Merge pull request PaddlePaddle#398 from duyiqi17/mind

model MIND

datasets/AmazonBook/preprocess.py

@@ -0,0 +1,76 @@
+# 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 json
+import numpy as np
+import argparse
+import random
+
+parser = argparse.ArgumentParser()
+parser.add_argument(
+    "-type", type=str, default="train", help="train|valid|test")
+parser.add_argument("-maxlen", type=int, default=20)
+
+
+def load_graph(source):
+    graph = {}
+    with open(source) as fr:
+        for line in fr:
+            conts = line.strip().split(',')
+            user_id = int(conts[0])
+            item_id = int(conts[1])
+            time_stamp = int(conts[2])
+            if user_id not in graph:
+                graph[user_id] = []
+            graph[user_id].append((item_id, time_stamp))
+
+    for user_id, value in graph.items():
+        value.sort(key=lambda x: x[1])
+        graph[user_id] = [x[0] for x in value]
+    return graph
+
+
+if __name__ == "__main__":
+    args = parser.parse_args()
+    filelist = []
+    for i in range(10):
+        filelist.append(open(args.type + "/part-%d" % (i), "w"))
+    action_graph = load_graph("book_data/book_" + args.type + ".txt")
+    if args.type == "train":
+        for uid, item_list in action_graph.items():
+            for i in range(4, len(item_list)):
+                if i >= args.maxlen:
+                    hist_item = item_list[i - args.maxlen:i]
+                else:
+                    hist_item = item_list[:i]
+                target_item = item_list[i]
+                print(
+                    " ".join(["user_id:" + str(uid)] + [
+                        "hist_item:" + str(n) for n in hist_item
+                    ] + ["target_item:" + str(target_item)]),
+                    file=random.choice(filelist))
+    else:
+        for uid, item_list in action_graph.items():
+            k = int(len(item_list) * 0.8)
+            if k >= args.maxlen:
+                hist_item = item_list[k - args.maxlen:k]
+            else:
+                hist_item = item_list[:k]
+            target_item = item_list[k:]
+            print(
+                " ".join(["user_id:" + str(uid), "target_item:0"] + [
+                    "hist_item:" + str(n) for n in hist_item
+                ] + ["eval_item:" + str(n) for n in target_item]),
+                file=random.choice(filelist))

datasets/AmazonBook/run.sh

@@ -0,0 +1,11 @@
+
+wget https://paddlerec.bj.bcebos.com/datasets/AmazonBook/AmazonBook.tar.gz
+
+tar -xvf AmazonBook.tar.gz
+
+rm -rf train valid
+mkdir train
+mkdir valid
+
+mv book_data/book_train.txt  train
+python preprocess.py -type valid -maxlen 20

datasets/readme.md

@@ -26,3 +26,4 @@ sh data_process.sh
  |[one_billion](http://www.statmt.org/lm-benchmark/)|拥有十亿个单词基准，为语言建模实验提供标准的训练和测试|[One Billion Word Benchmark for Measuring Progress in Statistical Language Modeling](https://arxiv.org/abs/1312.3005)|
  |[MIND](https://paddlerec.bj.bcebos.com/datasets/MIND/bigdata.zip)|MIND即MIcrosoft News Dataset的简写，MIND里的数据来自Microsoft News用户的行为日志。MIND的数据集里包含了1,000,000的用户以及这些用户与160,000的文章的交互行为。|[Microsoft(2020)](https://msnews.github.io)|
  |[movielens_pinterest_NCF](https://paddlerec.bj.bcebos.com/ncf/Data.zip)|论文原作者处理过的movielens数据集和pinterest数据集|[《Neural Collaborative Filtering 》](https://arxiv.org/pdf/1708.05031.pdf)|
+|[AmazonBook](https://paddlerec.bj.bcebos.com/datasets/AmazonBook/AmazonBook.tar.gz)|论文原作者处理过的AmazonBook数据集 |[《Controllable Multi-Interest Framework for Recommendation》](https://arxiv.org/abs/2005.09347)|

