Merge branch 'master' of https://github.com/PaddlePaddle/PaddleRec into gru4rec

core/trainers/framework/runner.py

@@ -209,9 +209,13 @@ class RunnerBase(object):
 
                     if save_step_interval >= 1 and batch_id % save_step_interval == 0 and context[
                             "is_infer"] == False:
-                        if context["fleet_mode"].upper() == "PS":
-                            train_prog = context["model"][model_dict["name"]][
-                                "main_program"]
+                        if context["is_fleet"]:
+                            if context["fleet_mode"].upper() == "PS":
+                                train_prog = context["model"][model_dict[
+                                    "name"]]["main_program"]
+                            else:
+                                train_prog = context["model"][model_dict[
+                                    "name"]]["default_main_program"]
                         else:
                             train_prog = context["model"][model_dict["name"]][
                                 "default_main_program"]

doc/custom_reader.md

+# PaddleRec 自定义数据集及Reader
+
+用户自定义数据集及配置异步Reader，需要关注以下几个步骤：
+
+* [数据集整理](#数据集整理)
+* [在模型组网中加入输入占位符](#在模型组网中加入输入占位符)
+* [Reader实现](#Reader的实现)
+* [在yaml文件中配置Reader](#在yaml文件中配置reader)
+
+我们以CTR-DNN模型为例，给出了从数据整理，变量定义，Reader写法，调试的完整历程。
+
+* [数据及Reader示例-DNN](#数据及Reader示例-DNN)
+
+
+## 数据集整理
+
+PaddleRec支持模型自定义数据集。
+
+关于数据的tips：
+1. 数据量：
+
+    PaddleRec面向大规模数据设计，可以轻松支持亿级的数据读取，工业级的数据读写api：`dataset`在搜索、推荐、信息流等业务得到了充分打磨。
+2. 文件类型:
+
+    支持任意直接可读的文本数据，`dataset`同时支持`.gz`格式的文本压缩数据，无需额外代码，可直接读取。数据样本应以`\n`为标志，按行组织。
+
+3. 文件存放位置：
+
+    文件通常存放在训练节点本地，但同时，`dataset`支持使用`hadoop`远程读取数据，数据无需下载到本地，为dataset配置hadoop相关账户及地址即可。
+4. 数据类型
+
+    Reader处理的是以行为单位的`string`数据，喂入网络的数据需要转为`int`,`float`的数值数据，不支持`string`喂入网络，不建议明文保存及处理训练数据。
+5. Tips
+
+    Dataset模式下，训练线程与数据读取线程的关系强相关，为了多线程充分利用，`强烈建议将文件合理的拆为多个小文件`，尤其是在分布式训练场景下，可以均衡各个节点的数据量，同时加快数据的下载速度。
+
+## 在模型组网中加入输入占位符
+
+Reader读取文件后，产出的数据喂入网络，需要有占位符进行接收。占位符在Paddle中使用`fluid.data`或`fluid.layers.data`进行定义。`data`的定义可以参考[fluid.data](https://www.paddlepaddle.org.cn/documentation/docs/zh/api_cn/fluid_cn/data_cn.html#data)以及[fluid.layers.data](https://www.paddlepaddle.org.cn/documentation/docs/zh/api_cn/layers_cn/data_cn.html#data)。
+
+加入您希望输入三个数据，分别是维度32的数据A，维度变长的稀疏数据B，以及一个一维的标签数据C，并希望梯度可以经过该变量向前传递，则示例如下：
+
+数据A的定义：
+```python
+var_a = fluid.data(name='A', shape= [-1, 32], dtype='float32')
+```
+
+数据B的定义，变长数据的使用可以参考[LoDTensor](https://www.paddlepaddle.org.cn/documentation/docs/zh/beginners_guide/basic_concept/lod_tensor.html#cn-user-guide-lod-tensor)：
+```python
+var_b = fluid.data(name='B', shape=[-1, 1], lod_level=1, dtype='int64')
+```
+
+数据C的定义：
+```python
+var_c = fluid.data(name='C', shape=[-1, 1], dtype='int32')
+var_c.stop_gradient = False
+```
+
+当我们完成以上三个数据的定义后，在PaddleRec的模型定义中，还需将其加入model基类成员变量`self._data_var`
+
+```python
+self._data_var.append(var_a)
+self._data_var.append(var_b)
+self._data_var.append(var_c)
+```
+至此，我们完成了在组网中定义输入数据的工作。
+
+## Reader的实现
+
+### Reader的实现范式
+
+Reader的逻辑需要一个单独的python文件进行描述。我们试写一个`test_reader.py`，实现的具体流程如下：
+1. 首先我们需要引入Reader基类
+
+    ```python
+    from paddlerec.core.reader import ReaderBase
+    ```
+2. 创建一个子类，继承Reader的基类，训练所需Reader命名为`TrainerReader`
+    ```python
+    class Reader(ReaderBase):
+        def init(self):
+            pass
+
+        def generator_sample(self, line):
+            pass
+    ```
+
+3. 在`init(self)`函数中声明一些在数据读取中会用到的变量，必要时可以在`config.yaml`文件中配置变量，利用`env.get_global_env()`拿到。
+   
+    比如，我们希望从yaml文件中读取一个数据预处理变量`avg=10`，目的是将数据A的数据缩小10倍，可以这样实现：
+
+    首先更改yaml文件，在某个hyper_parameters下加入该变量
+
+    ```yaml
+    ...
+    hyper_parameters:
+        reader:
+            avg: 10
+    ...
+    ```
+
+
+    再更改Reader的init函数
+
+    ```python
+    from paddlerec.core.utils import envs
+    class Reader(ReaderBase):
+        def init(self):
+            self.avg = envs.get_global_env("avg", None, "hyper_parameters.reader")
+
+        def generator_sample(self, line):
+            pass
+    ```
+
+4. 继承并实现基类中的`generate_sample(self, line)`函数，逐行读取数据。
+   - 该函数应返回一个可以迭代的reader方法(带有yield的函数不再是一个普通的函数，而是一个生成器generator，成为了可以迭代的对象，等价于一个数组、链表、文件、字符串etc.)
+   - 在这个可以迭代的函数中，如示例代码中的`def reader()`，我们定义数据读取的逻辑。以行为单位的数据进行截取，转换及预处理。
+   - 最后，我们需要将数据整理为特定的格式，才能够被PaddleRec的Reader正确读取，并灌入的训练的网络中。简单来说，数据的输出顺序与我们在网络中创建的`inputs`必须是严格一一对应的，并转换为类似字典的形式。
+    
+    示例： 假设数据ABC在文本数据中，每行以这样的形式存储：
+    ```shell
+    0.1,0.2,0.3...3.0,3.1,3.2 \t 99999,99998,99997 \t 1 \n
+    ```
+
+    则示例代码如下：
+    ```python
+    from paddlerec.core.utils import envs
+    class Reader(ReaderBase):
+        def init(self):
+            self.avg = envs.get_global_env("avg", None, "hyper_parameters.reader")
+
+        def generator_sample(self, line):
+            
+            def reader(self, line):
+                # 先分割 '\n'， 再以 '\t'为标志分割为list
+                variables = (line.strip('\n')).split('\t')
+
+                # A是第一个元素，并且每个数据之间使用','分割
+                var_a = variables[0].split(',') # list
+                var_a = [float(i) / self.avg for i in var_a] # 将str数据转换为float
+                
+
+                # B是第二个元素，同样以 ',' 分割
+                var_b = variables[1].split(',') # list
+                var_b = [int(i) for i in var_b] # 将str数据转换为int
+
+                # C是第三个元素, 只有一个元素，没有分割符
+                var_c = variables[2]
+                var_c = int(var_c) # 将str数据转换为int
+                var_c = [var_c] # 将单独的数据元素置入list中
+
+                # 将数据与数据名结合，组织为dict的形式
+                # 如下，output形式为{ A: var_a, B: var_b, C: var_c}
+                variable_name = ['A', 'B', 'C']
+                output = zip(variable_name, [var_a] + [var_b] + [var_c])
+
+                # 将数据输出，使用yield方法，将该函数变为了一个可迭代的对象
+                yield output
+
+    ```
+    
+    至此，我们完成了Reader的实现。
+
+
+### 在yaml文件中配置Reader
+
+在模型的yaml配置文件中，主要的修改是三个，如下
+
+```yaml
+reader:
+    batch_size: 2
+    class: "{workspace}/criteo_reader.py"
+    train_data_path: "{workspace}/data/train_data"
+```
+
+batch_size: 顾名思义，是小批量训练时的样本大小
+class: 运行改模型所需reader的路径
+train_data_path: 训练数据所在文件夹
+reader_debug_mode: 测试reader语法，及输出是否符合预期的debug模式的开关
+
+
+## 数据及Reader示例-DNN
+
+
+### Criteo数据集格式
+
+CTR-DNN训练及测试数据集选用[Display Advertising Challenge](https://www.kaggle.com/c/criteo-display-ad-challenge/)所用的Criteo数据集。该数据集包括两部分：训练集和测试集。训练集包含一段时间内Criteo的部分流量，测试集则对应训练数据后一天的广告点击流量。
+每一行数据格式如下所示：
+```bash
+<label> <integer feature 1> ... <integer feature 13> <categorical feature 1> ... <categorical feature 26>
+```
+其中```<label>```表示广告是否被点击，点击用1表示，未点击用0表示。```<integer feature>```代表数值特征（连续特征），共有13个连续特征。```<categorical feature>```代表分类特征（离散特征），共有26个离散特征。相邻两个特征用```\t```分隔，缺失特征用空格表示。测试集中```<label>```特征已被移除。
+
+### Criteo数据集的预处理
+
+数据预处理共包括两步：
+- 将原始训练集按9:1划分为训练集和验证集
+- 数值特征（连续特征）需进行归一化处理，但需要注意的是，对每一个特征```<integer feature i>```，归一化时用到的最大值并不是用全局最大值，而是取排序后95%位置处的特征值作为最大值，同时保留极值。
+
+### CTR网络输入的定义
+
+正如前所述，Criteo数据集中，分为连续数据与离散（稀疏）数据，所以整体而言，CTR-DNN模型的数据输入层包括三个，分别是：`dense_input`用于输入连续数据，维度由超参数`dense_feature_dim`指定，数据类型是归一化后的浮点型数据。`sparse_input_ids`用于记录离散数据，在Criteo数据集中，共有26个slot，所以我们创建了名为`C1~C26`的26个稀疏参数输入，并设置`lod_level=1`，代表其为变长数据，数据类型为整数；最后是每条样本的`label`，代表了是否被点击，数据类型是整数，0代表负样例，1代表正样例。
+
+在Paddle中数据输入的声明使用`paddle.fluid.layers.data()`，会创建指定类型的占位符，数据IO会依据此定义进行数据的输入。
+
+稀疏参数输入的定义:
+```python
+def sparse_inputs():
+    ids = envs.get_global_env("hyper_parameters.sparse_inputs_slots", None)
+
+    sparse_input_ids = [
+        fluid.layers.data(name="S" + str(i),
+                            shape=[1],
+                            lod_level=1,
+                            dtype="int64") for i in range(1, ids)
+    ]
+    return sparse_input_ids
+```
+
+稠密参数输入的定义：
+```python
+def dense_input():
+    dim = envs.get_global_env("hyper_parameters.dense_input_dim", None)
+
+    dense_input_var = fluid.layers.data(name="D",
+                                        shape=[dim],
+                                        dtype="float32")
+    return dense_input_var
+```
+
+标签的定义：
+```python
+def label_input():
+    label = fluid.layers.data(name="click", shape=[1], dtype="int64")
+    return label
+```
+
+组合起来，正确的声明他们：
+```python
+self.sparse_inputs = sparse_inputs()
+self.dense_input = dense_input()
+self.label_input = label_input()
+
+self._data_var.append(self.dense_input)
+
+for input in self.sparse_inputs:
+    self._data_var.append(input)
+
+self._data_var.append(self.label_input)
+
+```
+
+
+### Criteo Reader写法
+
+```python
+# 引入PaddleRec的Reader基类
+from paddlerec.core.reader import ReaderBase
+# 引入PaddleRec的读取yaml配置文件的方法
+from paddlerec.core.utils import envs
+
+# 定义TrainReader，需要继承 paddlerec.core.reader.Reader
+class Reader(ReaderBase):
+
+    # 数据预处理逻辑，继承自基类
+    # 如果无需处理， 使用pass跳过该函数的执行
+    def init(self):
+        self.cont_min_ = [0, -3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
+        self.cont_max_ = [20, 600, 100, 50, 64000, 500, 100, 50, 500, 10, 10, 10, 50]
+        self.cont_diff_ = [20, 603, 100, 50, 64000, 500, 100, 50, 500, 10, 10, 10, 50]
+        self.hash_dim_ = envs.get_global_env("hyper_parameters.sparse_feature_number", None)
+        self.continuous_range_ = range(1, 14)
+        self.categorical_range_ = range(14, 40)
+
+    # 读取数据方法，继承自基类
+    # 实现可以迭代的reader函数，逐行处理数据
+    def generate_sample(self, line):
+        """
+        Read the data line by line and process it as a dictionary
+        """
+
+        def reader():
+            """
+            This function needs to be implemented by the user, based on data format
+            """
+            features = line.rstrip('\n').split('\t')
+
+            dense_feature = []
+            sparse_feature = []
+            for idx in self.continuous_range_:
+                if features[idx] == "":
+                    dense_feature.append(0.0)
+                else:
+                    dense_feature.append(
+                        (float(features[idx]) - self.cont_min_[idx - 1]) /
+                        self.cont_diff_[idx - 1])
+
+            for idx in self.categorical_range_:
+                sparse_feature.append(
+                    [hash(str(idx) + features[idx]) % self.hash_dim_])
+            label = [int(features[0])]
+            feature_name = ["D"]
+            for idx in self.categorical_range_:
+                feature_name.append("S" + str(idx - 13))
+            feature_name.append("label")
+            yield zip(feature_name, [dense_feature] + sparse_feature + [label])
+
+        return reader
+```

