(BaseNet):
+
+    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]
+
+    def gnn_layer(self, gw, feature, hidden_size, act, initializer, learning_rate, name):
+        def ernie_send(src_feat, dst_feat, edge_feat):
+            """doc"""
+            cls = L.fill_constant_batch_size_like(src_feat["term_ids"], [-1, 1, 1], "int64", 1)
+            src_ids = L.concat([cls, src_feat["term_ids"]], 1)
+            dst_ids = dst_feat["term_ids"]
+
+            sent_ids = L.concat([L.zeros_like(src_ids), L.ones_like(dst_ids)], 1)
+            term_ids = L.concat([src_ids, dst_ids], 1)
+
+            term_ids.stop_gradient = True
+            sent_ids.stop_gradient = True
+            ernie = ErnieModel(
+                term_ids, sent_ids,
+                config=self.config.ernie_config)
+            feature = ernie.get_pooled_output()
+            return feature
+
+        def erniesage_v2_aggregator(gw, feature, hidden_size, act, initializer, learning_rate, name):
+            feature = L.unsqueeze(feature, [-1])
+            msg = gw.send(ernie_send, nfeat_list=[("term_ids", feature)])
+            neigh_feature = gw.recv(msg, lambda feat: F.layers.sequence_pool(feat, pool_type="sum"))
+
+            term_ids = feature
+            cls = L.fill_constant_batch_size_like(term_ids, [-1, 1, 1], "int64", 1)
+            term_ids = L.concat([cls, term_ids], 1)
+            term_ids.stop_gradient = True
+            ernie = ErnieModel(
+                term_ids, L.zeros_like(term_ids),
+                config=self.config.ernie_config)
+            self_feature = ernie.get_pooled_output()
+
+            self_feature = L.fc(self_feature,
+                                           hidden_size,
+                                           act=act,
+                                           param_attr=F.ParamAttr(name=name + "_l",
+                                           learning_rate=learning_rate),
+                                           )
+            neigh_feature = L.fc(neigh_feature,
+                                            hidden_size,
+                                            act=act,
+                                            param_attr=F.ParamAttr(name=name + "_r",
+                                           learning_rate=learning_rate),
+                                            )
+            output = L.concat([self_feature, neigh_feature], axis=1)
+            output = L.l2_normalize(output, axis=1)
+            return output
+        return erniesage_v2_aggregator(gw, feature, hidden_size, act, initializer, learning_rate, name)
+
+    def gnn_layers(self, graph_wrappers, feature):
+        features = [feature]
+
+        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 = self.gnn_layer(
+                graph_wrappers[i],
+                feature,
+                self.config.hidden_size,
+                act,
+                initializer,
+                learning_rate=fc_lr,
+                name="%s_%s" % ("erniesage_v2", 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
+
+
+class ErnieSageModelV2(BaseGNNModel):
+    def gen_net_fn(self, config):
+        return ErnieSageV2(config)

examples/erniesage/models/erniesage_v3.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 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.message_passing import copy_send
+
+
+class ErnieSageV3(BaseNet):
+    def __init__(self, config):
+        super(ErnieSageV3, self).__init__(config)
+        self.config.layer_type = "ernie_recv_sum"
+
+    def build_inputs(self):
+        inputs = super(ErnieSageV3, 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]
+
+    def gnn_layer(self, gw, feature, hidden_size, act, initializer, learning_rate, name):
+        def ernie_recv(feat):
+            """doc"""
+            # TODO maxlen  400
+            #pad_value = L.cast(L.assign(input=np.array([0], dtype=np.int32)), "int64")
+            pad_value = L.zeros([1], "int64")
+            out, _ = L.sequence_pad(feat, pad_value=pad_value, maxlen=10)
+            out = L.reshape(out, [0, 400])
+            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]
+
+        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 = self.gnn_layer(
+                graph_wrappers[i],
+                feature,
+                self.config.hidden_size,
+                act,
+                initializer,
+                learning_rate=fc_lr,
+                name="%s_%s" % (self.config.layer_type, i))
+            features.append(feature)
+        return features
+
+    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,
+            name="student_")
+        feat = ernie.get_pooled_output()
+        fc_lr = self.config.lr / 0.001
+        feat= L.fc(feat,
+                   self.config.hidden_size,
+                   act="relu",
+                   param_attr=F.ParamAttr(name=name + "_l",
+                   learning_rate=fc_lr),
+                   )
+        feat = L.l2_normalize(feat, axis=1)
+
+        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 __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
+    
+
+class ErnieSageModelV3(BaseGNNModel):
+    def gen_net_fn(self, config):
+        return ErnieSageV3(config)

examples/erniesage/models/message_passing.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 numpy as np
+import paddle
+import paddle.fluid as fluid
+import paddle.fluid.layers as L
+
+
+def copy_send(src_feat, dst_feat, edge_feat):
+    """doc"""
+    return src_feat["h"]
+
+def weighted_copy_send(src_feat, dst_feat, edge_feat):
+    """doc"""
+    return src_feat["h"] * edge_feat["weight"]
+
+def mean_recv(feat):
+    """doc"""
+    return fluid.layers.sequence_pool(feat, pool_type="average")
+
+
+def sum_recv(feat):
+    """doc"""
+    return fluid.layers.sequence_pool(feat, pool_type="sum")
+
+
+def max_recv(feat):
+    """doc"""
+    return fluid.layers.sequence_pool(feat, pool_type="max")
+
+
+def lstm_recv(feat):
+    """doc"""
+    hidden_dim = 128
+    forward, _ = fluid.layers.dynamic_lstm(
+        input=feat, size=hidden_dim * 4, use_peepholes=False)
+    output = fluid.layers.sequence_last_step(forward)
+    return output
+
+
+def graphsage_sum(gw, feature, hidden_size, act, initializer, learning_rate, name):
+    """doc"""
+    msg = gw.send(copy_send, nfeat_list=[("h", feature)])
+    neigh_feature = gw.recv(msg, sum_recv)
+    self_feature = feature
+    self_feature = fluid.layers.fc(self_feature,
+                                   hidden_size,
+                                   act=act,
+                                   param_attr=fluid.ParamAttr(name=name + "_l", initializer=initializer,
+                                   learning_rate=learning_rate),
+                                   )
+    neigh_feature = fluid.layers.fc(neigh_feature,
+                                    hidden_size,
+                                    act=act,
+                                    param_attr=fluid.ParamAttr(name=name + "_r", initializer=initializer,
+                                   learning_rate=learning_rate),
+                                    )
+    output = fluid.layers.concat([self_feature, neigh_feature], axis=1)
+    output = fluid.layers.l2_normalize(output, axis=1)
+    return output
+
+
+def graphsage_mean(gw, feature, hidden_size, act, initializer, learning_rate, name):
+    """doc"""
+    msg = gw.send(copy_send, nfeat_list=[("h", feature)])
+    neigh_feature = gw.recv(msg, mean_recv)
+    self_feature = feature
+    self_feature = fluid.layers.fc(self_feature,
+                                   hidden_size,
+                                   act=act,
+                                   param_attr=fluid.ParamAttr(name=name + "_l", initializer=initializer,
+                                   learning_rate=learning_rate),
+                                   )
+    neigh_feature = fluid.layers.fc(neigh_feature,
+                                    hidden_size,
+                                    act=act,
+                                    param_attr=fluid.ParamAttr(name=name + "_r", initializer=initializer,
+                                   learning_rate=learning_rate),
+                                    )
+    output = fluid.layers.concat([self_feature, neigh_feature], axis=1)
+    output = fluid.layers.l2_normalize(output, axis=1)
+    return output
+
+
+def pinsage_mean(gw, feature, hidden_size, act, initializer, learning_rate, name):
+    """doc"""
+    msg = gw.send(weighted_copy_send, nfeat_list=[("h", feature)], efeat_list=["weight"])
+    neigh_feature = gw.recv(msg, mean_recv)
+    self_feature = feature
+    self_feature = fluid.layers.fc(self_feature,
+                                   hidden_size,
+                                   act=act,
+                                   param_attr=fluid.ParamAttr(name=name + "_l", initializer=initializer,
+                                   learning_rate=learning_rate),
+                                   )
+    neigh_feature = fluid.layers.fc(neigh_feature,
+                                    hidden_size,
+                                    act=act,
+                                    param_attr=fluid.ParamAttr(name=name + "_r", initializer=initializer,
+                                   learning_rate=learning_rate),
+                                    )
+    output = fluid.layers.concat([self_feature, neigh_feature], axis=1)
+    output = fluid.layers.l2_normalize(output, axis=1)
+    return output
+
+
+def pinsage_sum(gw, feature, hidden_size, act, initializer, learning_rate, name):
+    """doc"""
+    msg = gw.send(weighted_copy_send, nfeat_list=[("h", feature)], efeat_list=["weight"])
+    neigh_feature = gw.recv(msg, sum_recv)
+    self_feature = feature
+    self_feature = fluid.layers.fc(self_feature,
+                                   hidden_size,
+                                   act=act,
+                                   param_attr=fluid.ParamAttr(name=name + "_l", initializer=initializer,
+                                   learning_rate=learning_rate),
+                                   )
+    neigh_feature = fluid.layers.fc(neigh_feature,
+                                    hidden_size,
+                                    act=act,
+                                    param_attr=fluid.ParamAttr(name=name + "_r", initializer=initializer,
+                                   learning_rate=learning_rate),
+                                    )
+    output = fluid.layers.concat([self_feature, neigh_feature], axis=1)
+    output = fluid.layers.l2_normalize(output, axis=1)
+    return output