models/recall/mind/README.md

@@ -0,0 +1,125 @@
+# MIND(Multi-Interest Network with Dynamic Routing)
+
+以下是本例的简要目录结构及说明： 
+```
+├── data #样例数据
+│   ├── demo                    #demo训练数据
+│   │   └── demo.txt     
+│   └── valid                    #demo测试数据
+│       └── part-0    
+├── config.yaml                 #demo数据配置
+├── config_bigdata.yaml         #全量数据配置
+├── infer.py                    #评测动态图
+├── dygraph_model.py            #构建动态图
+├── mind_reader.py              #训练数据reader
+├── mind_infer_reader.py        #评测数据reader
+├── net.py                      #模型核心组网（动静合一）
+├── static_infer.py             #评测静态图
+└── static_model.py             #构建静态图
+```
+
+注：在阅读该示例前，建议您先了解以下内容：
+
+[paddlerec入门教程](https://github.com/PaddlePaddle/PaddleRec/blob/master/README.md)
+
+## 内容
+- [模型简介](#模型简介)
+- [数据准备](#数据准备)
+- [运行环境](#运行环境)
+- [快速开始](#快速开始)
+- [模型组网](#模型组网)
+- [效果复现](#效果复现)
+- [进阶使用](#进阶使用)
+- [FAQ](#FAQ)
+
+## 模型简介
+本例实现了基于动态路由的用户多兴趣网络，如下图所示：
+<p align="center">
+<img align="center" src="../../../doc/imgs/mind.png">
+<p>
+Multi-Interest Network with Dynamic Routing (MIND) 是通过构建用户和商品向量在统一的向量空间的多个用户兴趣向量，以表达用户多样的兴趣分布。然后通过向量召回技术，利用这多个兴趣向量去检索出TopK个与其近邻的商品向量，得到 TopK个 用户感兴趣的商品。其核心是一个基于胶囊网络和动态路由的（B2I Dynamic Routing）Multi-Interest Extractor Layer。
+
+推荐参考论文:[http://cn.arxiv.org/abs/1904.08030](http://cn.arxiv.org/abs/1904.08030)
+
+## 数据准备
+在模型目录的data目录下为您准备了快速运行的示例数据，训练数据、测试数据、词表文件依次保存在data/train, data/test文件夹中。若需要使用全量数据可以参考下方效果复现部分。
+
+训练数据的格式如下：
+```
+0,17978,0
+0,901,1
+0,97224,2
+0,774,3
+0,85757,4
+```
+分别表示uid、item_id和点击的顺序(时间戳)
+
+测试数据的格式如下：
+```
+user_id:487766 target_item:0 hist_item:17784 hist_item:126 hist_item:36 hist_item:124 hist_item:34 hist_item:1 hist_item:134 hist_item:6331 hist_item:141 hist_item:4336 hist_item:1373 eval_item:1062 eval_item:867 eval_item:62
+user_id:487793 target_item:0 hist_item:153428 hist_item:132997 hist_item:155723 hist_item:66546 hist_item:335397 hist_item:1926 eval_item:1122 eval_item:10105
+user_id:487805 target_item:0 hist_item:291025 hist_item:25190 hist_item:2820 hist_item:26047 hist_item:47259 hist_item:36376 eval_item:260145 eval_item:83865
+user_id:487811 target_item:0 hist_item:180837 hist_item:202701 hist_item:184587 hist_item:211642 eval_item:101621 eval_item:55716
+user_id:487820 target_item:0 hist_item:268524 hist_item:44318 hist_item:35153 hist_item:70847 eval_item:238318
+user_id:487825 target_item:0 hist_item:35602 hist_item:4353 hist_item:1540 hist_item:72921 eval_item:501
+```
+其中`hist_item`和`eval_item`均是变长序列，读取方式可以看`mind_infer_reader.py`
+
+## 运行环境
+PaddlePaddle>=2.0
+
+python 2.7/3.5/3.6/3.7
+
+os : windows/linux/macos
+
+## 快速开始
+
+在mind模型目录的快速执行命令如下：
+```
+# 进入模型目录
+# cd models/recall/mind # 在任意目录均可运行
+# 动态图训练
+python -u ../../../tools/trainer.py -m config.yaml 
+# 动态图预测
+python -u infer.py -m config.yaml -top_n 50  #对测试数据进行预测，并通过faiss召回候选结果评测Reacll、NDCG、HitRate指标
+
+# 静态图训练
+python -u ../../../tools/static_trainer.py -m config.yaml # 全量数据运行config_bigdata.yaml 
+# 静态图预测
+python -u static_infer.py -m config.yaml -top_n 50  #对测试数据进行预测，并通过faiss召回候选结果评测Reacll、NDCG、HitRate指标
+```
+
+## 模型组网
+
+细节见上面[模型简介](#模型简介)部分
+
+### 效果复现
+由于原始论文没有提供实验的复现细节，为了方便使用者能够快速的跑通每一个模型，我们使用论文[ComiRec](https://arxiv.org/abs/2005.09347)提供的AmazonBook数据集和训练任务进行复现。我们在每个模型下都提供了样例数据。如果需要复现readme中的效果,请按如下步骤依次操作即可。 
+
+在全量数据下模型的指标如下：
+| 模型 |  batch_size | epoch_num| Recall@50 | NDCG@50 | HitRate@50 |Time of each epoch |
+| :------| :------ | :------ | :------| :------ | :------|  :------ | 
+| mind(静态图) | 128 | 6 | 5.61% | 8.96% | 11.81% | -- |
+| mind(动态图) | 128 | 6 | 5.54% | 8.85% | 11.75% | -- |
+
+1. 确认您当前所在目录为PaddleRec/models/recall/mind
+2. 进入paddlerec/datasets/AmazonBook目录下执行run.sh脚本，会下载处理完成的AmazonBook数据集，并解压到指定目录
+```bash
+cd ../../../datasets/AmazonBook
+sh run.sh
+``` 
+3. 切回模型目录,执行命令运行全量数据
+```bash
+cd - # 切回模型目录
+# 动态图训练
+python -u ../../../tools/trainer.py -m config_bigdata.yaml # 全量数据运行config_bigdata
+python -u infer.py -m config_bigdata.yaml # 全量数据运行config_bigdata
+```
+
+## 进阶使用
+
+## FAQ
+
+## 参考
+
+数据集及训练任务参考了[ComiRec](https://github.com/THUDM/ComiRec)