models/contentunderstanding/tagspace/data/text2paddle.py

@@ -18,6 +18,7 @@ import collections
 import os
 import csv
 import re
+import io
 import sys
 if six.PY2:
     reload(sys)
@@ -45,11 +46,11 @@ def build_dict(column_num=2, min_word_freq=0, train_dir="", test_dir=""):
     word_freq = collections.defaultdict(int)
     files = os.listdir(train_dir)
     for fi in files:
-        with open(os.path.join(train_dir, fi), "r", encoding='utf-8') as f:
+        with io.open(os.path.join(train_dir, fi), "r", encoding='utf-8') as f:
             word_freq = word_count(column_num, f, word_freq)
     files = os.listdir(test_dir)
     for fi in files:
-        with open(os.path.join(test_dir, fi), "r", encoding='utf-8') as f:
+        with io.open(os.path.join(test_dir, fi), "r", encoding='utf-8') as f:
             word_freq = word_count(column_num, f, word_freq)
 
     word_freq = [x for x in six.iteritems(word_freq) if x[1] > min_word_freq]
@@ -65,51 +66,51 @@ def write_paddle(text_idx, tag_idx, train_dir, test_dir, output_train_dir,
     if not os.path.exists(output_train_dir):
         os.mkdir(output_train_dir)
     for fi in files:
-        with open(os.path.join(train_dir, fi), "r", encoding='utf-8') as f:
-            with open(
+        with io.open(os.path.join(train_dir, fi), "r", encoding='utf-8') as f:
+            with io.open(
                     os.path.join(output_train_dir, fi), "w",
                     encoding='utf-8') as wf:
                 data_file = csv.reader(f)
                 for row in data_file:
                     tag_raw = re.split(r'\W+', row[0].strip())
                     pos_index = tag_idx.get(tag_raw[0])
-                    wf.write(str(pos_index) + ",")
+                    wf.write(u"{},".format(str(pos_index)))
                     text_raw = re.split(r'\W+', row[2].strip())
                     l = [text_idx.get(w) for w in text_raw]
                     for w in l:
-                        wf.write(str(w) + " ")
-                    wf.write("\n")
+                        wf.write(u"{} ".format(str(w)))
+                    wf.write(u"\n")
 
     files = os.listdir(test_dir)
     if not os.path.exists(output_test_dir):
         os.mkdir(output_test_dir)
     for fi in files:
-        with open(os.path.join(test_dir, fi), "r", encoding='utf-8') as f:
-            with open(
+        with io.open(os.path.join(test_dir, fi), "r", encoding='utf-8') as f:
+            with io.open(
                     os.path.join(output_test_dir, fi), "w",
                     encoding='utf-8') as wf:
                 data_file = csv.reader(f)
                 for row in data_file:
                     tag_raw = re.split(r'\W+', row[0].strip())
                     pos_index = tag_idx.get(tag_raw[0])
-                    wf.write(str(pos_index) + ",")
+                    wf.write(u"{},".format(str(pos_index)))
                     text_raw = re.split(r'\W+', row[2].strip())
                     l = [text_idx.get(w) for w in text_raw]
                     for w in l:
-                        wf.write(str(w) + " ")
-                    wf.write("\n")
+                        wf.write(u"{} ".format(str(w)))
+                    wf.write(u"\n")
 
 
 def text2paddle(train_dir, test_dir, output_train_dir, output_test_dir,
                 output_vocab_text, output_vocab_tag):
     print("start constuct word dict")
     vocab_text = build_dict(2, 0, train_dir, test_dir)
-    with open(output_vocab_text, "w", encoding='utf-8') as wf:
-        wf.write(str(len(vocab_text)) + "\n")
+    with io.open(output_vocab_text, "w", encoding='utf-8') as wf:
+        wf.write(u"{}\n".format(str(len(vocab_text))))
 
     vocab_tag = build_dict(0, 0, train_dir, test_dir)
-    with open(output_vocab_tag, "w", encoding='utf-8') as wf:
-        wf.write(str(len(vocab_tag)) + "\n")
+    with io.open(output_vocab_tag, "w", encoding='utf-8') as wf:
+        wf.write(u"{}\n".format(str(len(vocab_tag))))
 
     print("construct word dict done\n")
     write_paddle(vocab_text, vocab_tag, train_dir, test_dir, output_train_dir,

models/match/dssm/config.yaml

@@ -29,11 +29,12 @@ dataset:
 
 hyper_parameters:
   optimizer:
-    class: sgd
+    class: adam
     learning_rate: 0.001
-    strategy: async
-  trigram_d: 1439
+    strategy: sync
+  trigram_d: 2900
   neg_num: 1
+  slice_end: 8
   fc_sizes: [300, 300, 128]
   fc_acts: ['tanh', 'tanh', 'tanh']
 
@@ -44,7 +45,7 @@ runner:
 - name: train_runner
   class: train
   # num of epochs
-  epochs: 3
+  epochs: 1
   # device to run training or infer
   device: cpu
   save_checkpoint_interval: 1 # save model interval of epochs
@@ -54,14 +55,14 @@ runner:
   save_inference_feed_varnames: ["query", "doc_pos"] # feed vars of save inference
   save_inference_fetch_varnames: ["cos_sim_0.tmp_0"] # fetch vars of save inference
   init_model_path: "" # load model path
-  print_interval: 2
+  print_interval: 10
   phases: phase1
 - name: infer_runner
   class: infer
   # device to run training or infer
   device: cpu
   print_interval: 1
-  init_model_path: "increment/2" # load model path
+  init_model_path: "increment/0" # load model path
   phases: phase2
 
 # runner will run all the phase in each epoch

models/match/dssm/data/data_process.sh

+# 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.
+#!/bin/bash
+
+wget https://paddlerec.bj.bcebos.com/dssm%2Fbq.tar.gz
+tar xzf dssm%2Fbq.tar.gz
+rm -f dssm%2Fbq.tar.gz
+mv bq/train.txt ./raw_data.txt
+python3 preprocess.py
+mkdir big_train
+mv train.txt ./big_train
+mkdir big_test
+mv test.txt ./big_test

models/match/dssm/data/preprocess.py

+#encoding=utf-8
 # Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -11,14 +12,14 @@
 # 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.
-#encoding=utf-8
 
 import os
 import sys
+import jieba
 import numpy as np
 import random
 
-f = open("./zhidao", "r")
+f = open("./raw_data.txt", "r")
 lines = f.readlines()
 f.close()
 
@@ -26,14 +27,15 @@ f.close()
 word_dict = {}
 for line in lines:
     line = line.strip().split("\t")
-    text = line[0].split(" ") + line[1].split(" ")
+    text = line[0].strip("") + " " + line[1].strip("")
+    text = jieba.cut(text)
     for word in text:
         if word in word_dict:
             continue
         else:
             word_dict[word] = len(word_dict) + 1
 
-f = open("./zhidao", "r")
+f = open("./raw_data.txt", "r")
 lines = f.readlines()
 f.close()
 
@@ -57,12 +59,13 @@ for line in lines:
         else:
             pos_dict[line[0]] = [line[1]]
 