examples/erniesage/models/model_factory.py

+from models.base import BaseGNNModel
+from models.ernie import ErnieModel
+from models.erniesage_v1 import ErnieSageModelV1
+from models.erniesage_v2 import ErnieSageModelV2
+from models.erniesage_v3 import ErnieSageModelV3
+
+class Model(object):
+    @classmethod
+    def factory(cls, config):
+        name = config.model_type
+        if name == "BaseGNNModel":
+            return BaseGNNModel(config)
+        if name == "ErnieModel":
+            return ErnieModel(config)
+        if name == "ErnieSageModelV1":
+            return ErnieSageModelV1(config)
+        if name == "ErnieSageModelV2":
+            return ErnieSageModelV2(config)
+        if name == "ErnieSageModelV3":
+            return ErnieSageModelV3(config)
+        else:
+            raise ValueError
+
+

examples/erniesage/preprocessing/dump_graph.py

+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+########################################################################
+#
+# Copyright (c) 2020 Baidu.com, Inc. All Rights Reserved
+#
+# File: dump_graph.py
+# Author: suweiyue(suweiyue@baidu.com)
+# Date: 2020/03/01 22:17:13
+#
+########################################################################
+"""
+    Comment.
+"""
+from __future__ import division
+from __future__ import absolute_import
+from __future__ import print_function
+#from __future__ import unicode_literals
+
+import io
+import os
+import sys
+import argparse
+import logging
+import multiprocessing
+from functools import partial
+from io import open
+
+import numpy as np
+import tqdm
+import pgl
+from pgl.graph_kernel import alias_sample_build_table
+from pgl.utils.logger import log
+
+from tokenization import FullTokenizer
+
+
+def term2id(string, tokenizer, max_seqlen):
+    string = string.split("\t")[1]
+    tokens = tokenizer.tokenize(string)
+    ids = tokenizer.convert_tokens_to_ids(tokens)
+    ids = ids[:max_seqlen-1]
+    ids = ids + [2] # ids + [sep]
+    ids = ids + [0] * (max_seqlen - len(ids))
+    return ids
+
+
+def dump_graph(args):
+    if not os.path.exists(args.outpath):
+        os.makedirs(args.outpath)
+    neg_samples = []
+    str2id = dict()
+    term_file = io.open(os.path.join(args.outpath, "terms.txt"), "w", encoding=args.encoding)
+    terms = []
+    count = 0
+
+    with io.open(args.inpath, encoding=args.encoding) as f:
+        edges = []
+        for idx, line in enumerate(f):
+            if idx % 100000 == 0:
+                log.info("%s readed %s lines" % (args.inpath, idx))
+            slots = []
+            for col_idx, col in enumerate(line.strip("\n").split("\t")):
+                s = col[:args.max_seqlen]
+                if s not in str2id:
+                    str2id[s] = count
+                    count += 1
+                    term_file.write(str(col_idx) + "\t" + col + "\n")
+                    
+                slots.append(str2id[s])
+
+            src = slots[0]
+            dst = slots[1]
+            neg_samples.append(slots[2:])
+            edges.append((src, dst))
+            edges.append((dst, src))
+
+        term_file.close()
+        edges = np.array(edges, dtype="int64")
+        num_nodes = len(str2id)
+        str2id.clear()
+    log.info("building graph...")
+    graph = pgl.graph.Graph(num_nodes=num_nodes, edges=edges)
+    indegree = graph.indegree()
+    graph.outdegree()
+    graph.dump(args.outpath)
+    
+    # dump alias sample table
+    sqrt_indegree = np.sqrt(indegree)
+    distribution = 1. * sqrt_indegree / sqrt_indegree.sum()
+    alias, events = alias_sample_build_table(distribution)
+    np.save(os.path.join(args.outpath, "alias.npy"), alias)
+    np.save(os.path.join(args.outpath, "events.npy"), events)
+    np.save(os.path.join(args.outpath, "neg_samples.npy"), np.array(neg_samples))
+    log.info("End Build Graph")
+
+def dump_id2str_map(args):
+    log.info("Dump id2str map starting...")
+    id2str = np.array([line.strip("\n") for line in open(os.path.join(args.outpath, "terms.txt"), "r", encoding=args.encoding)])
+    np.save(os.path.join(args.outpath, "id2str.npy"), id2str)
+    log.info("Dump id2str map done.")
+
+def dump_node_feat(args):
+    log.info("Dump node feat starting...")
+    id2str = np.load(os.path.join(args.outpath, "id2str.npy"), mmap_mode="r")
+    pool = multiprocessing.Pool()
+    tokenizer = FullTokenizer(args.vocab_file)
+    term_ids = pool.map(partial(term2id, tokenizer=tokenizer, max_seqlen=args.max_seqlen), id2str)
+    np.save(os.path.join(args.outpath, "term_ids.npy"), np.array(term_ids))
+    log.info("Dump node feat done.")
+    pool.terminate()
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='main')
+    parser.add_argument("-i", "--inpath", type=str, default=None)
+    parser.add_argument("-l", "--max_seqlen", type=int, default=30)
+    parser.add_argument("--vocab_file", type=str, default="./vocab.txt")
+    parser.add_argument("--encoding", type=str, default="utf8")
+    parser.add_argument("-o", "--outpath", type=str, default=None)
+    args = parser.parse_args()
+    dump_graph(args)
+    dump_id2str_map(args)
+    dump_node_feat(args)

