add gen_tree (#214)

Co-authored-by: Nwuzhihua <35824027+fuyinno4@users.noreply.github.com>

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)