+print("build dict done")
 #划分训练集和测试集
 query_list = list(pos_dict.keys())
-#print(len(query))
-random.shuffle(query_list)
-train_query = query_list[:90]
-test_query = query_list[90:]
+#print(len(query_list))
+#random.shuffle(query_list)
+train_query = query_list[:11600]
+test_query = query_list[11600:]
 
 #获得训练集
 train_set = []
@@ -73,6 +76,7 @@ for query in train_query:
         for neg in neg_dict[query]:
             train_set.append([query, pos, neg])
 random.shuffle(train_set)
+print("get train_set done")
 
 #获得测试集
 test_set = []
@@ -84,13 +88,14 @@ for query in test_query:
     for neg in neg_dict[query]:
         test_set.append([query, neg, 0])
 random.shuffle(test_set)
+print("get test_set done")
 
 #训练集中的query,pos,neg转化为词袋
 f = open("train.txt", "w")
 for line in train_set:
-    query = line[0].strip().split(" ")
-    pos = line[1].strip().split(" ")
-    neg = line[2].strip().split(" ")
+    query = jieba.cut(line[0].strip())
+    pos = jieba.cut(line[1].strip())
+    neg = jieba.cut(line[2].strip())
     query_token = [0] * (len(word_dict) + 1)
     for word in query:
         query_token[word_dict[word]] = 1
@@ -109,8 +114,8 @@ f.close()
 f = open("test.txt", "w")
 fa = open("label.txt", "w")
 for line in test_set:
-    query = line[0].strip().split(" ")
-    pos = line[1].strip().split(" ")
+    query = jieba.cut(line[0].strip())
+    pos = jieba.cut(line[1].strip())
     label = line[2]
     query_token = [0] * (len(word_dict) + 1)
     for word in query:

models/match/dssm/model.py

@@ -29,6 +29,7 @@ class Model(ModelBase):
         self.hidden_acts = envs.get_global_env("hyper_parameters.fc_acts")
         self.learning_rate = envs.get_global_env(
             "hyper_parameters.learning_rate")
+        self.slice_end = envs.get_global_env("hyper_parameters.slice_end")
 
     def input_data(self, is_infer=False, **kwargs):
         query = fluid.data(
@@ -94,7 +95,7 @@ class Model(ModelBase):
         prob = fluid.layers.softmax(concat_Rs, axis=1)
 
         hit_prob = fluid.layers.slice(
-            prob, axes=[0, 1], starts=[0, 0], ends=[8, 1])
+            prob, axes=[0, 1], starts=[0, 0], ends=[self.slice_end, 1])
         loss = -fluid.layers.reduce_sum(fluid.layers.log(hit_prob))
         avg_cost = fluid.layers.mean(x=loss)
         self._cost = avg_cost

models/match/dssm/readme.md

@@ -4,11 +4,12 @@
 
 ```
 ├── data #样例数据
-	├── train
-		├── train.txt #训练数据样例
-	├── test
-    	├── test.txt #测试数据样例
-	├── preprocess.py #数据处理程序
+    ├── train
+        ├── train.txt #训练数据样例
+    ├── test
+        ├── test.txt #测试数据样例
+    ├── preprocess.py #数据处理程序
+    ├── data_process #数据一键处理脚本
 ├── __init__.py
 ├── README.md #文档
 ├── model.py #模型文件
@@ -46,13 +47,19 @@ Query 和 Doc 的语义相似性可以用这两个向量的 cosine 距离表示
 <p>
 
 ## 数据准备
-我们公开了自建的测试集，包括百度知道、ECOM、QQSIM、UNICOM 四个数据集。这里我们选取百度知道数据集来进行训练。执行以下命令可以获取上述数据集。
+BQ是一个智能客服中文问句匹配数据集，该数据集是自动问答系统语料，共有120,000对句子对，并标注了句子对相似度值。数据中存在错别字、语法不规范等问题，但更加贴近工业场景。执行以下命令可以获取上述数据集。
 ```
-wget --no-check-certificate https://baidu-nlp.bj.bcebos.com/simnet_dataset-1.0.0.tar.gz
-tar xzf simnet_dataset-1.0.0.tar.gz
-rm simnet_dataset-1.0.0.tar.gz
+wget https://paddlerec.bj.bcebos.com/dssm%2Fbq.tar.gz
+tar xzf dssm%2Fbq.tar.gz
+rm -f dssm%2Fbq.tar.gz
+```
+数据集样例：
+```
+请问一天是否都是限定只能转入或转出都是五万。    微众多少可以赎回短期理财        0
+微粒咨询电话号码多少    你们的人工客服电话是多少        1
+已经在银行换了新预留号码。      我现在换了电话号码，这个需要更换吗      1
+每个字段以tab键分隔，第1，2列表示两个文本。第3列表示类别（0或1，0表示两个文本不相似，1表示两个文本相似）。
 ```
-
 ## 运行环境
 PaddlePaddle>=1.7.2
 
@@ -120,21 +127,24 @@ PaddleRec Finish
 2. 在data目录下载并解压数据集，命令如下：  
 ``` 
 cd data
-wget --no-check-certificate https://baidu-nlp.bj.bcebos.com/simnet_dataset-1.0.0.tar.gz
-tar xzf simnet_dataset-1.0.0.tar.gz
-rm simnet_dataset-1.0.0.tar.gz
+wget https://paddlerec.bj.bcebos.com/dssm%2Fbq.tar.gz
+tar xzf dssm%2Fbq.tar.gz
+rm -f dssm%2Fbq.tar.gz
 ```
-3. 本文提供了快速将数据集中的汉字数据处理为可训练格式数据的脚本，您在解压数据集后，可以看见目录中存在一个名为zhidao的文件。然后能可以在python3环境下运行我们提供的preprocess.py文件。即可生成可以直接用于训练的数据目录test.txt,train.txt和label.txt。将其放入train和test目录下以备训练时调用。命令如下：
+3. 本文提供了快速将数据集中的汉字数据处理为可训练格式数据的脚本，您在解压数据集后，可以看见目录中存在一个名为bq的目录。将其中的train.txt文件移动到data目录下，然后可以在python3环境下运行我们提供的preprocess.py文件。即可生成可以直接用于训练的数据目录test.txt,train.txt和label.txt。将其放入train和test目录下以备训练时调用。生成时间较长，请耐心等待。命令如下：
 ```
-mv data/zhidao ./
-rm -rf data
+mv bq/train.txt ./raw_data.txt
 python3 preprocess.py
-rm -f ./train/train.txt
-mv train.txt ./train
-rm -f ./test/test.txt
-mv test.txt test
+mkdir big_train
+mv train.txt ./big_train
+mkdir big_test
+mv test.txt ./big_test
 cd ..
 ```
+也可以使用我们提供的一键数据处理脚本data_process.sh
+```
+sh data_process.sh
+```
 经过预处理的格式：  
 训练集为三个稀疏的BOW方式的向量：query,pos,neg  
 测试集为两个稀疏的BOW方式的向量：query,pos  
@@ -144,8 +154,10 @@ label.txt中对应的测试集中的标签
 
 将workspace改为您当前的绝对路径。（可用pwd命令获取绝对路径）  
 将dataset_train中的batch_size从8改为128
-将文件model.py中的 hit_prob = fluid.layers.slice(prob, axes=[0, 1], starts=[0, 0], ends=[8, 1])  
-    改为hit_prob = fluid.layers.slice(prob, axes=[0, 1], starts=[0, 0], ends=[128, 1]).当您需要改变batchsize的时候，end中第一个参数也需要随之变化
+将hyper_parameters中的slice_end从8改为128.当您需要改变batchsize的时候，这个参数也需要随之变化
+将dataset_train中的data_path改为{workspace}/data/big_train
+将dataset_infer中的data_path改为{workspace}/data/big_test
+将hyper_parameters中的trigram_d改为5913
 
 5.  执行脚本，开始训练.脚本会运行python -m paddlerec.run -m ./config.yaml启动训练，并将结果输出到result文件中。然后启动transform.py整合数据，最后计算出正逆序指标：
 ```
@@ -155,26 +167,14 @@ sh run.sh
 输出结果示例：
 ```
 ................run.................
-!!! The CPU_NUM is not specified, you should set CPU_NUM in the environment variable list.
-CPU_NUM indicates that how many CPUPlace are used in the current task.
-And if this parameter are set as N (equal to the number of physical CPU core) the program may be faster.
-
-export CPU_NUM=32 # for example, set CPU_NUM as number of physical CPU core which is 32.
-
-!!! The default number of CPU_NUM=1.
-I0821 07:16:04.512531 32200 parallel_executor.cc:440] The Program will be executed on CPU using ParallelExecutor, 1 cards are used, so 1 programs are executed in parallel.
-I0821 07:16:04.515708 32200 build_strategy.cc:365] SeqOnlyAllReduceOps:0, num_trainers:1
-I0821 07:16:04.518872 32200 parallel_executor.cc:307] Inplace strategy is enabled, when build_strategy.enable_inplace = True
-I0821 07:16:04.520995 32200 parallel_executor.cc:375] Garbage collection strategy is enabled, when FLAGS_eager_delete_tensor_gb = 0
-75
-pnr: 2.25581395349
-query_num: 11
-pair_num: 184 184
-equal_num: 44
-正序率： 0.692857142857
-97 43
-```
-6. 提醒：因为采取较小的数据集进行训练和测试，得到指标的浮动程度会比较大。如果得到的指标不合预期，可以多次执行步骤5，即可获得合理的指标。
+8989
+pnr:2.75621659307
+query_num:1369
+pair_num:16240 , 16240
+equal_num:77
+正序率: 0.733774670544
+pos_num: 11860 , neg_num: 4303
+```
 
 ## 进阶使用
   

models/match/dssm/run.sh

@@ -13,7 +13,7 @@
 # limitations under the License.
 #!/bin/bash
 echo "................run................."
-python -m paddlerec.run -m ./config.yaml >result1.txt
+python -m paddlerec.run -m ./config.yaml &> result1.txt
 grep -i "query_doc_sim" ./result1.txt >./result2.txt
 sed '$d' result2.txt >result.txt
 rm -f result1.txt

models/match/dssm/transform.py

@@ -32,13 +32,13 @@ filename = './result.txt'
 sim = []
 for line in open(filename):
     line = line.strip().split(",")
-    line[1] = line[1].split(":")
-    line = line[1][1].strip(" ")
+    line[3] = line[3].split(":")
+    line = line[3][1].strip(" ")
     line = line.strip("[")
     line = line.strip("]")
     sim.append(float(line))
 
-filename = './data/test/test.txt'
+filename = './data/big_test/test.txt'
 f = open(filename, "r")
 f.readline()
 query = []

models/match/match-pyramid/data/process.py

@@ -106,7 +106,7 @@ def make_train():
                         pair_list.append((d1, high_d2, low_d2))
     print('Pair Instance Count:', len(pair_list))
 
-    f = open("./data/train/train.txt", "w")
+    f = open("./data/big_train/train.txt", "w")
     for batch in range(800):
         X1 = np.zeros((batch_size * 2, data1_maxlen), dtype=np.int32)
         X2 = np.zeros((batch_size * 2, data2_maxlen), dtype=np.int32)
@@ -131,7 +131,7 @@ def make_train():
 def make_test():
     rel = read_relation(filename=os.path.join(Letor07Path,
                                               'relation.test.fold1.txt'))
-    f = open("./data/test/test.txt", "w")
+    f = open("./data/big_test/test.txt", "w")
     for label, d1, d2 in rel:
         X1 = np.zeros(data1_maxlen, dtype=np.int32)
         X2 = np.zeros(data2_maxlen, dtype=np.int32)

models/match/match-pyramid/data_process.sh

@@ -3,7 +3,9 @@
 echo "...........load  data................."
 wget --no-check-certificate 'https://paddlerec.bj.bcebos.com/match_pyramid/match_pyramid_data.tar.gz'
 mv ./match_pyramid_data.tar.gz ./data
-rm -rf ./data/relation.test.fold1.txt ./data/realtion.train.fold1.txt
+rm -rf ./data/relation.test.fold1.txt 
 tar -xvf ./data/match_pyramid_data.tar.gz
+mkdir ./data/big_train
+mkdir ./data/big_test
 echo "...........data process..............."
 python ./data/process.py

models/match/match-pyramid/eval.py

@@ -49,8 +49,8 @@ filename = './result.txt'
 pred = []
 for line in open(filename):
     line = line.strip().split(",")
-    line[1] = line[1].split(":")
-    line = line[1][1].strip(" ")
+    line[3] = line[3].split(":")
+    line = line[3][1].strip(" ")
     line = line.strip("[")
     line = line.strip("]")
     pred.append(float(line))

models/match/match-pyramid/readme.md

@@ -56,10 +56,10 @@
 4.嵌入层文件：我们将预训练的词向量存储在嵌入文件中。例如：embed_wiki-pdc_d50_norm  
 
 ## 运行环境
-PaddlePaddle>=1.7.2
-python 2.7/3.5/3.6/3.7
-PaddleRec >=0.1
-os : windows/linux/macos
+PaddlePaddle>=1.7.2  
+python 2.7/3.5/3.6/3.7  
+PaddleRec >=0.1  
+os : windows/linux/macos  
 
 ## 快速开始
 
@@ -72,7 +72,7 @@ python -m paddlerec.run -m models/match/match-pyramid/config.yaml
 ## 论文复现
 1. 确认您当前所在目录为PaddleRec/models/match/match-pyramid
 2. 本文提供了原数据集的下载以及一键生成训练和测试数据的预处理脚本，您可以直接一键运行:bash data_process.sh  
-执行该脚本，会从国内源的服务器上下载Letor07数据集，删除掉data文件夹中原有的relation.test.fold1.txt和relation.train.fold1.txt，并将完整的数据集解压到data文件夹。随后运行 process.py 将全量训练数据放置于`./data/train`，全量测试数据放置于`./data/test`。并生成用于初始化embedding层的embedding.npy文件  
+执行该脚本，会从国内源的服务器上下载Letor07数据集，并将完整的数据集解压到data文件夹。随后运行 process.py 将全量训练数据放置于`./data/big_train`，全量测试数据放置于`./data/big_test`。并生成用于初始化embedding层的embedding.npy文件  
 执行该脚本的理想输出为：  
 ```
 bash data_process.sh
@@ -123,6 +123,8 @@ data/embed_wiki-pdc_d50_norm
 3. 打开文件config.yaml,更改其中的参数  
 