examples/erniesage/preprocessing/tokenization.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# 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.
+"""Tokenization classes."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import collections
+import unicodedata
+import six
+import sentencepiece as sp
+
+def convert_to_unicode(text):
+    """Converts `text` to Unicode (if it's not already), assuming utf-8 input."""
+    if six.PY3:
+        if isinstance(text, str):
+            return text
+        elif isinstance(text, bytes):
+            return text.decode("utf-8", "ignore")
+        else:
+            raise ValueError("Unsupported string type: %s" % (type(text)))
+    elif six.PY2:
+        if isinstance(text, str):
+            return text.decode("utf-8", "ignore")
+        elif isinstance(text, unicode):
+            return text
+        else:
+            raise ValueError("Unsupported string type: %s" % (type(text)))
+    else:
+        raise ValueError("Not running on Python2 or Python 3?")
+
+
+def printable_text(text):
+    """Returns text encoded in a way suitable for print or `tf.logging`."""
+
+    # These functions want `str` for both Python2 and Python3, but in one case
+    # it's a Unicode string and in the other it's a byte string.
+    if six.PY3:
+        if isinstance(text, str):
+            return text
+        elif isinstance(text, bytes):
+            return text.decode("utf-8", "ignore")
+        else:
+            raise ValueError("Unsupported string type: %s" % (type(text)))
+    elif six.PY2:
+        if isinstance(text, str):
+            return text
+        elif isinstance(text, unicode):
+            return text.encode("utf-8")
+        else:
+            raise ValueError("Unsupported string type: %s" % (type(text)))
+    else:
+        raise ValueError("Not running on Python2 or Python 3?")
+
+
+def load_vocab(vocab_file):
+    """Loads a vocabulary file into a dictionary."""
+    vocab = collections.OrderedDict()
+    fin = open(vocab_file, 'rb')
+    for num, line in enumerate(fin):
+        items = convert_to_unicode(line.strip()).split("\t")
+        if len(items) > 2:
+            break
+        token = items[0]
+        index = items[1] if len(items) == 2 else num
+        token = token.strip()
+        vocab[token] = int(index)
+    return vocab
+
+
+def convert_by_vocab(vocab, items):
+    """Converts a sequence of [tokens|ids] using the vocab."""
+    output = []
+    for item in items:
+        output.append(vocab[item])
+    return output
+
+
+def convert_tokens_to_ids_include_unk(vocab, tokens, unk_token="[UNK]"):
+    output = []
+    for token in tokens:
+        if token in vocab:
+            output.append(vocab[token])
+        else:
+            output.append(vocab[unk_token])
+    return output
+
+
+def convert_tokens_to_ids(vocab, tokens):
+    return convert_by_vocab(vocab, tokens)
+
+
+def convert_ids_to_tokens(inv_vocab, ids):
+    return convert_by_vocab(inv_vocab, ids)
+
+
+def whitespace_tokenize(text):
+    """Runs basic whitespace cleaning and splitting on a peice of text."""
+    text = text.strip()
+    if not text:
+        return []
+    tokens = text.split()
+    return tokens
+
+
+class FullTokenizer(object):
+    """Runs end-to-end tokenziation."""
+
+    def __init__(self, vocab_file, do_lower_case=True):
+        self.vocab = load_vocab(vocab_file)
+        self.inv_vocab = {v: k for k, v in self.vocab.items()}
+        self.basic_tokenizer = BasicTokenizer(do_lower_case=do_lower_case)
+        self.wordpiece_tokenizer = WordpieceTokenizer(vocab=self.vocab)
+
+    def tokenize(self, text):
+        split_tokens = []
+        for token in self.basic_tokenizer.tokenize(text):
+            for sub_token in self.wordpiece_tokenizer.tokenize(token):
+                split_tokens.append(sub_token)
+
+        return split_tokens
+
+    def convert_tokens_to_ids(self, tokens):
+        return convert_by_vocab(self.vocab, tokens)
+
+    def convert_ids_to_tokens(self, ids):
+        return convert_by_vocab(self.inv_vocab, ids)
+
+
+class CharTokenizer(object):
+    """Runs end-to-end tokenziation."""
+
+    def __init__(self, vocab_file, do_lower_case=True):
+        self.vocab = load_vocab(vocab_file)
+        self.inv_vocab = {v: k for k, v in self.vocab.items()}
+        self.tokenizer = WordpieceTokenizer(vocab=self.vocab)
+
+    def tokenize(self, text):
+        split_tokens = []
+        for token in text.lower().split(" "):
+            for sub_token in self.tokenizer.tokenize(token):
+                split_tokens.append(sub_token)
+        return split_tokens
+
+    def convert_tokens_to_ids(self, tokens):
+        return convert_by_vocab(self.vocab, tokens)
+
+    def convert_ids_to_tokens(self, ids):
+        return convert_by_vocab(self.inv_vocab, ids)
+
+
+class BasicTokenizer(object):
+    """Runs basic tokenization (punctuation splitting, lower casing, etc.)."""
+
+    def __init__(self, do_lower_case=True):
+        """Constructs a BasicTokenizer.
+
+        Args:
+            do_lower_case: Whether to lower case the input.
+        """
+        self.do_lower_case = do_lower_case
+
+    def tokenize(self, text):
+        """Tokenizes a piece of text."""
+        text = convert_to_unicode(text)
+        text = self._clean_text(text)
+
+        # This was added on November 1st, 2018 for the multilingual and Chinese
+        # models. This is also applied to the English models now, but it doesn't
+        # matter since the English models were not trained on any Chinese data
+        # and generally don't have any Chinese data in them (there are Chinese
+        # characters in the vocabulary because Wikipedia does have some Chinese
+        # words in the English Wikipedia.).
+        text = self._tokenize_chinese_chars(text)
+
+        orig_tokens = whitespace_tokenize(text)
+        split_tokens = []
+        for token in orig_tokens:
+            if self.do_lower_case:
+                token = token.lower()
+                token = self._run_strip_accents(token)
+            split_tokens.extend(self._run_split_on_punc(token))
+
+        output_tokens = whitespace_tokenize(" ".join(split_tokens))
+        return output_tokens
+
+    def _run_strip_accents(self, text):
+        """Strips accents from a piece of text."""
+        text = unicodedata.normalize("NFD", text)
+        output = []
+        for char in text:
+            cat = unicodedata.category(char)
+            if cat == "Mn":
+                continue
+            output.append(char)
+        return "".join(output)
+
+    def _run_split_on_punc(self, text):
+        """Splits punctuation on a piece of text."""
+        chars = list(text)
+        i = 0
+        start_new_word = True
+        output = []
+        while i < len(chars):
+            char = chars[i]
+            if _is_punctuation(char):
+                output.append([char])
+                start_new_word = True
+            else:
+                if start_new_word:
+                    output.append([])
+                start_new_word = False
+                output[-1].append(char)
+            i += 1
+
+        return ["".join(x) for x in output]
+
+    def _tokenize_chinese_chars(self, text):
+        """Adds whitespace around any CJK character."""
+        output = []
+        for char in text:
+            cp = ord(char)
+            if self._is_chinese_char(cp):
+                output.append(" ")
+                output.append(char)
+                output.append(" ")
+            else:
+                output.append(char)
+        return "".join(output)
+
+    def _is_chinese_char(self, cp):
+        """Checks whether CP is the codepoint of a CJK character."""
+        # This defines a "chinese character" as anything in the CJK Unicode block:
+        #     https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_(Unicode_block)
+        #
+        # Note that the CJK Unicode block is NOT all Japanese and Korean characters,
+        # despite its name. The modern Korean Hangul alphabet is a different block,
+        # as is Japanese Hiragana and Katakana. Those alphabets are used to write
+        # space-separated words, so they are not treated specially and handled
+        # like the all of the other languages.
+        if ((cp >= 0x4E00 and cp <= 0x9FFF) or  #
+            (cp >= 0x3400 and cp <= 0x4DBF) or  #
+            (cp >= 0x20000 and cp <= 0x2A6DF) or  #
+            (cp >= 0x2A700 and cp <= 0x2B73F) or  #
+            (cp >= 0x2B740 and cp <= 0x2B81F) or  #
+            (cp >= 0x2B820 and cp <= 0x2CEAF) or
+            (cp >= 0xF900 and cp <= 0xFAFF) or  #
+            (cp >= 0x2F800 and cp <= 0x2FA1F)):  #
+            return True
+
+        return False
+
+    def _clean_text(self, text):
+        """Performs invalid character removal and whitespace cleanup on text."""
+        output = []
+        for char in text:
+            cp = ord(char)
+            if cp == 0 or cp == 0xfffd or _is_control(char):
+                continue
+            if _is_whitespace(char):
+                output.append(" ")
+            else:
+                output.append(char)
+        return "".join(output)
+
+
+class SentencepieceTokenizer(object):
+    """Runs SentencePiece tokenziation."""
+
+    def __init__(self, vocab_file, do_lower_case=True, unk_token="[UNK]"):
+        self.vocab = load_vocab(vocab_file)
+        self.inv_vocab = {v: k for k, v in self.vocab.items()}
+        self.do_lower_case = do_lower_case
+        self.tokenizer = sp.SentencePieceProcessor()
+        self.tokenizer.Load(vocab_file + ".model")
+        self.sp_unk_token = "<unk>"
+        self.unk_token = unk_token
+
+    def tokenize(self, text):
+        """Tokenizes a piece of text into its word pieces.
+
+        Returns:
+            A list of wordpiece tokens.
+        """
+        text = text.lower() if self.do_lower_case else text 
+        text = convert_to_unicode(text.replace("\1", " "))
+        tokens = self.tokenizer.EncodeAsPieces(text)
+        
+        output_tokens = []
+        for token in tokens:
+            if token == self.sp_unk_token:
+                token = self.unk_token
+            
+            if token in self.vocab:
+                output_tokens.append(token)
+            else:
+                output_tokens.append(self.unk_token)
+        
+        return output_tokens
+    
+    def convert_tokens_to_ids(self, tokens):
+        return convert_by_vocab(self.vocab, tokens)
+
+    def convert_ids_to_tokens(self, ids):
+        return convert_by_vocab(self.inv_vocab, ids)
+
+
+class WordsegTokenizer(object):
+    """Runs Wordseg tokenziation."""
+
+    def __init__(self, vocab_file, do_lower_case=True, unk_token="[UNK]", 
+            split_token="\1"):
+        self.vocab = load_vocab(vocab_file)
+        self.inv_vocab = {v: k for k, v in self.vocab.items()}
+        self.tokenizer = sp.SentencePieceProcessor()
+        self.tokenizer.Load(vocab_file + ".model")
+        
+        self.do_lower_case = do_lower_case
+        self.unk_token = unk_token
+        self.split_token = split_token
+
+    def tokenize(self, text):
+        """Tokenizes a piece of text into its word pieces.
+
+        Returns:
+            A list of wordpiece tokens.
+        """
+        text = text.lower() if self.do_lower_case else text 
+        text = convert_to_unicode(text)
+        
+        output_tokens = []
+        for token in text.split(self.split_token):
+            if token in self.vocab:
+                output_tokens.append(token)
+            else:
+                sp_tokens = self.tokenizer.EncodeAsPieces(token)
+                for sp_token in sp_tokens:
+                    if sp_token in self.vocab:
+                        output_tokens.append(sp_token)
+        return output_tokens
+    
+    def convert_tokens_to_ids(self, tokens):
+        return convert_by_vocab(self.vocab, tokens)
+
+    def convert_ids_to_tokens(self, ids):
+        return convert_by_vocab(self.inv_vocab, ids)
+
+
+class WordpieceTokenizer(object):
+    """Runs WordPiece tokenziation."""
+
+    def __init__(self, vocab, unk_token="[UNK]", max_input_chars_per_word=100):
+        self.vocab = vocab
+        self.unk_token = unk_token
+        self.max_input_chars_per_word = max_input_chars_per_word
+
+    def tokenize(self, text):
+        """Tokenizes a piece of text into its word pieces.
+
+        This uses a greedy longest-match-first algorithm to perform tokenization
+        using the given vocabulary.
+
+        For example:
+            input = "unaffable"
+            output = ["un", "##aff", "##able"]
+
+        Args:
+            text: A single token or whitespace separated tokens. This should have
+                already been passed through `BasicTokenizer.
+
+        Returns:
+            A list of wordpiece tokens.
+        """
+
+        text = convert_to_unicode(text)
+
+        output_tokens = []
+        for token in whitespace_tokenize(text):
+            chars = list(token)
+            if len(chars) > self.max_input_chars_per_word:
+                output_tokens.append(self.unk_token)
+                continue
+
+            is_bad = False
+            start = 0
+            sub_tokens = []
+            while start < len(chars):
+                end = len(chars)
+                cur_substr = None
+                while start < end:
+                    substr = "".join(chars[start:end])
+                    if start > 0:
+                        substr = "##" + substr
+                    if substr in self.vocab:
+                        cur_substr = substr
+                        break
+                    end -= 1
+                if cur_substr is None:
+                    is_bad = True
+                    break
+                sub_tokens.append(cur_substr)
+                start = end
+
+            if is_bad:
+                output_tokens.append(self.unk_token)
+            else:
+                output_tokens.extend(sub_tokens)
+        return output_tokens
+
+
+def _is_whitespace(char):
+    """Checks whether `chars` is a whitespace character."""
+    # \t, \n, and \r are technically contorl characters but we treat them
+    # as whitespace since they are generally considered as such.
+    if char == " " or char == "\t" or char == "\n" or char == "\r":
+        return True
+    cat = unicodedata.category(char)
+    if cat == "Zs":
+        return True
+    return False
+
+
+def _is_control(char):
+    """Checks whether `chars` is a control character."""
+    # These are technically control characters but we count them as whitespace
+    # characters.
+    if char == "\t" or char == "\n" or char == "\r":
+        return False
+    cat = unicodedata.category(char)
+    if cat.startswith("C"):
+        return True
+    return False
+
+
+def _is_punctuation(char):
+    """Checks whether `chars` is a punctuation character."""
+    cp = ord(char)
+    # We treat all non-letter/number ASCII as punctuation.
+    # Characters such as "^", "$", and "`" are not in the Unicode
+    # Punctuation class but we treat them as punctuation anyways, for
+    # consistency.
+    if ((cp >= 33 and cp <= 47) or (cp >= 58 and cp <= 64) or
+        (cp >= 91 and cp <= 96) or (cp >= 123 and cp <= 126)):
+        return True
+    cat = unicodedata.category(char)
+    if cat.startswith("P"):
+        return True
+    return False