models/recall/mind/config.yaml

@@ -0,0 +1,50 @@
+# 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.
+
+runner:
+  train_data_dir: "data/train"
+  train_reader_path: "mind_reader" # importlib format
+  use_gpu: True
+  use_auc: False
+  train_batch_size: 128
+  epochs: 2
+  print_interval: 500
+  model_save_path: "output_model_mind"
+  infer_batch_size: 128
+  infer_reader_path: "mind_infer_reader" # importlib format
+  test_data_dir: "data/valid"
+  infer_load_path: "output_model_mind"
+  infer_start_epoch: 0
+  infer_end_epoch: 1
+
+  # distribute_config
+  # sync_mode: "async"
+  # split_file_list: False
+  # thread_num: 1
+
+
+  # hyper parameters of user-defined network
+hyper_parameters:
+  # optimizer config
+  optimizer:
+    class: Adam
+    learning_rate: 0.005
+    # strategy: async
+  # user-defined <key, value> pairs
+  item_count: 367983
+  embedding_dim: 64
+  hidden_size: 64
+  neg_samples: 1280
+  maxlen: 20
+  pow_p: 1.0

models/recall/mind/config_bigdata.yaml

@@ -0,0 +1,50 @@
+# 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.
+
+runner:
+  train_data_dir: "../../../datasets/AmazonBook/train"
+  train_reader_path: "mind_reader" # importlib format
+  use_gpu: True
+  use_auc: False
+  train_batch_size: 128
+  epochs: 6
+  print_interval: 500
+  model_save_path: "output_model_mind"
+  infer_batch_size: 128
+  infer_reader_path: "mind_infer_reader" # importlib format
+  test_data_dir: "../../../datasets/AmazonBook/valid"
+  infer_load_path: "output_model_mind"
+  infer_start_epoch: 0
+  infer_end_epoch: 1
+
+  # distribute_config
+  # sync_mode: "async"
+  # split_file_list: False
+  # thread_num: 1
+
+
+  # hyper parameters of user-defined network
+hyper_parameters:
+  # optimizer config
+  optimizer:
+    class: Adam
+    learning_rate: 0.005
+    # strategy: async
+  # user-defined <key, value> pairs
+  item_count: 367983
+  embedding_dim: 64
+  hidden_size: 64
+  neg_samples: 1280
+  maxlen: 20
+  pow_p: 1.0