 将workspace改为您当前的绝对路径。（可用pwd命令获取绝对路径）
+将dataset_train下的data_path参数改为{workspace}/data/big_train
+将dataset_infer下的data_path参数改为{workspace}/data/big_test
 
 4. 随后，您直接一键运行：bash run.sh  即可得到复现的论文效果
 执行该脚本后，会执行python -m paddlerec.run -m ./config.yaml 命令开始训练并测试模型，将测试的结果保存到result.txt文件，最后通过执行eval.py进行评估得到数据的map指标  
@@ -131,7 +133,7 @@ data/embed_wiki-pdc_d50_norm
 ..............test.................
 13651
 336
-('map=', 0.420878322843591)
+('map=', 0.3993127885738651)
 ```  
 ## 进阶使用
   

models/match/match-pyramid/run.sh

 #!/bin/bash
 echo "................run................."
-python -m paddlerec.run -m ./config.yaml >result1.txt
-grep -i "prediction" ./result1.txt >./result.txt
+python -m paddlerec.run -m ./config.yaml &>result1.txt
+grep -i "prediction" ./result1.txt >./result2.txt
+sed '$d' result2.txt >result.txt
+rm -f result2.txt
 rm -f result1.txt
 python eval.py

models/match/multiview-simnet/config.yaml

@@ -26,19 +26,19 @@ dataset:
   batch_size: 1
   type: DataLoader # or QueueDataset
   data_path: "{workspace}/data/test"
-  sparse_slots: "1 2"
+  sparse_slots: "0 1"
 
 # hyper parameters of user-defined network
 hyper_parameters:
   optimizer:
     class: Adam
-    learning_rate: 0.0001
-    strategy: async
+    learning_rate: 0.001
+    strategy: sync
   query_encoder: "gru"
   title_encoder: "gru"
   query_encode_dim: 128
   title_encode_dim: 128
-  sparse_feature_dim: 1439
+  sparse_feature_dim: 6327
   embedding_dim: 128
   hidden_size: 128
   margin: 0.1

models/match/multiview-simnet/data/data_process.sh

+# 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.
+#!/bin/bash
+
+wget https://paddlerec.bj.bcebos.com/dssm%2Fbq.tar.gz
+tar xzf dssm%2Fbq.tar.gz
+rm -f dssm%2Fbq.tar.gz
+mv bq/train.txt ./raw_data.txt
+python3 preprocess.py
+mkdir big_train
+mv train.txt ./big_train
+mkdir big_test
+mv test.txt ./big_test

models/match/multiview-simnet/data/preprocess.py

+#encoding=utf-8
 # Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -11,14 +12,14 @@
 # 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.
-#encoding=utf-8
 
 import os
 import sys
+import jieba
 import numpy as np
 import random
 
-f = open("./zhidao", "r")
+f = open("./raw_data.txt", "r")
 lines = f.readlines()
 f.close()
 
@@ -26,14 +27,15 @@ f.close()
 word_dict = {}
 for line in lines:
     line = line.strip().split("\t")
-    text = line[0].split(" ") + line[1].split(" ")
+    text = line[0].strip("") + line[1].strip("")
+    text = jieba.cut(text)
     for word in text:
         if word in word_dict:
             continue
         else:
             word_dict[word] = len(word_dict) + 1
 
-f = open("./zhidao", "r")
+f = open("./raw_data.txt", "r")
 lines = f.readlines()
 f.close()
 
@@ -59,10 +61,10 @@ for line in lines:
 
 #划分训练集和测试集
 query_list = list(pos_dict.keys())
-#print(len(query_list))
+print(len(query_list))
 random.shuffle(query_list)
-train_query = query_list[:90]
-test_query = query_list[90:]
+train_query = query_list[:11600]
+test_query = query_list[11600:]
 
 #获得训练集
 train_set = []
@@ -88,9 +90,9 @@ random.shuffle(test_set)
 #训练集中的query,pos,neg转化格式
 f = open("train.txt", "w")
 for line in train_set:
-    query = line[0].strip().split(" ")
-    pos = line[1].strip().split(" ")
-    neg = line[2].strip().split(" ")
+    query = jieba.cut(line[0].strip())
+    pos = jieba.cut(line[1].strip())
+    neg = jieba.cut(line[2].strip())
     query_list = []
     for word in query:
         query_list.append(word_dict[word])
@@ -110,8 +112,8 @@ f = open("test.txt", "w")
 fa = open("label.txt", "w")
 fb = open("testquery.txt", "w")
 for line in test_set:
-    query = line[0].strip().split(" ")
-    pos = line[1].strip().split(" ")
+    query = jieba.cut(line[0].strip())
+    pos = jieba.cut(line[1].strip())
     label = line[2]
     query_list = []
     for word in query:

models/match/multiview-simnet/readme.md

@@ -4,11 +4,12 @@
 
 ```
 ├── data #样例数据
-	├── train
-		├── train.txt #训练数据样例
-	├── test
-    	├── test.txt #测试数据样例
-	├── preprocess.py #数据处理程序
+    ├── train
+        ├── train.txt #训练数据样例
+    ├── test
+        ├── test.txt #测试数据样例
+    ├── preprocess.py #数据处理程序
+    ├── data_process.sh #一键数据处理脚本
 ├── __init__.py
 ├── README.md #文档
 ├── model.py #模型文件
@@ -42,14 +43,20 @@
 <p>
 