examples/erniesage/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 os
+import argparse
+import traceback
+
+import yaml
+import numpy as np
+from easydict import EasyDict as edict
+from pgl.utils.logger import log
+from pgl.utils import paddle_helper
+
+from learner import Learner
+from models.model_factory import Model
+from dataset.graph_reader import GraphGenerator 
+
+
+class TrainData(object):
+    def __init__(self, graph_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))
+
+        edges = np.load(os.path.join(graph_path, "edges.npy"), allow_pickle=True)
+        # edges is bidirectional.
+        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 neg_samples.size != 0:
+                train_negs = neg_samples[trainer_id::trainer_count]
+                returns["train_data"].append(train_negs)
+        log.info("Load train_data done.")
+        self.data = returns
+
+    def __getitem__(self, index):
+        return [ data[index] for data in self.data["train_data"]]
+
+    def __len__(self):
+        return len(self.data["train_data"][0])
+
+
+def main(config):
+    # Select Model
+    model = Model.factory(config)
+
+    # Build Train Edges
+    data = TrainData(config.graph_path)
+
+    # Build Train Data
+    train_iter = GraphGenerator(
+        graph_wrappers=model.graph_wrappers,
+        batch_size=config.batch_size,
+        data=data,
+        samples=config.samples,
+        num_workers=config.sample_workers,
+        feed_name_list=[var.name for var in model.feed_list],
+        use_pyreader=config.use_pyreader,
+        phase="train",
+        graph_data_path=config.graph_path,
+        shuffle=True)
+
+    log.info("build graph reader done.")
+
+    learner = Learner.factory(config.learner_type)
+    learner.build(model, train_iter, config)
+
+    learner.start()
+    learner.stop()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='main')
+    parser.add_argument("--conf", type=str, default="./config.yaml")
+    args = parser.parse_args()
+    config = edict(yaml.load(open(args.conf), Loader=yaml.FullLoader))
+    print(config)
+    main(config)

examples/gat/train.py

@@ -44,7 +44,6 @@ def main(args):
     with fluid.program_guard(train_program, startup_program):
         gw = pgl.graph_wrapper.GraphWrapper(
             name="graph",
-            place=place,
             node_feat=dataset.graph.node_feat_info())
 
         output = pgl.layers.gat(gw,

examples/gin/README.md

@@ -4,18 +4,19 @@
 
 ### Datasets
 
-The dataset can be downloaded from [here](https://github.com/weihua916/powerful-gnns/blob/master/dataset.zip)
+The dataset can be downloaded from [here](https://github.com/weihua916/powerful-gnns/blob/master/dataset.zip).
+After downloading the data，uncompress them, then a directory named `./dataset/` can be found in current directory. Note that the current directory is the root directory of GIN model.
 
 ### Dependencies
 
-- paddlepaddle 1.6
+- paddlepaddle >= 1.6
 - pgl 1.0.2
 
 ### How to run
 
 For examples, use GPU to train GIN model on MUTAG dataset.
 ```
-python main.py --use_cuda --dataset_name MUTAG 
+python main.py --use_cuda --dataset_name MUTAG  --data_path ./dataset
 ```
 
 ### Hyperparameters

examples/gin/model.py

@@ -50,7 +50,16 @@ class GINModel(object):
                     init_eps=0.0,
                     train_eps=self.train_eps)
 
-            h = fl.batch_norm(h)
+            h = fl.layer_norm(
+                h,
+                begin_norm_axis=1,
+                param_attr=fluid.ParamAttr(
+                    name="norm_scale_%s" % (i),
+                    initializer=fluid.initializer.Constant(1.0)),
+                bias_attr=fluid.ParamAttr(
+                    name="norm_bias_%s" % (i),
+                    initializer=fluid.initializer.Constant(0.0)), )
+
             h = fl.relu(h)
 
             features_list.append(h)

examples/graphsage/train.py

@@ -204,8 +204,8 @@ def main(args):
 
         graph_wrapper = pgl.graph_wrapper.GraphWrapper(
             "sub_graph",
-            fluid.CPUPlace(),
             node_feat=data['graph'].node_feat_info())
+
         model_loss, model_acc = build_graph_model(
             graph_wrapper,
             num_class=data["num_class"],

examples/graphsage/train_multi.py

@@ -231,7 +231,6 @@ def main(args):
     with fluid.program_guard(train_program, startup_program):
         graph_wrapper = pgl.graph_wrapper.GraphWrapper(
             "sub_graph",
-            fluid.CPUPlace(),
             node_feat=data['graph'].node_feat_info())
 
         model_loss, model_acc = build_graph_model(

examples/graphsage/train_scale.py

@@ -227,7 +227,6 @@ def main(args):
     with fluid.program_guard(train_program, startup_program):
         graph_wrapper = pgl.graph_wrapper.GraphWrapper(
             "sub_graph",
-            fluid.CPUPlace(),
             node_feat=data['graph'].node_feat_info())
 
         model_loss, model_acc = build_graph_model(

examples/stgcn/main.py

@@ -49,7 +49,6 @@ def main(args):
     with fluid.program_guard(train_program, startup_program):
         gw = pgl.graph_wrapper.GraphWrapper(
             "gw",
-            place,
             node_feat=[('norm', [None, 1], "float32")],
             edge_feat=[('weights', [None, 1], "float32")])
 

examples/unsup_graphsage/train.py

@@ -88,7 +88,7 @@ def build_graph_model(args):
 
     graph_wrappers.append(
         pgl.graph_wrapper.GraphWrapper(
-            "layer_0", fluid.CPUPlace(), node_feat=node_feature_info))
+            "layer_0", node_feat=node_feature_info))
     #edge_feat=[("f", [None, 1], "float32")]))
 
     num_embed = args.num_nodes

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/graphproppred/mol/README.md