 ## 数据准备
-我们公开了自建的测试集，包括百度知道、ECOM、QQSIM、UNICOM 四个数据集。这里我们选取百度知道数据集来进行训练。执行以下命令可以获取上述数据集。
+BQ是一个智能客服中文问句匹配数据集，该数据集是自动问答系统语料，共有120,000对句子对，并标注了句子对相似度值。数据中存在错别字、语法不规范等问题，但更加贴近工业场景。执行以下命令可以获取上述数据集。
 ```
-wget --no-check-certificate https://baidu-nlp.bj.bcebos.com/simnet_dataset-1.0.0.tar.gz
-tar xzf simnet_dataset-1.0.0.tar.gz
-rm simnet_dataset-1.0.0.tar.gz
+wget https://paddlerec.bj.bcebos.com/dssm%2Fbq.tar.gz
+tar xzf dssm%2Fbq.tar.gz
+rm -f dssm%2Fbq.tar.gz
 ```
-
-数据格式为一个标识句子的slot，后跟一个句子中词的token。两者形成{slot：token}的形式标识一个词：  
+数据集样例：
+```
+请问一天是否都是限定只能转入或转出都是五万。    微众多少可以赎回短期理财        0
+微粒咨询电话号码多少    你们的人工客服电话是多少        1
+已经在银行换了新预留号码。      我现在换了电话号码，这个需要更换吗      1
+每个字段以tab键分隔，第1，2列表示两个文本。第3列表示类别（0或1，0表示两个文本不相似，1表示两个文本相似）。
+```
+最终输出的数据格式为一个标识句子的slot，后跟一个句子中词的token。两者形成{slot：token}的形式标识一个词：  
 ```
 0:358 0:206 0:205 0:250 0:9 0:3 0:207 0:10 0:330 0:164 1:1144 1:217 1:206 1:9 1:3 1:207 1:10 1:398 1:2 2:217 2:206 2:9 2:3 2:207 2:10 2:398 2:2
 0:358 0:206 0:205 0:250 0:9 0:3 0:207 0:10 0:330 0:164 1:951 1:952 1:206 1:9 1:3 1:207 1:10 1:398 2:217 2:206 2:9 2:3 2:207 2:10 2:398 2:2
@@ -75,24 +82,29 @@ python -m paddlerec.run -m models/match/multiview-simnet/config.yaml
 2. 在data目录下载并解压数据集，命令如下：  
 ``` 
 cd data
-wget --no-check-certificate https://baidu-nlp.bj.bcebos.com/simnet_dataset-1.0.0.tar.gz
-tar xzf simnet_dataset-1.0.0.tar.gz
-rm -f simnet_dataset-1.0.0.tar.gz
-mv data/zhidao ./
-rm -rf data
+wget https://paddlerec.bj.bcebos.com/dssm%2Fbq.tar.gz
+tar xzf dssm%2Fbq.tar.gz
+rm -f dssm%2Fbq.tar.gz
+mv bq/train.txt ./raw_data.txt
 ```
-3. 本文提供了快速将数据集中的汉字数据处理为可训练格式数据的脚本，您在解压数据集后，可以看见目录中存在一个名为zhidao的文件。然后能可以在python3环境下运行我们提供的preprocess.py文件。即可生成可以直接用于训练的数据目录test.txt,train.txt,label.txt和testquery.txt。将其放入train和test目录下以备训练时调用。命令如下：
+3. 本文提供了快速将数据集中的汉字数据处理为可训练格式数据的脚本，您在解压数据集后，可以看见目录中存在一个名为bq的目录。将其中的train.txt文件移动到data目录下。然后可以在python3环境下运行我们提供的preprocess.py文件。即可生成可以直接用于训练的数据目录test.txt,train.txt,label.txt和testquery.txt。将其放入train和test目录下以备训练时调用。生成时间较长，请耐心等待。命令如下：
 ```
 python3 preprocess.py
-rm -f ./train/train.txt
-mv train.txt ./train
-rm -f ./test/test.txt
-mv test.txt ./test
+mkdir big_train
+mv train.txt ./big_train
+mkdir big_test
+mv test.txt ./big_test
 cd ..
 ```
-4. 退回tagspace目录中，打开文件config.yaml,更改其中的参数  
+也可以使用我们提供的一键数据处理脚本data_process.sh
+```
+sh data_process.sh
+```
+4. 退回multiview-simnet目录中，打开文件config.yaml,更改其中的参数  
 
-    将workspace改为您当前的绝对路径。（可用pwd命令获取绝对路径）  
+    将workspace改为您当前的绝对路径。（可用pwd命令获取绝对路径） 
+	将dataset_train中的data_path改为{workspace}/data/big_train
+    将dataset_infer中的data_path改为{workspace}/data/big_test 
 
 5.  执行脚本，开始训练.脚本会运行python -m paddlerec.run -m ./config.yaml启动训练，并将结果输出到result文件中。然后启动格式整理程序transform，最后计算正逆序比：
 ```
@@ -102,26 +114,14 @@ sh run.sh
 运行结果大致如下：
 ```
 ................run.................
-!!! The CPU_NUM is not specified, you should set CPU_NUM in the environment variable list.
-CPU_NUM indicates that how many CPUPlace are used in the current task.
-And if this parameter are set as N (equal to the number of physical CPU core) the program may be faster.
-
-export CPU_NUM=32 # for example, set CPU_NUM as number of physical CPU core which is 32.
-
-!!! The default number of CPU_NUM=1.
-I0821 14:24:57.255358  7888 parallel_executor.cc:440] The Program will be executed on CPU using ParallelExecutor, 1 cards are used, so 1 programs are executed in parallel.
-I0821 14:24:57.259166  7888 build_strategy.cc:365] SeqOnlyAllReduceOps:0, num_trainers:1
-I0821 14:24:57.262634  7888 parallel_executor.cc:307] Inplace strategy is enabled, when build_strategy.enable_inplace = True
-I0821 14:24:57.264791  7888 parallel_executor.cc:375] Garbage collection strategy is enabled, when FLAGS_eager_delete_tensor_gb = 0
-103
-pnr: 1.17674418605
-query_num: 11
-pair_num: 468 468
+8902
+pnr: 13.6785350966
+query_num: 1371
+pair_num: 14429 14429
 equal_num: 0
-正序率： 0.540598290598
-253 215
+正序率： 0.931873310694
+13446 983
 ```
-6. 提醒：因为采取较小的数据集进行训练和测试，得到指标的浮动程度会比较大。如果得到的指标不合预期，可以多次执行步骤5，即可获得合理的指标。
 ## 进阶使用
   
 ## FAQ

models/match/multiview-simnet/run.sh

@@ -14,7 +14,7 @@
 
 #!/bin/bash
 echo "................run................."
-python -m paddlerec.run -m ./config.yaml >result1.txt
+python -m paddlerec.run -m ./config.yaml &>result1.txt
 grep -i "query_pt_sim" ./result1.txt >./result2.txt
 sed '$d' result2.txt >result.txt
 rm -f result1.txt

models/match/multiview-simnet/transform.py

@@ -31,8 +31,9 @@ filename = './result.txt'
 sim = []
 for line in open(filename):
     line = line.strip().split(",")
-    line[1] = line[1].split(":")
-    line = line[1][1].strip(" ")
+    print(line)
+    line[3] = line[3].split(":")
+    line = line[3][1].strip(" ")
     line = line.strip("[")
     line = line.strip("]")
     sim.append(float(line))
@@ -49,5 +50,6 @@ f.close()
 filename = 'pair.txt'
 f = open(filename, "w")
 for i in range(len(sim)):
+    print(i)
     f.write(str(query[i]) + "\t" + str(sim[i]) + "\t" + str(label[i]) + "\n")
 f.close()

models/treebased/tdm/README.md

@@ -13,6 +13,7 @@ cd paddle-rec
 
 python -m paddlerec.run -m models/treebased/tdm/config.yaml
 ```
+3. 建树及自定义训练的细节可以查阅[TDM-Demo建树及训练](./gen_tree/README.md)
 
 ## 树结构的准备
 ### 名词概念

models/treebased/tdm/build_tree.md

-
-
-wget https://paddlerec.bj.bcebos.com/utils/tree_build_utils.tar.gz --no-check-certificate
-
-# input_path: embedding的路径
-# emb_shape: embedding中key-value，value的维度
-# emb格式要求： embedding_id(int64),embedding(float),embedding(float),......,embedding(float)
-# cluster_threads: 建树聚类所用线程
-python_172_anytree/bin/python -u main.py --input_path=./gen_emb/item_emb.txt --output_path=./ --emb_shape=24 --cluster_threads=4
-
-建树流程是：1、读取emb -> 2、kmeans聚类 -> 3、聚类结果整理为树 -> 4、基于树结构得到模型所需的4个文件
-    1    Layer_list：记录了每一层都有哪些节点。训练用
-    2    Travel_list：记录每个叶子节点的Travel路径。训练用
-    3    Tree_Info：记录了每个节点的信息，主要为：是否是item/item_id，所在层级，父节点，子节点。检索用
-    4    Tree_Embedding：记录所有节点的Embedding。训练及检索用
-
-注意一下训练数据输入的item是建树之前用的item id，还是基于树的node id，还是基于叶子的leaf id，在tdm_reader.py中，可以加载字典，做映射。
-用厂内版建树得到的输出文件夹里，有名为id2nodeid.txt的映射文件，格式是『hash值』+ 『树节点ID』+『叶子节点ID（表示第几个叶子节点，tdm_sampler op 所需的输入）』
-在另一个id2bidword.txt中，也有映射关系，格式是『hash值』+『原始item ID』，这个文件中仅存储了叶子节点的信息。

models/treebased/tdm/config.yaml

@@ -59,49 +59,39 @@ hyper_parameters:
     tree_emb_path: "{workspace}/tree/tree_emb.npy"
 
 # select runner by name
-mode: runner1
-# config of each runner.
-# runner is a kind of paddle training class, which wraps the train/infer process.
+mode: [runner1]
+
 runner:
 - name: runner1
   class: train
   startup_class_path: "{workspace}/tdm_startup.py"
-  # num of epochs
   epochs: 10
-  # device to run training or infer
   device: cpu
   save_checkpoint_interval: 2 # save model interval of epochs
-  save_inference_interval: 4 # save inference
   save_checkpoint_path: "increment" # save checkpoint path
-  save_inference_path: "inference" # save inference path
-  save_inference_feed_varnames: [] # feed vars of save inference
-  save_inference_fetch_varnames: [] # fetch vars of save inference
   init_model_path: "" # load model path
   print_interval: 10
+  phases: [phase1]
 
 - name: runner2
   class: infer
   startup_class_path: "{workspace}/tdm_startup.py"
-  # device to run training or infer
   device: cpu
   init_model_path: "increment/0" # load model path
   print_interval: 1
+  phases: [phase2]
   
 - name: runner3
   class: local_cluster_train
   startup_class_path: "{workspace}/tdm_startup.py"
   fleet_mode: ps
   epochs: 10
-  # device to run training or infer
   device: cpu
   save_checkpoint_interval: 2 # save model interval of epochs
-  save_inference_interval: 4 # save inference
   save_checkpoint_path: "increment" # save checkpoint path
-  save_inference_path: "inference" # save inference path
-  save_inference_feed_varnames: [] # feed vars of save inference
-  save_inference_fetch_varnames: [] # fetch vars of save inference
   init_model_path: "init_model" # load model path
   print_interval: 10
+  phases: [phase1]
 
 # runner will run all the phase in each epoch
 phase:
@@ -109,7 +99,7 @@ phase:
   model: "{workspace}/model.py" # user-defined model
   dataset_name: dataset_train # select dataset by name
   thread_num: 1
-# - name: phase2
-#   model: "{workspace}/model.py"
-#   dataset_name: dataset_infer
-#   thread_num: 2
+- name: phase2
+  model: "{workspace}/model.py"
+  dataset_name: dataset_infer
+  thread_num: 2

models/treebased/tdm/gen_tree/README.md

+# TDM-Demo建树及训练
+
+## 建树所需环境
+Requirements:
+- python >= 2.7
+- paddlepaddle >= 1.7.2（建议1.7.2）
+- paddle-rec （克隆github paddlerec，执行python setup.py install）
+- sklearn
+- anytree
+
+
+## 建树流程
+
+### 生成建树所需Embedding
+
+- 生成Fake的emb
+
+```shell
+cd gen_tree
+python -u emb_util.py
+```
+
+生成的emb维度是[13, 64]，含义是共有13个item，每个item的embedding维度是64，生成的item_emb位于`gen_tree/item_emb.txt`
+
+格式为`emb_value_0(float) 空格 emb_value_1(float) ... emb_value_63(float) \t item_id `
+
+在demo中，要求item的编号从0开始，范围 [0, item_nums-1]
+
+真实场景可以通过各种hash映射满足该要求
+
+### 对Item_embedding进行聚类建树
+
+执行
+
+```shell
+cd gen_tree
+# emd_path: item_emb的地址
+# emb_size: item_emb的第二个维度，即每个item的emb的size（示例中为64）
+# threads: 多线程建树配置的线程数
+# n_clusters: 最终建树为几叉树，此处设置为2叉树
+python gen_tree.py --emd_path item_emb.txt --emb_size 64 --output_dir ./output --threads 1 --n_clusters 2
+```
+
+生成的训练所需树结构文件位于`gen_tree/output`
+```shell
+.
+├── id2item.json         # 树节点id到item id的映射表
+├── layer_list.txt       # 树的每个层级都有哪些节点
+├── travel_list.npy      # 每个item从根到叶子的遍历路径，按item顺序排序
+├── travel_list.txt      # 上个文件的明文txt
+├── tree_embedding.txt   # 所有节点按节点id排列组成的embedding
+├── tree_emb.npy         # 上个文件的.npy版本
+├── tree_info.npy        # 每个节点：是否对应item/父/层级/子节点，按节点顺序排列
+├── tree_info.txt        # 上个文件的明文txt
+└── tree.pkl             # 聚类得到的树结构
+```
+
+我们最终需要使用建树生成的以下四个文件，参与网络训练，参考`models/treebased/tdm/config.yaml`
+
+1. layer_list.txt
+2. travel_list.npy 
+3. tree_info.npy
+4. tree_emb.npy
+
+
+### 执行训练
+
+- 更改`config.yaml`中的配置
+
+首先更改
+```yaml
+hyper_parameters:
+    # ...
+    tree:
+        # 单机训练建议tree只load一次，保存为paddle tensor，之后从paddle模型热启
+        # 分布式训练trainer需要独立load 
+        # 预测时也改为从paddle模型加载
+        load_tree_from_numpy: True # only once
+        load_paddle_model: False # train & infer need, after load from npy, change it to True
+        tree_layer_path: "{workspace}/tree/layer_list.txt"
+        tree_travel_path: "{workspace}/tree/travel_list.npy"
+        tree_info_path: "{workspace}/tree/tree_info.npy"
+        tree_emb_path: "{workspace}/tree/tree_emb.npy"
+```
+将上述几个path改为建树得到的文件所在的地址
+
+再更改
+```yaml
+hyper_parameters:
+  max_layers: 4                          # 不含根节点，树的层数
+  node_nums: 26                          # 树共有多少个节点，数量与tree_info文件的行数相等
+  leaf_node_nums: 13                     # 树共有多少个叶子节点
+  layer_node_num_list: [2, 4, 8, 10]     # 树的每层有多少个节点
+  child_nums: 2                          # 每个节点最多有几个孩子结点（几叉树）
+  neg_sampling_list: [1, 2, 3, 4]        # 在树的每层做多少负采样，训练自定义的参数
+```
+
+若并不知道对上面几个参数具体值，可以试运行一下，paddlerec读取建树生成的文件后，会将具体信息打印到屏幕上，如下所示：
+```shell
+...
+File_list: ['models/treebased/tdm/data/train/demo_fake_input.txt']
+2020-09-10 15:17:19,259 - INFO - Run TDM Trainer Startup Pass
+2020-09-10 15:17:19,283 - INFO - load tree from numpy
+2020-09-10 15:17:19,284 - INFO - TDM Tree leaf node nums: 13
+2020-09-10 15:17:19,284 - INFO - TDM Tree max layer: 4
+2020-09-10 15:17:19,284 - INFO - TDM Tree layer_node_num_list: [2, 4, 8, 10]
+2020-09-10 15:17:19,285 - INFO - Begin Save Init model.
+2020-09-10 15:17:19,394 - INFO - End Save Init model.
+Running SingleRunner.
+...
+```
+将其抄到配置中即可
+
+- 训练
+
+执行
+```
+cd /PaddleRec # PaddleRec 克隆的根目录
+python -m paddlerec.run -m models/treebased/tdm/config.yaml
+```

models/treebased/tdm/gen_tree/__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.
+from . import cluster
+
+__all__ = []
+__all__ += cluster.__all__

models/treebased/tdm/gen_tree/cluster.py

+# Copyright (C) 2016-2018 Alibaba Group Holding Limited
+#
+# 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 print_function
+import codecs
+import os
+import time
+import collections
+import argparse
+
+import multiprocessing as mp
+import numpy as np
+
+from sklearn.cluster import KMeans
+import tree_builder
+
+__all__ = ['Cluster']
+
+
+class Cluster:
+    def __init__(self,
+                 filename,
+                 emb_size,
+                 id_offset=None,
+                 parall=16,
+                 prev_result=None,
+                 output_dir='./',
+                 _n_clusters=2):
+        self.filename = filename
+        self.emb_size = emb_size
+        self.mini_batch = 256
+        self.ids = None
+        self.data = None
+        self.items = None
+        self.parall = parall
+        self.queue = None
+        self.timeout = 5
+        self.id_offset = id_offset
+        self.codes = None
+        self.prev_result = prev_result
+        self.output_dir = output_dir
+        self.n_clusters = _n_clusters
+
+    def _read(self):
+        t1 = time.time()
+        ids = list()
+        data = list()
+        items = list()
+        count = 0
+        with codecs.open(self.filename, 'r', encoding='utf-8') as f:
+            for line in f:
+                arr = line.rstrip().split('\t')
+                if not arr:
+                    break
+                elif len(arr) == 1:
+                    label = arr[0]
+                    emb_vec = (np.random.random_sample(
+                        (self.emb_size, ))).tolist()
+                elif len(arr) == 2:
+                    label = arr[1]
+                    emb_vec = arr[0].split()
+                if len(emb_vec) != self.emb_size:
+                    continue
+                if label in items:
+                    index = items.index(label)
+                    for i in range(0, len(emb_vec)):
+                        data[index][i + 1] += float(emb_vec[i])
+                    data[index][0] += 1
+                else:
+                    items.append(label)
+                    ids.append(count)
+                    count += 1
+                    vector = list()
+                    vector.append(1)
+                    for i in range(0, len(emb_vec)):
+                        vector.append(float(emb_vec[i]))
+                    data.append(vector)
+        for i in range(len(data)):
+            data_len = len(data[0])
+            for j in range(1, data_len):
+                data[i][j] /= data[i][0]
+            data[i] = data[i][1:]
+        self.ids = np.array(ids)
+        self.data = np.array(data)
+        self.items = np.array(items)
+        t2 = time.time()
+
+        print("Read data done, {} records read, elapsed: {}".format(
+            len(ids), t2 - t1))
+
+    def train(self):
+        ''' Cluster data '''
+        self._read()
+        queue = mp.Queue()
+        self.process_prev_result(queue)
+        processes = []
+        pipes = []
+        for _ in range(self.parall):
+            a, b = mp.Pipe()
+            p = mp.Process(target=self._train, args=(b, queue))
+            processes.append(p)
+            pipes.append(a)
+            p.start()
+
+        self.codes = np.zeros((len(self.ids), ), dtype=np.int64)
+        for pipe in pipes:
+            codes = pipe.recv()
+            for i in range(len(codes)):
+                if codes[i] > 0:
+                    self.codes[i] = codes[i]
+
+        for p in processes:
+            p.join()
+
+        assert (queue.empty())
+        builder = tree_builder.TreeBuilder(self.output_dir, self.n_clusters)
+        builder.build(self.ids, self.codes, items=self.items, data=self.data)
+
+    def process_prev_result(self, queue):
+        if not self.prev_result:
+            queue.put((0, np.array(range(len(self.ids)))))
+            return True
+
+        di = dict()
+        for i, node_id in enumerate(self.ids):
+            di[node_id] = i
+
+        indexes = []
+        clusters = []
+        with open(self.prev_result) as f:
+            for line in f:
+                arr = line.split(",")
+                if arr < 2:
+                    break
+                ni = [di[int(m)] for m in arr]
+                clusters.append(ni)
+                indexes += ni
+        assert len(set(indexes)) == len(self.ids), \
+            "ids count: {}, index count: {}".format(len(self.ids),
+                                                    len(set(indexes)))
+        count = len(clusters)
+        assert (count & (count - 1)) == 0, \
+            "Prev cluster count: {}".format(count)
+        for i, ni in enumerate(clusters):