+# Graph Property Prediction for Open Graph Benchmark (OGB)
+
+[The Open Graph Benchmark (OGB)](https://ogb.stanford.edu/) is a collection of benchmark datasets, data loaders, and evaluators for graph machine learning. Here we complete the Graph Property Prediction task based on PGL.
+
+### Requirements
+
+- paddlpaddle >= 1.7.1
+- pgl 1.0.2
+- ogb
+
+NOTE: To install ogb that is fited for this project, run below command to install ogb
+```
+git clone https://github.com/snap-stanford/ogb.git
+git checkout 482c40bc9f31fe25f9df5aa11c8fb657bd2b1621
+python setup.py install
+```
+
+### How to run
+For example, use GPU to train model on ogbg-molhiv dataset and ogb-molpcba dataset.
+```
+CUDA_VISIBLE_DEVICES=1 python -u main.py --config hiv_config.yaml --use_cuda
+
+CUDA_VISIBLE_DEVICES=2 python -u main.py --config pcba_config.yaml --use_cuda
+```
+
+If you want to use CPU to train model, environment variables `CPU_NUM` should be specified and should be in the range of 1 to N, where N is the total CPU number on your machine.
+```
+CPU_NUM=1 python -u main.py --config hiv_config.yaml
+
+CPU_NUM=1 python -u main.py --config pcba_config.yaml
+```
+
+### Experiment results
+
+| model | hiv (rocauc)| pcba (prcauc)|
+|-------|-------------|--------------|
+| GIN   |0.7719 (0.0079) | 0.2232 (0.0018) |

ogb_examples/graphproppred/mol/args.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.
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+from __future__ import unicode_literals
+from __future__ import absolute_import
+
+import os
+import time
+import argparse
+
+from utils.args import ArgumentGroup
+
+# yapf: disable
+parser = argparse.ArgumentParser(__doc__)
+parser.add_argument('--use_cuda', action='store_true')
+model_g = ArgumentGroup(parser, "model", "model configuration and paths.")
+model_g.add_arg("init_checkpoint",          str,  None,           "Init checkpoint to resume training from.")
+model_g.add_arg("init_pretraining_params",  str,  None,
+                "Init pre-training params which preforms fine-tuning from. If the "
+                 "arg 'init_checkpoint' has been set, this argument wouldn't be valid.")
+model_g.add_arg("./save_dir",              str,  "./checkpoints",  "Path to save checkpoints.")
+model_g.add_arg("hidden_size",             int,    128,       "hidden size.")
+
+
+train_g = ArgumentGroup(parser, "training", "training options.")
+train_g.add_arg("epoch",             int,    3,       "Number of epoches for fine-tuning.")
+train_g.add_arg("learning_rate",     float,  5e-5,    "Learning rate used to train with warmup.")
+train_g.add_arg("lr_scheduler",      str,    "linear_warmup_decay",
+                "scheduler of learning rate.", choices=['linear_warmup_decay', 'noam_decay'])
+train_g.add_arg("weight_decay",      float,  0.01,    "Weight decay rate for L2 regularizer.")
+train_g.add_arg("warmup_proportion", float,  0.1,
+                "Proportion of training steps to perform linear learning rate warmup for.")
+train_g.add_arg("save_steps",        int,    10000,   "The steps interval to save checkpoints.")
+train_g.add_arg("validation_steps",  int,    1000,    "The steps interval to evaluate model performance.")
+train_g.add_arg("use_dynamic_loss_scaling",    bool,   True,   "Whether to use dynamic loss scaling.")
+train_g.add_arg("init_loss_scaling",           float,  102400,
+                "Loss scaling factor for mixed precision training, only valid when use_fp16 is enabled.")
+
+train_g.add_arg("test_save",            str,    "./checkpoints/test_result",       "test_save")
+train_g.add_arg("metric",               str,    "simple_accuracy",   "metric")
+train_g.add_arg("incr_every_n_steps",          int,    100,   "Increases loss scaling every n consecutive.")
+train_g.add_arg("decr_every_n_nan_or_inf",     int,    2,
+                "Decreases loss scaling every n accumulated steps with nan or inf gradients.")
+train_g.add_arg("incr_ratio",                  float,  2.0,
+                "The multiplier to use when increasing the loss scaling.")
+train_g.add_arg("decr_ratio",                  float,  0.8,
+                "The less-than-one-multiplier to use when decreasing.")
+
+
+
+
+log_g = ArgumentGroup(parser,     "logging", "logging related.")
+log_g.add_arg("skip_steps",          int,    10,    "The steps interval to print loss.")
+log_g.add_arg("verbose",             bool,   False, "Whether to output verbose log.")
+log_g.add_arg("log_dir",             str,   './logs/', "Whether to output verbose log.")
+
+data_g = ArgumentGroup(parser, "data", "Data paths, vocab paths and data processing options")
+data_g.add_arg("tokenizer",           str, "FullTokenizer",
+              "ATTENTION: the INPUT must be splited by Word with blank while using SentencepieceTokenizer or WordsegTokenizer")
+data_g.add_arg("train_set",           str,  None,  "Path to training data.")
+data_g.add_arg("test_set",            str,  None,  "Path to test data.")
+data_g.add_arg("dev_set",             str,  None,  "Path to validation data.")
+data_g.add_arg("aug1_type",           str,  "scheme1",  "augment type")
+data_g.add_arg("aug2_type",           str,  "scheme1",  "augment type")
+data_g.add_arg("batch_size",          int,  32,    "Total examples' number in batch for training. see also --in_tokens.")
+data_g.add_arg("predict_batch_size",  int,  None,    "Total examples' number in batch for predict. see also --in_tokens.")
+data_g.add_arg("random_seed",         int,  None,     "Random seed.")
+data_g.add_arg("buf_size",         int,  1000,     "Random seed.")
+
+run_type_g = ArgumentGroup(parser, "run_type", "running type options.")
+run_type_g.add_arg("num_iteration_per_drop_scope", int,    10,    "Iteration intervals to drop scope.")
+run_type_g.add_arg("do_train",                     bool,   True,  "Whether to perform training.")
+run_type_g.add_arg("do_val",                       bool,   True,  "Whether to perform evaluation on dev data set.")
+run_type_g.add_arg("do_test",                      bool,   True,  "Whether to perform evaluation on test data set.")
+run_type_g.add_arg("metrics",                      bool,   True,  "Whether to perform evaluation on test data set.")
+run_type_g.add_arg("shuffle",                      bool,   True,  "")
+run_type_g.add_arg("for_cn",                       bool,   True,  "model train for cn or for other langs.")
+run_type_g.add_arg("num_workers",                       int,   1,  "use multiprocess to generate graph")
+run_type_g.add_arg("output_dir",                       str,   None,  "path to save model")
+run_type_g.add_arg("config",                       str,   None,  "configure yaml file")
+run_type_g.add_arg("n",                       str,   None,  "task name")
+run_type_g.add_arg("task_name", str,   None,  "task name")
+run_type_g.add_arg("pretrain", bool,   False,  "Whether do pretrian")
+run_type_g.add_arg("pretrain_name", str,   None,  "pretrain task name")
+run_type_g.add_arg("pretrain_config", str,   None,  "pretrain config.yaml file")
+run_type_g.add_arg("pretrain_model_step", str,   None,  "pretrain model step")
+run_type_g.add_arg("model_type", str,   "BaseLineModel",  "pretrain model step")
+run_type_g.add_arg("num_class", int,   1,  "number class")
+run_type_g.add_arg("dataset_name", str,   None,  "finetune dataset name")
+run_type_g.add_arg("eval_metrics", str,   None,  "evaluate metrics")
+run_type_g.add_arg("task_type", str,   None,  "regression or classification")

ogb_examples/graphproppred/mol/data/__init__.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.

ogb_examples/graphproppred/mol/data/base_dataset.py

+#   Copyright (c) 2018 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 sys
+import os
+
+from ogb.graphproppred import GraphPropPredDataset
+import pgl
+from pgl.utils.logger import log
+
+
+class BaseDataset(object):
+    def __init__(self):
+        pass
+
+    def __getitem__(self, idx):
+        raise NotImplementedError
+
+    def __len__(self):
+        raise NotImplementedError
+
+
+class Subset(BaseDataset):
+    r"""
+    Subset of a dataset at specified indices.
+    Arguments:
+        dataset (Dataset): The whole Dataset
+        indices (sequence): Indices in the whole set selected for subset
+    """
+
+    def __init__(self, dataset, indices):
+        self.dataset = dataset
+        self.indices = indices
+
+    def __getitem__(self, idx):
+        return self.dataset[self.indices[idx]]
+
+    def __len__(self):
+        return len(self.indices)
+
+
+class Dataset(BaseDataset):
+    def __init__(self, args):
+        self.args = args
+        self.raw_dataset = GraphPropPredDataset(name=args.dataset_name)
+        self.num_tasks = self.raw_dataset.num_tasks
+        self.eval_metrics = self.raw_dataset.eval_metric
+        self.task_type = self.raw_dataset.task_type
+
+        self.pgl_graph_list = []
+        self.graph_label_list = []
+        for i in range(len(self.raw_dataset)):
+            graph, label = self.raw_dataset[i]
+            edges = list(zip(graph["edge_index"][0], graph["edge_index"][1]))
+            g = pgl.graph.Graph(num_nodes=graph["num_nodes"], edges=edges)
+
+            if graph["edge_feat"] is not None:
+                g.edge_feat["feat"] = graph["edge_feat"]
+
+            if graph["node_feat"] is not None:
+                g.node_feat["feat"] = graph["node_feat"]
+
+            self.pgl_graph_list.append(g)
+            self.graph_label_list.append(label)
+
+    def __getitem__(self, idx):
+        return self.pgl_graph_list[idx], self.graph_label_list[idx]
+
+    def __len__(self):
+        return len(slef.pgl_graph_list)
+
+    def get_idx_split(self):
+        return self.raw_dataset.get_idx_split()

ogb_examples/graphproppred/mol/data/dataloader.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 file implement the graph dataloader.
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+from __future__ import unicode_literals
+from __future__ import absolute_import
+
+import ssl
+ssl._create_default_https_context = ssl._create_unverified_context
+# SSL
+
+import torch
+import sys
+import six
+from io import open
+import collections
+from collections import namedtuple
+import numpy as np
+import tqdm
+import time
+
+import paddle
+import paddle.fluid as fluid
+import paddle.fluid.layers as fl
+import pgl
+from pgl.utils import mp_reader
+from pgl.utils.logger import log
+
+from ogb.graphproppred import GraphPropPredDataset
+
+
+def batch_iter(data, batch_size, fid, num_workers):
+    """node_batch_iter
+    """
+    size = len(data)
+    perm = np.arange(size)
+    np.random.shuffle(perm)
+    start = 0
+    cc = 0
+    while start < size:
+        index = perm[start:start + batch_size]
+        start += batch_size
+        cc += 1
+        if cc % num_workers != fid:
+            continue
+        yield data[index]
+
+
+def scan_batch_iter(data, batch_size, fid, num_workers):
+    """scan_batch_iter
+    """
+    batch = []
+    cc = 0
+    for line_example in data.scan():
+        cc += 1
+        if cc % num_workers != fid:
+            continue
+        batch.append(line_example)
+        if len(batch) == batch_size:
+            yield batch
+            batch = []
+
+    if len(batch) > 0:
+        yield batch
+
+
+class GraphDataloader(object):
+    """Graph Dataloader
+    """
+
+    def __init__(self,
+                 dataset,
+                 graph_wrapper,
+                 batch_size,
+                 seed=0,
+                 num_workers=1,
+                 buf_size=1000,
+                 shuffle=True):
+
+        self.shuffle = shuffle
+        self.seed = seed
+        self.num_workers = num_workers
+        self.buf_size = buf_size
+        self.batch_size = batch_size
+        self.dataset = dataset
+        self.graph_wrapper = graph_wrapper
+
+    def batch_fn(self, batch_examples):
+        """ batch_fn batch producer"""
+        graphs = [b[0] for b in batch_examples]
+        labels = [b[1] for b in batch_examples]
+        join_graph = pgl.graph.MultiGraph(graphs)
+        labels = np.array(labels)
+
+        feed_dict = self.graph_wrapper.to_feed(join_graph)
+        batch_valid = (labels == labels).astype("float32")
+        labels = np.nan_to_num(labels).astype("float32")
+        feed_dict['labels'] = labels
+        feed_dict['unmask'] = batch_valid
+        return feed_dict
+
+    def batch_iter(self, fid):
+        """batch_iter"""
+        if self.shuffle:
+            for batch in batch_iter(self, self.batch_size, fid,
+                                    self.num_workers):
+                yield batch
+        else:
+            for batch in scan_batch_iter(self, self.batch_size, fid,
+                                         self.num_workers):
+                yield batch
+
+    def __len__(self):
+        """__len__"""
+        return len(self.dataset)
+
+    def __getitem__(self, idx):
+        """__getitem__"""
+        if isinstance(idx, collections.Iterable):
+            return [self[bidx] for bidx in idx]
+        else:
+            return self.dataset[idx]
+
+    def __iter__(self):
+        """__iter__"""
+
+        def worker(filter_id):
+            def func_run():
+                for batch_examples in self.batch_iter(filter_id):
+                    batch_dict = self.batch_fn(batch_examples)
+                    yield batch_dict
+
+            return func_run
+
+        if self.num_workers == 1:
+            r = paddle.reader.buffered(worker(0), self.buf_size)
+        else:
+            worker_pool = [worker(wid) for wid in range(self.num_workers)]
+            worker = mp_reader.multiprocess_reader(
+                worker_pool, use_pipe=True, queue_size=1000)
+            r = paddle.reader.buffered(worker, self.buf_size)
+
+        for batch in r():
+            yield batch
+
+    def scan(self):
+        """scan"""
+        for example in self.dataset:
+            yield example
+
+
+if __name__ == "__main__":
+    from base_dataset import BaseDataset, Subset
+    dataset = GraphPropPredDataset(name="ogbg-molhiv")
+    splitted_index = dataset.get_idx_split()
+    train_dataset = Subset(dataset, splitted_index['train'])
+    valid_dataset = Subset(dataset, splitted_index['valid'])
+    test_dataset = Subset(dataset, splitted_index['test'])
+    log.info("Train Examples: %s" % len(train_dataset))
+    log.info("Val Examples: %s" % len(valid_dataset))
+    log.info("Test Examples: %s" % len(test_dataset))
+
+    #  train_loader = GraphDataloader(train_dataset, batch_size=3)
+    #  for batch_data in train_loader:
+    #      graphs, labels = batch_data
+    #      print(labels.shape)
+    #      time.sleep(4)