+            queue.put((i + count - 1, np.array(ni)))
+        return True
+
+    def _train(self, pipe, queue):
+        last_size = -1
+        catch_time = 0
+        processed = False
+        code = np.zeros((len(self.ids), ), dtype=np.int64)
+        while True:
+            for _ in range(3):
+                try:
+                    pcode, index = queue.get(timeout=self.timeout)
+                except:
+                    index = None
+                if index is not None:
+                    break
+
+            if index is None:
+                if processed and (last_size <= self.mini_batch or
+                                  catch_time >= 3):
+                    print("Process {} exits".format(os.getpid()))
+                    break
+                else:
+                    print("Got empty job, pid: {}, time: {}".format(os.getpid(
+                    ), catch_time))
+                    catch_time += 1
+                    continue
+
+            processed = True
+            catch_time = 0
+            last_size = len(index)
+            if last_size <= self.mini_batch:
+                self._minbatch(pcode, index, code)
+            else:
+                start = time.time()
+                sub_index = self._cluster(index)
+                if last_size > self.mini_batch:
+                    print("Train iteration done, pcode:{}, "
+                          "data size: {}, elapsed time: {}"
+                          .format(pcode, len(index), time.time() - start))
+                self.timeout = int(0.4 * self.timeout + 0.6 * (time.time() -
+                                                               start))
+                if self.timeout < 5:
+                    self.timeout = 5
+
+                for i in range(self.n_clusters):
+                    if len(sub_index[i]) > 1:
+                        queue.put(
+                            (self.n_clusters * pcode + i + 1, sub_index[i]))
+
+        process_count = 0
+        for c in code:
+            if c > 0:
+                process_count += 1
+        print("Process {} process {} items".format(os.getpid(), process_count))
+        pipe.send(code)
+
+    def _minbatch(self, pcode, index, code):
+        dq = collections.deque()
+        dq.append((pcode, index))
+        batch_size = len(index)
+        tstart = time.time()
+        while dq:
+            pcode, index = dq.popleft()
+            if len(index) <= self.n_clusters:
+                for i in range(len(index)):
+                    code[index[i]] = self.n_clusters * pcode + i + 1
+                continue
+
+            sub_index = self._cluster(index)
+            for i in range(self.n_clusters):
+                if len(sub_index[i]) > 1:
+                    dq.append((self.n_clusters * pcode + i + 1, sub_index[i]))
+                elif len(sub_index[i]) > 0:
+                    for j in range(len(sub_index[i])):
+                        code[sub_index[i][j]] = self.n_clusters * \
+                            pcode + i + j + 1
+        print("Minbatch, batch size: {}, elapsed: {}".format(
+            batch_size, time.time() - tstart))
+
+    def _cluster(self, index):
+        data = self.data[index]
+        kmeans = KMeans(n_clusters=self.n_clusters, random_state=0).fit(data)
+        labels = kmeans.labels_
+        sub_indexes = []
+        remain_index = []
+        ave_num = len(index) / self.n_clusters
+
+        for i in range(self.n_clusters):
+            sub_i = np.where(labels == i)[0]
+            sub_index = index[sub_i]
+            if len(sub_index) <= ave_num:
+                sub_indexes.append(sub_index)
+            else:
+                distances = kmeans.transform(data[sub_i])[:, i]
+                sorted_index = sub_index[np.argsort(distances)]
+                sub_indexes.append(sorted_index[:ave_num])
+                remain_index.extend(list(sorted_index[ave_num:]))
+        idx = 0
+        while idx < self.n_clusters and len(remain_index) > 0:
+            if len(sub_indexes[idx]) >= ave_num:
+                idx += 1
+            else:
+                diff = min(len(remain_index), ave_num - len(sub_indexes[idx]))
+                sub_indexes[idx] = np.append(sub_indexes[idx],
+                                             np.array(remain_index[0:diff]))
+                remain_index = remain_index[diff:]
+                idx += 1
+        if len(remain_index) > 0:
+            sub_indexes[0] = np.append(sub_indexes[0], np.array(remain_index))
+
+        return sub_indexes
+
+    def _cluster1(self, index):
+        pass
+
+    def _rebalance(self, lindex, rindex, distances):
+        sorted_index = rindex[np.argsort(distances)]
+        idx = np.concatenate((lindex, sorted_index))
+        mid = int(len(idx) / 2)
+        return idx[mid:], idx[:mid]
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Tree cluster")
+    parser.add_argument(
+        "--embed_file",
+        required=True,
+        help="filename of the embedded vector file")
+    parser.add_argument(
+        "--emb_size",
+        type=int,
+        default=64,
+        help="dimension of input embedded vector")
+    parser.add_argument(
+        "--id_offset",
+        default=None,
+        help="id offset of the generated tree internal node")
+    parser.add_argument(
+        "--parall",
+        type=int,
+        default=16,
+        help="Parall execution process number")
+    parser.add_argument(
+        "--prev_result",
+        default=None,
+        help="filename of the previous cluster reuslt")
+
+    argments = parser.parse_args()
+    t1 = time.time()
+    cluster = Cluster(argments.embed_file, argments.emb_size,
+                      argments.id_offset, argments.parall,
+                      argments.prev_result)
+    cluster.train()
+    t2 = time.time()
+    print("Train complete successfully, elapsed: {}".format(t2 - t1))

models/treebased/tdm/gen_tree/emb_util.py

+# -*- coding=utf8 -*-
+# 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 os
+import paddle
+import paddle.fluid as fluid
+import numpy as np
+import json
+import argparse
+
+parser = argparse.ArgumentParser()
+parser.add_argument(
+    "--mode",
+    default="create_fake_emb",
+    choices=["create_fake_emb", "save_item_emb"],
+    type=str,
+    help=".")
+parser.add_argument("--emb_id_nums", default=13, type=int, help=".")
+parser.add_argument("--emb_shape", default=64, type=int, help=".")
+parser.add_argument("--emb_path", default='./item_emb.txt', type=str, help='.')
+args = parser.parse_args()
+
+
+def create_fake_emb(emb_id_nums, emb_shape, emb_path):
+    x = fluid.data(name="item", shape=[1], lod_level=1, dtype="int64")
+
+    # use layers.embedding to init emb value
+    item_emb = fluid.layers.embedding(
+        input=x,
+        is_sparse=True,
+        size=[emb_id_nums, emb_shape],
+        param_attr=fluid.ParamAttr(
+            name="Item_Emb",
+            initializer=fluid.initializer.TruncatedNormal(
+                loc=0.0, scale=2.0)))
+
+    # run startup to init emb tensor
+    exe = fluid.Executor(fluid.CPUPlace())
+    exe.run(fluid.default_startup_program())
+
+    # get np.array(emb_tensor)
+    print("Get Emb")
+    item_emb_array = np.array(fluid.global_scope().find_var("Item_Emb")
+                              .get_tensor())
+    with open(emb_path, 'w+') as f:
+        emb_str = ""
+        for index, value in enumerate(item_emb_array):
+            line = []
+            for v in value:
+                line.append(str(v))
+            line_str = " ".join(line)
+            line_str += "\t"
+            line_str += str(index)
+            line_str += "\n"
+            emb_str += line_str
+        f.write(emb_str)
+    print("Item Emb write Finish")
+
+
+if __name__ == "__main__":
+    create_fake_emb(args.emb_id_nums, args.emb_shape, args.emb_path)

models/treebased/tdm/gen_tree/gen_tree.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 os
+import argparse
+from cluster import Cluster
+import time
+import argparse
+from tree_search_util import tree_search_main
+parser = argparse.ArgumentParser()
+parser.add_argument("--emd_path", default='', type=str, help=".")
+parser.add_argument("--emb_size", default=64, type=int, help=".")
+parser.add_argument("--threads", default=1, type=int, help=".")
+parser.add_argument("--n_clusters", default=3, type=int, help=".")
+parser.add_argument("--output_dir", default='', type=str, help='.')
+args = parser.parse_args()
+
+
+def main():
+    cur_time = time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time()))
+    if not os.path.exists(args.output_dir):
+        os.system("mkdir -p " + args.output_dir)
+    print('%s start build tree' % cur_time)
+    # 1. Tree clustering, generating two files in current directory, tree.pkl, id2item.json
+    cluster = Cluster(
+        args.emd_path,
+        args.emb_size,
+        parall=args.threads,
+        output_dir=args.output_dir,
+        _n_clusters=args.n_clusters)
+    cluster.train()
+
+    # 2. Tree searching, generating tree_info, travel_list, layer_list for train process.
+    tree_search_main(
+        os.path.join(args.output_dir, "tree.pkl"),
+        os.path.join(args.output_dir, "id2item.json"), args.output_dir,
+        args.n_clusters)
+
+
+if __name__ == "__main__":
+    main()

models/treebased/tdm/gen_tree/tree_builder.py

+# Copyright (C) 2016-2018 Alibaba Group Holding Limited
+#
+# 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 print_function
+import numpy as np
+
+import sys
+import os
+import codecs
+from tree_impl import _build
+
+_CUR_DIR = os.path.dirname(os.path.realpath(__file__))
+sys.path.append(os.path.join(_CUR_DIR, ".."))
+
+
+class TreeBuilder:
+    def __init__(self, output_dir='./', n_clusters=2):
+        self.output_dir = output_dir
+        self.n_clusters = n_clusters
+
+    def build(
+            self,
+            ids,
+            codes,
+            data=None,
+            items=None,
+            id_offset=None, ):
+        _build(ids, codes, data, items, self.output_dir, self.n_clusters)
+
+    def _ancessors(self, code):
+        ancs = []
+        while code > 0:
+            code = int((code - 1) / 2)
+            ancs.append(code)
+        return ancs

models/treebased/tdm/gen_tree/tree_impl.py

+# Copyright (C) 2016-2018 Alibaba Group Holding Limited
+#
+# 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 anytree import NodeMixin, RenderTree
+import numpy as np
+from anytree.exporter.dictexporter import DictExporter
+import pickle
+import json
+import os
+import time
+
+
+class BaseClass(object):
+    pass
+
+
+class TDMTreeClass(BaseClass, NodeMixin):
+    def __init__(self,
+                 key_code,
+                 emb_vec,
+                 ids=None,
+                 text=None,
+                 parent=None,
+                 children=None):
+        super(TDMTreeClass, self).__init__()
+        self.key_code = key_code
+        self.ids = ids
+        self.emb_vec = emb_vec
+        self.text = text
+        self.parent = parent
+        if children:
+            self.children = children
+
+    def set_parent(self, parent):
+        self.parent = parent
+
+    def set_children(self, children):
+        self.children = children
+
+
+def _build(ids, codes, data, items, output_dir, n_clusters=2):
+    code_list = [0] * 50000000
+    node_dict = {}
+    max_code = 0
+    id2item = {}
+    curtime = time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time()))