ogb_examples/graphproppred/mol/data/splitters.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.
+
+import sys
+import os
+import logging
+from random import random
+import pandas as pd
+import numpy as np
+from itertools import compress
+
+import scipy.sparse as sp
+from sklearn.model_selection import StratifiedKFold
+from sklearn.preprocessing import StandardScaler
+from rdkit.Chem.Scaffolds import MurckoScaffold
+
+import pgl
+from pgl.utils import paddle_helper
+try:
+    from dataset.Dataset import Subset
+    from dataset.Dataset import ChemDataset
+except:
+    from Dataset import Subset
+    from Dataset import ChemDataset
+
+log = logging.getLogger("logger")
+
+
+def random_split(dataset, args):
+    total_precent = args.frac_train + args.frac_valid + args.frac_test
+    np.testing.assert_almost_equal(total_precent, 1.0)
+
+    length = len(dataset)
+    perm = list(range(length))
+    np.random.shuffle(perm)
+    num_train = int(args.frac_train * length)
+    num_valid = int(args.frac_valid * length)
+    num_test = int(args.frac_test * length)
+
+    train_indices = perm[0:num_train]
+    valid_indices = perm[num_train:(num_train + num_valid)]
+    test_indices = perm[(num_train + num_valid):]
+    assert (len(train_indices) + len(valid_indices) + len(test_indices)
+            ) == length
+
+    train_dataset = Subset(dataset, train_indices)
+    valid_dataset = Subset(dataset, valid_indices)
+    test_dataset = Subset(dataset, test_indices)
+    return train_dataset, valid_dataset, test_dataset
+
+
+def scaffold_split(dataset, args, return_smiles=False):
+    total_precent = args.frac_train + args.frac_valid + args.frac_test
+    np.testing.assert_almost_equal(total_precent, 1.0)
+
+    smiles_list_file = os.path.join(args.data_dir, "smiles.csv")
+    smiles_list = pd.read_csv(smiles_list_file, header=None)[0].tolist()
+
+    non_null = np.ones(len(dataset)) == 1
+    smiles_list = list(compress(enumerate(smiles_list), non_null))
+
+    # create dict of the form {scaffold_i: [idx1, idx....]}
+    all_scaffolds = {}
+    for i, smiles in smiles_list:
+        scaffold = MurckoScaffold.MurckoScaffoldSmiles(
+            smiles=smiles, includeChirality=True)
+        #  scaffold = generate_scaffold(smiles, include_chirality=True)
+        if scaffold not in all_scaffolds:
+            all_scaffolds[scaffold] = [i]
+        else:
+            all_scaffolds[scaffold].append(i)
+
+    # sort from largest to smallest sets
+    all_scaffolds = {
+        key: sorted(value)
+        for key, value in all_scaffolds.items()
+    }
+    all_scaffold_sets = [
+        scaffold_set
+        for (scaffold, scaffold_set) in sorted(
+            all_scaffolds.items(),
+            key=lambda x: (len(x[1]), x[1][0]),
+            reverse=True)
+    ]
+
+    # get train, valid test indices
+    train_cutoff = args.frac_train * len(smiles_list)
+    valid_cutoff = (args.frac_train + args.frac_valid) * len(smiles_list)
+    train_idx, valid_idx, test_idx = [], [], []
+    for scaffold_set in all_scaffold_sets:
+        if len(train_idx) + len(scaffold_set) > train_cutoff:
+            if len(train_idx) + len(valid_idx) + len(
+                    scaffold_set) > valid_cutoff:
+                test_idx.extend(scaffold_set)
+            else:
+                valid_idx.extend(scaffold_set)
+        else:
+            train_idx.extend(scaffold_set)
+
+    assert len(set(train_idx).intersection(set(valid_idx))) == 0
+    assert len(set(test_idx).intersection(set(valid_idx))) == 0
+    #  log.info(len(scaffold_set))
+    #  log.info(["train_idx", train_idx])
+    #  log.info(["valid_idx", valid_idx])
+    #  log.info(["test_idx", test_idx])
+
+    train_dataset = Subset(dataset, train_idx)
+    valid_dataset = Subset(dataset, valid_idx)
+    test_dataset = Subset(dataset, test_idx)
+
+    if return_smiles:
+        train_smiles = [smiles_list[i][1] for i in train_idx]
+        valid_smiles = [smiles_list[i][1] for i in valid_idx]
+        test_smiles = [smiles_list[i][1] for i in test_idx]
+
+        return train_dataset, valid_dataset, test_dataset, (
+            train_smiles, valid_smiles, test_smiles)
+
+    return train_dataset, valid_dataset, test_dataset
+
+
+if __name__ == "__main__":
+    file_path = os.path.dirname(os.path.realpath(__file__))
+    proj_path = os.path.join(file_path, '../')
+    sys.path.append(proj_path)
+    from utils.config import Config
+    from dataset.Dataset import Subset
+    from dataset.Dataset import ChemDataset
+
+    config_file = "./finetune_config.yaml"
+    args = Config(config_file)
+    log.info("loading dataset")
+    dataset = ChemDataset(args)
+
+    train_dataset, valid_dataset, test_dataset = scaffold_split(dataset, args)
+
+    log.info("Train Examples: %s" % len(train_dataset))
+    log.info("Val Examples: %s" % len(valid_dataset))
+    log.info("Test Examples: %s" % len(test_dataset))
+    import ipdb
+    ipdb.set_trace()
+    log.info("preprocess finish")

ogb_examples/graphproppred/mol/hiv_config.yaml

+task_name: hiv
+seed: 15391
+dataset_name: ogbg-molhiv
+eval_metrics: null 
+task_type: null
+num_class: null
+pool_type: average
+train_eps: True
+norm_type: layer_norm
+
+model_type: GNNModel
+embed_dim: 128
+num_layers: 5
+hidden_size: 256
+save_dir: ./checkpoints
+
+
+# finetune model config
+init_checkpoint: null
+init_pretraining_params: null
+
+# data config
+data_dir: ./dataset/
+symmetry: True
+batch_size: 32
+buf_size: 1000
+metrics: True
+shuffle: True
+num_workers: 12
+output_dir: ./outputs/
+
+# trainging config
+epoch: 50
+learning_rate: 0.0001
+lr_scheduler: linear_warmup_decay
+weight_decay: 0.01
+warmup_proportion: 0.1
+save_steps: 10000
+validation_steps: 1000
+use_dynamic_loss_scaling: True
+init_loss_scaling: 102400
+metric: simple_accuracy
+incr_every_n_steps: 100
+decr_every_n_nan_or_inf: 2
+incr_ratio: 2.0
+decr_ratio: 0.8
+log_dir: ./logs
+eval_step: 400
+train_log_step: 20
+
+# log config
+skip_steps: 10
+verbose: False

ogb_examples/graphproppred/mol/main.py

+#   Copyright (c) 2018 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 ssl
+ssl._create_default_https_context = ssl._create_unverified_context
+# SSL
+
+import torch
+import os
+import re
+import time
+from random import random
+from functools import reduce, partial
+import numpy as np
+import multiprocessing
+
+from ogb.graphproppred import Evaluator
+import paddle
+import paddle.fluid as F
+import paddle.fluid.layers as L
+import pgl
+from pgl.utils import paddle_helper
+from pgl.utils.logger import log
+
+from utils.args import print_arguments, check_cuda, prepare_logger
+from utils.init import init_checkpoint, init_pretraining_params
+from utils.config import Config
+from optimization import optimization
+from monitor.train_monitor import train_and_evaluate
+from args import parser
+
+import model as Model
+from data.base_dataset import Subset, Dataset
+from data.dataloader import GraphDataloader
+
+
+def main(args):
+    log.info('loading data')
+    dataset = Dataset(args)
+    args.num_class = dataset.num_tasks
+    args.eval_metrics = dataset.eval_metrics
+    args.task_type = dataset.task_type
+    splitted_index = dataset.get_idx_split()
+    train_dataset = Subset(dataset, splitted_index['train'])
+    valid_dataset = Subset(dataset, splitted_index['valid'])
+    test_dataset = Subset(dataset, splitted_index['test'])
+
+    log.info("preprocess finish")
+    log.info("Train Examples: %s" % len(train_dataset))
+    log.info("Val Examples: %s" % len(valid_dataset))
+    log.info("Test Examples: %s" % len(test_dataset))
+
+    train_prog = F.Program()
+    startup_prog = F.Program()
+
+    if args.use_cuda:
+        dev_list = F.cuda_places()
+        place = dev_list[0]
+        dev_count = len(dev_list)
+    else:
+        place = F.CPUPlace()
+        dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
+        #  dev_count = args.cpu_num
+
+    log.info("building model")
+    with F.program_guard(train_prog, startup_prog):
+        with F.unique_name.guard():
+            graph_model = getattr(Model, args.model_type)(args, dataset)
+            train_ds = GraphDataloader(
+                train_dataset,
+                graph_model.graph_wrapper,
+                batch_size=args.batch_size)
+
+            num_train_examples = len(train_dataset)
+            max_train_steps = args.epoch * num_train_examples // args.batch_size // dev_count
+            warmup_steps = int(max_train_steps * args.warmup_proportion)
+
+            scheduled_lr, loss_scaling = optimization(
+                loss=graph_model.loss,
+                warmup_steps=warmup_steps,
+                num_train_steps=max_train_steps,
+                learning_rate=args.learning_rate,
+                train_program=train_prog,
+                startup_prog=startup_prog,
+                weight_decay=args.weight_decay,
+                scheduler=args.lr_scheduler,
+                use_fp16=False,
+                use_dynamic_loss_scaling=args.use_dynamic_loss_scaling,
+                init_loss_scaling=args.init_loss_scaling,
+                incr_every_n_steps=args.incr_every_n_steps,
+                decr_every_n_nan_or_inf=args.decr_every_n_nan_or_inf,
+                incr_ratio=args.incr_ratio,
+                decr_ratio=args.decr_ratio)
+
+    test_prog = F.Program()
+    with F.program_guard(test_prog, startup_prog):
+        with F.unique_name.guard():
+            _graph_model = getattr(Model, args.model_type)(args, dataset)
+
+    test_prog = test_prog.clone(for_test=True)
+
+    valid_ds = GraphDataloader(
+        valid_dataset,
+        graph_model.graph_wrapper,
+        batch_size=args.batch_size,
+        shuffle=False)
+    test_ds = GraphDataloader(
+        test_dataset,
+        graph_model.graph_wrapper,
+        batch_size=args.batch_size,
+        shuffle=False)
+
+    exe = F.Executor(place)
+    exe.run(startup_prog)
+    for init in graph_model.init_vars:
+        init(place)
+    for init in _graph_model.init_vars:
+        init(place)
+
+    if args.init_pretraining_params is not None:
+        init_pretraining_params(
+            exe, args.init_pretraining_params, main_program=startup_prog)
+
+    nccl2_num_trainers = 1
+    nccl2_trainer_id = 0
+    if dev_count > 1:
+
+        exec_strategy = F.ExecutionStrategy()
+        exec_strategy.num_threads = dev_count
+
+        train_exe = F.ParallelExecutor(
+            use_cuda=args.use_cuda,
+            loss_name=graph_model.loss.name,
+            exec_strategy=exec_strategy,
+            main_program=train_prog,
+            num_trainers=nccl2_num_trainers,
+            trainer_id=nccl2_trainer_id)
+
+        test_exe = exe
+    else:
+        train_exe, test_exe = exe, exe
+
+    evaluator = Evaluator(args.dataset_name)
+
+    train_and_evaluate(
+        exe=exe,
+        train_exe=train_exe,
+        valid_exe=test_exe,
+        train_ds=train_ds,
+        valid_ds=valid_ds,
+        test_ds=test_ds,
+        train_prog=train_prog,
+        valid_prog=test_prog,
+        args=args,
+        dev_count=dev_count,
+        evaluator=evaluator,
+        model=graph_model)
+
+
+if __name__ == "__main__":
+    args = parser.parse_args()
+    if args.config is not None:
+        config = Config(args.config, isCreate=True, isSave=True)
+
+    config['use_cuda'] = args.use_cuda
+
+    log.info(config)
+
+    main(config)