+    print('%s start gen code_list' % curtime)
+    for _id, code, datum, item in zip(ids, codes, data, items):
+        code_list[code] = [datum, _id]
+        id2item[str(_id)] = item
+        max_code = max(code, max_code)
+        ancessors = _ancessors(code, n_clusters)
+        for ancessor in ancessors:
+            code_list[ancessor] = [[]]
+
+    for code in range(max_code, -1, -1):
+        if code_list[code] == 0:
+            continue
+        if len(code_list[code]) > 1:
+            pass
+        elif len(code_list[code]) == 1:
+            code_list[code][0] = np.mean(code_list[code][0], axis=0)
+        if code > 0:
+            ancessor = int((code - 1) / n_clusters)
+            code_list[ancessor][0].append(code_list[code][0])
+
+    print('start gen node_dict')
+    for code in range(0, max_code + 1):
+        if code_list[code] == 0:
+            continue
+        if len(code_list[code]) > 1:
+            [datum, _id] = code_list[code]
+            node_dict[code] = TDMTreeClass(code, emb_vec=datum, ids=_id)
+        elif len(code_list[code]) == 1:
+            [datum] = code_list[code]
+            node_dict[code] = TDMTreeClass(code, emb_vec=datum)
+        if code > 0:
+            ancessor = int((code - 1) / n_clusters)
+            node_dict[code].set_parent(node_dict[ancessor])
+
+    save_tree(node_dict[0], os.path.join(output_dir, 'tree.pkl'))
+    save_dict(id2item, os.path.join(output_dir, 'id2item.json'))
+
+
+def render(root):
+    for row in RenderTree(root, childiter=reversed):
+        print("%s%s" % (row.pre, row.node.text))
+
+
+def save_tree(root, path):
+    print('save tree to %s' % path)
+    exporter = DictExporter()
+    data = exporter.export(root)
+    f = open(path, 'wb')
+    pickle.dump(data, f)
+    f.close()
+
+
+def save_dict(dic, filename):
+    """save dict into json file"""
+    print('save dict to %s' % filename)
+    with open(filename, "w") as json_file:
+        json.dump(dic, json_file, ensure_ascii=False)
+
+
+def _ancessors(code, n_clusters):
+    ancs = []
+    while code > 0:
+        code = int((code - 1) / n_clusters)
+        ancs.append(code)
+    return ancs

models/treebased/tdm/gen_tree/tree_search_util.py

+# -*- coding=utf8 -*-
+# 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 json
+import pickle
+import time
+import os
+import numpy as np
+
+from anytree import (AsciiStyle, LevelOrderGroupIter, LevelOrderIter, Node,
+                     NodeMixin, RenderTree)
+from anytree.importer.dictimporter import DictImporter
+from anytree.iterators.abstractiter import AbstractIter
+from anytree.walker import Walker
+from tree_impl import TDMTreeClass
+
+
+class myLevelOrderIter(AbstractIter):
+    @staticmethod
+    def _iter(children, filter_, stop, maxlevel):
+        level = 1
+        while children:
+            next_children = []
+            for child in children:
+                if filter_(child):
+                    yield child, level
+                next_children += AbstractIter._get_children(child.children,
+                                                            stop)
+            children = next_children
+            level += 1
+            if AbstractIter._abort_at_level(level, maxlevel):
+                break
+
+
+class Tree_search(object):
+    def __init__(self, tree_path, id2item_path, child_num=2):
+        self.root = None
+        self.id2item = None
+        self.item2id = None
+        self.child_num = child_num
+
+        self.load(tree_path)
+        # self.load_id2item(id2item_path)
+
+        self.level_code = [[]]
+        self.max_level = 0
+        self.keycode_id_dict = {}
+        #  embedding
+        self.keycode_nodeid_dict = {}
+        self.tree_info = []
+        self.id_node_dict = {}
+
+        self.get_keycode_mapping()
+        self.travel_tree()
+        self.get_children()
+
+    def get_keycode_mapping(self):
+        nodeid = 0
+        self.embedding = []
+        print("Begin Keycode Mapping")
+        for node in myLevelOrderIter(self.root):
+            node, level = node
+            if level - 1 > self.max_level:
+                self.max_level = level - 1
+                self.level_code.append([])
+            if node.ids is not None:
+                self.keycode_id_dict[node.key_code] = node.ids
+                self.id_node_dict[node.ids] = node
+            self.keycode_nodeid_dict[node.key_code] = nodeid
+            self.level_code[self.max_level].append(nodeid)
+
+            node_infos = []
+            if node.ids is not None:  # item_id
+                node_infos.append(node.ids)
+            else:
+                node_infos.append(0)
+            node_infos.append(self.max_level)  # layer_id
+            if node.parent:  # ancestor_id
+                node_infos.append(self.keycode_nodeid_dict[
+                    node.parent.key_code])
+            else:
+                node_infos.append(0)
+            self.tree_info.append(node_infos)
+            self.embedding.append(node.emb_vec)
+            nodeid += 1
+            if nodeid % 1000 == 0:
+                print("travel node id {}".format(nodeid))
+
+    def load(self, path):
+        print("Begin Load Tree")
+        f = open(path, "rb")
+        data = pickle.load(f)
+        pickle.dump(data, open(path, "wb"), protocol=2)
+        importer = DictImporter()
+        self.root = importer.import_(data)
+        f.close()
+
+    def load_id2item(self, path):
+        """load dict from json file"""
+        with open(path, "rb") as json_file:
+            self.id2item = json.load(json_file)
+
+        self.item2id = {value: int(key) for key, value in self.id2item.items()}
+
+    def get_children(self):
+        """get every node children info"""
+        print("Begin Keycode Mapping")
+        for node in myLevelOrderIter(self.root):
+            node, level = node
+            node_id = self.keycode_nodeid_dict[node.key_code]
+            child_idx = 0
+            if node.children:
+                for child in node.children:
+                    self.tree_info[node_id].append(self.keycode_nodeid_dict[
+                        child.key_code])
+                    child_idx += 1
+            while child_idx < self.child_num:
+                self.tree_info[node_id].append(0)
+                child_idx += 1
+            if node_id % 1000 == 0:
+                print("get children  node id {}".format(node_id))
+
+    def travel_tree(self):
+        self.travel_list = []
+        tree_walker = Walker()
+        print("Begin Travel Tree")
+        for item in sorted(self.id_node_dict.keys()):
+            node = self.id_node_dict[int(item)]
+            paths, _, _ = tree_walker.walk(node, self.root)
+            paths = list(paths)
+            paths.reverse()
+            travel = [self.keycode_nodeid_dict[i.key_code] for i in paths]
+            while len(travel) < self.max_level:
+                travel.append(0)
+            self.travel_list.append(travel)
+
+
+def tree_search_main(tree_path, id2item_path, output_dir, n_clusters=2):
+    print("Begin Tree Search")
+    t = Tree_search(tree_path, id2item_path, n_clusters)
+
+    # 1. Walk all leaf nodes, get travel path array
+    travel_list = np.array(t.travel_list)
+    np.save(os.path.join(output_dir, "travel_list.npy"), travel_list)
+    with open(os.path.join(output_dir, "travel_list.txt"), 'w') as fout:
+        for i, travel in enumerate(t.travel_list):
+            travel = map(str, travel)
+            fout.write(','.join(travel))
+            fout.write("\n")
+    print("End Save tree travel")
+
+    # 2. Walk all layer of tree, get layer array
+    layer_num = 0
+    with open(os.path.join(output_dir, "layer_list.txt"), 'w') as fout:
+        for layer in t.level_code:
+            # exclude layer 0
+            if layer_num == 0:
+                layer_num += 1
+                continue
+            for idx in range(len(layer) - 1):
+                fout.write(str(layer[idx]) + ',')
+            fout.write(str(layer[-1]) + "\n")
+            print("Layer {} has {} node, the first {}, the last {}".format(
+                layer_num, len(layer), layer[0], layer[-1]))
+            layer_num += 1
+    print("End Save tree layer")
+
+    # 3. Walk all node of tree, get tree info
+    tree_info = np.array(t.tree_info)
+    np.save(os.path.join(output_dir, "tree_info.npy"), tree_info)
+    with open(os.path.join(output_dir, "tree_info.txt"), 'w') as fout:
+        for i, node_infos in enumerate(t.tree_info):
+            node_infos = map(str, node_infos)
+            fout.write(','.join(node_infos))
+            fout.write("\n")
+    print("End Save tree info")
+
+    # 4. save embedding
+    embedding = np.array(t.embedding)
+    np.save(os.path.join(output_dir, "tree_emb.npy"), embedding)
+    with open(os.path.join(output_dir, "tree_embedding.txt"), "w") as fout:
+        for i, emb in enumerate(t.embedding):
+            emb = map(str, emb)
+            fout.write(','.join(emb))
+            fout.write("\n")
+
+
+if __name__ == "__main__":
+    tree_path = "./tree.pkl"
+    id2item_path = "./id2item.json"
+    output_dir = "./output"
+    if not os.path.exists(output_dir):
+        os.system("mkdir -p " + output_dir)
+    tree_search_main(tree_path, id2item_path, output_dir)

tools/cal_pos_neg.py

+#encoding=utf-8     
 # Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -11,8 +12,6 @@
 # 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.
-#!/usr/bin/python   
-#-*- coding:utf-8 -*-   
 """
 docstring
 """
@@ -21,10 +20,10 @@ import os
 import sys
 
 if len(sys.argv) < 2:
-    print "usage:python %s input" % (sys.argv[0])
+    print("usage:python {} input".format(sys.argv[0]))
     sys.exit(-1)
 
-fin = file(sys.argv[1])
+fin = open(sys.argv[1])
 pos_num = 0
 neg_num = 0
 
@@ -42,15 +41,15 @@ for line in fin:
     cols = line.strip().split("\t")
     cnt += 1
     if cnt % 500000 == 0:
-        print "cnt:", cnt, 1.0 * pos_num / neg_num
+        print("cnt:{}".format(1.0 * pos_num / neg_num))
     if len(cols) != 3:
         continue
 
     cur_query = cols[0]
     if cur_query != last_query:
         query_num += 1
-        for i in xrange(0, len(score_list)):
-            for j in xrange(i + 1, len(score_list)):
+        for i in range(0, len(score_list)):
+            for j in range(i + 1, len(score_list)):
                 if label_list[i] == label_list[j]:
                     continue
                 pair_num += 1
@@ -74,8 +73,8 @@ for line in fin:
 
 fin.close()
 
-for i in xrange(0, len(score_list)):
-    for j in xrange(i + 1, len(score_list)):
+for i in range(0, len(score_list)):
+    for j in range(i + 1, len(score_list)):
         if label_list[i] == label_list[j]:
             continue
         pair_num += 1
@@ -89,9 +88,9 @@ for i in xrange(0, len(score_list)):
             equal_num += 1
 
 if neg_num > 0:
-    print "pnr:", 1.0 * pos_num / neg_num
-    print "query_num:", query_num
-    print "pair_num:", pos_num + neg_num + equal_num, pair_num
-    print "equal_num:", equal_num
-    print "正序率：", 1.0 * pos_num / (pos_num + neg_num)
-print pos_num, neg_num
+    print("pnr:{}".format(1.0 * pos_num / neg_num))
+    print("query_num:{}".format(query_num))
+    print("pair_num:{} , {}".format(pos_num + neg_num + equal_num, pair_num))
+    print("equal_num:{}".format(equal_num))
+    print("正序率: {}".format(1.0 * pos_num / (pos_num + neg_num)))
+print("pos_num: {} , neg_num: {}".format(pos_num, neg_num))