ogb_examples/graphproppred/mol/model.py

+#-*- coding: utf-8 -*-
+import os
+import re
+import time
+import logging
+from random import random
+from functools import reduce, partial
+
+import numpy as np
+import multiprocessing
+
+import paddle
+import paddle.fluid as F
+import paddle.fluid.layers as L
+import pgl
+from pgl.graph_wrapper import GraphWrapper
+from pgl.layers.conv import gcn, gat
+from pgl.utils import paddle_helper
+from pgl.utils.logger import log
+
+from utils.args import print_arguments, check_cuda, prepare_logger
+from utils.init import init_checkpoint, init_pretraining_params
+
+from mol_encoder import AtomEncoder, BondEncoder
+
+
+def copy_send(src_feat, dst_feat, edge_feat):
+    return src_feat["h"]
+
+
+def mean_recv(feat):
+    return L.sequence_pool(feat, pool_type="average")
+
+
+def sum_recv(feat):
+    return L.sequence_pool(feat, pool_type="sum")
+
+
+def max_recv(feat):
+    return L.sequence_pool(feat, pool_type="max")
+
+
+def unsqueeze(tensor):
+    tensor = L.unsqueeze(tensor, axes=-1)
+    tensor.stop_gradient = True
+    return tensor
+
+
+class Metric:
+    def __init__(self, **args):
+        self.args = args
+
+    @property
+    def vars(self):
+        values = [self.args[k] for k in self.args.keys()]
+        return values
+
+    def parse(self, fetch_list):
+        tup = list(zip(self.args.keys(), [float(v[0]) for v in fetch_list]))
+        return dict(tup)
+
+
+def gin_layer(gw, node_features, edge_features, train_eps, name):
+    def send_func(src_feat, dst_feat, edge_feat):
+        """Send"""
+        return src_feat["h"] + edge_feat["h"]
+
+    epsilon = L.create_parameter(
+        shape=[1, 1],
+        dtype="float32",
+        attr=F.ParamAttr(name="%s_eps" % name),
+        default_initializer=F.initializer.ConstantInitializer(value=0.0))
+    if not train_eps:
+        epsilon.stop_gradient = True
+
+    msg = gw.send(
+        send_func,
+        nfeat_list=[("h", node_features)],
+        efeat_list=[("h", edge_features)])
+
+    node_feat = gw.recv(msg, "sum") + node_features * (epsilon + 1.0)
+
+    #  if apply_func is not None:
+    #      node_feat = apply_func(node_feat, name)
+    return node_feat
+
+
+class GNNModel(object):
+    def __init__(self, args, dataset):
+        self.args = args
+        self.dataset = dataset
+        self.hidden_size = self.args.hidden_size
+        self.embed_dim = self.args.embed_dim
+        self.dropout_prob = self.args.dropout_rate
+        self.pool_type = self.args.pool_type
+        self._init_vars = []
+
+        graph_data = []
+        g, label = self.dataset[0]
+        graph_data.append(g)
+        g, label = self.dataset[1]
+        graph_data.append(g)
+
+        batch_graph = pgl.graph.MultiGraph(graph_data)
+        graph_data = batch_graph
+        graph_data.edge_feat["feat"] = graph_data.edge_feat["feat"].astype(
+            "int64")
+        graph_data.node_feat["feat"] = graph_data.node_feat["feat"].astype(
+            "int64")
+        self.graph_wrapper = GraphWrapper(
+            name="graph",
+            place=F.CPUPlace(),
+            node_feat=graph_data.node_feat_info(),
+            edge_feat=graph_data.edge_feat_info())
+
+        self.atom_encoder = AtomEncoder(name="atom", emb_dim=self.embed_dim)
+        self.bond_encoder = BondEncoder(name="bond", emb_dim=self.embed_dim)
+
+        self.labels = L.data(
+            "labels",
+            shape=[None, self.args.num_class],
+            dtype="float32",
+            append_batch_size=False)
+
+        self.unmask = L.data(
+            "unmask",
+            shape=[None, self.args.num_class],
+            dtype="float32",
+            append_batch_size=False)
+
+        self.build_model()
+
+    def build_model(self):
+        node_features = self.atom_encoder(self.graph_wrapper.node_feat['feat'])
+        edge_features = self.bond_encoder(self.graph_wrapper.edge_feat['feat'])
+
+        self._enc_out = self.node_repr_encode(node_features, edge_features)
+
+        logits = L.fc(self._enc_out,
+                      self.args.num_class,
+                      act=None,
+                      param_attr=F.ParamAttr(name="final_fc"))
+
+        #  L.Print(self.labels, message="labels")
+        #  L.Print(self.unmask, message="unmask")
+        loss = L.sigmoid_cross_entropy_with_logits(x=logits, label=self.labels)
+        loss = loss * self.unmask
+        self.loss = L.reduce_sum(loss) / L.reduce_sum(self.unmask)
+        self.pred = L.sigmoid(logits)
+
+        self._metrics = Metric(loss=self.loss)
+
+    def node_repr_encode(self, node_features, edge_features):
+        features_list = [node_features]
+        for layer in range(self.args.num_layers):
+            feat = gin_layer(
+                self.graph_wrapper,
+                features_list[layer],
+                edge_features,
+                train_eps=self.args.train_eps,
+                name="gin_%s" % layer, )
+
+            feat = self.mlp(feat, name="mlp_%s" % layer)
+
+            feat = feat + features_list[layer]  # residual
+
+            features_list.append(feat)
+
+        output = pgl.layers.graph_pooling(
+            self.graph_wrapper, features_list[-1], self.args.pool_type)
+
+        return output
+
+    def mlp(self, features, name):
+        h = features
+        dim = features.shape[-1]
+        dim_list = [dim * 2, dim]
+        for i in range(2):
+            h = L.fc(h,
+                     size=dim_list[i],
+                     name="%s_fc_%s" % (name, i),
+                     act=None)
+            if self.args.norm_type == "layer_norm":
+                log.info("norm_type is %s" % self.args.norm_type)
+                h = L.layer_norm(
+                    h,
+                    begin_norm_axis=1,
+                    param_attr=F.ParamAttr(
+                        name="norm_scale_%s_%s" % (name, i),
+                        initializer=F.initializer.Constant(1.0)),
+                    bias_attr=F.ParamAttr(
+                        name="norm_bias_%s_%s" % (name, i),
+                        initializer=F.initializer.Constant(0.0)), )
+            else:
+                log.info("using batch_norm")
+                h = L.batch_norm(h)
+            h = pgl.layers.graph_norm(self.graph_wrapper, h)
+            h = L.relu(h)
+        return h
+
+    def get_enc_output(self):
+        return self._enc_out
+
+    @property
+    def init_vars(self):
+        return self._init_vars
+
+    @property
+    def metrics(self):
+        return self._metrics

ogb_examples/graphproppred/mol/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/__init__.py

@@ -13,7 +13,7 @@
 # limitations under the License.
 """Generate pgl apis
 """
-__version__ = "1.0.2"
+__version__ = "1.1.0"
 from pgl import layers
 from pgl import graph_wrapper
 from pgl import graph