add flen

models/rank/flen/README.md

+# FLEN
+
+ 以下是本例的简要目录结构及说明： 
+
+```
+├── data #样例数据
+	├── sample_data
+		├── train
+			├── sample_train.txt
+	├── run.sh
+	├── get_slot_data.py
+├── __init__.py
+├── README.md # 文档
+├── model.py #模型文件
+├── config.yaml #配置文件
+```
+
+## 简介
+
+[《FLEN: Leveraging Field for Scalable CTR Prediction》](https://arxiv.org/pdf/1911.04690.pdf)文章提出了field-wise bi-interaction pooling技术，解决了在大规模应用特征field信息时存在的时间复杂度和空间复杂度高的困境，同时提出了一种缓解梯度耦合问题的方法dicefactor。该模型已应用于美图的大规模推荐系统中，持续稳定地取得业务效果的全面提升。
+
+本项目在avazu数据集上验证模型效果
+
+## 数据下载及预处理
+
+## 环境
+
+PaddlePaddle 1.7.2
+
+python3.7 
+
+PaddleRec
+
+## 单机训练
+
+CPU环境
+
+在config.yaml文件中设置好设备，epochs等。
+
+```
+# select runner by name
+mode: [single_cpu_train, single_cpu_infer]
+# config of each runner.
+# runner is a kind of paddle training class, which wraps the train/infer process.
+runner:
+- name: single_cpu_train
+  class: train
+  # num of epochs
+  epochs: 4
+  # 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_model" # 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]
+```
+
+## 单机预测
+
+CPU环境
+
+在config.yaml文件中设置好epochs、device等参数。
+
+```
+- name: single_cpu_infer
+  class: infer
+  # num of epochs
+  epochs: 1
+  # device to run training or infer
+  device: cpu #选择预测的设备
+  init_model_path: "increment_dnn" # load model path
+  phases: [phase2]
+```
+
+## 运行
+
+```
+python -m paddlerec.run -m paddlerec.models.rank.flen
+```
+
+## 模型效果
+
+在样例数据上测试模型
+
+训练：
+
+```
+0702 13:38:20.903220  7368 parallel_executor.cc:440] The Program will be executed on CPU using ParallelExecutor, 2 cards are used, so 2 programs are executed in parallel.
+I0702 13:38:20.925912  7368 parallel_executor.cc:307] Inplace strategy is enabled, when build_strategy.enable_inplace = True
+I0702 13:38:20.933356  7368 parallel_executor.cc:375] Garbage collection strategy is enabled, when FLAGS_eager_delete_tensor_gb = 0
+batch: 2, AUC: [0.09090909 0.        ], BATCH_AUC: [0.09090909 0.        ]
+batch: 4, AUC: [0.31578947 0.29411765], BATCH_AUC: [0.31578947 0.29411765]
+batch: 6, AUC: [0.41333333 0.33333333], BATCH_AUC: [0.41333333 0.33333333]
+batch: 8, AUC: [0.4453125  0.44166667], BATCH_AUC: [0.4453125  0.44166667]
+batch: 10, AUC: [0.39473684 0.38888889], BATCH_AUC: [0.44117647 0.41176471]
+batch: 12, AUC: [0.41860465 0.45535714], BATCH_AUC: [0.5078125  0.54545455]
+batch: 14, AUC: [0.43413729 0.42746615], BATCH_AUC: [0.56666667 0.56      ]
+batch: 16, AUC: [0.46433566 0.47460087], BATCH_AUC: [0.53       0.59247649]
+batch: 18, AUC: [0.44009217 0.44642857], BATCH_AUC: [0.46 0.47]
+batch: 20, AUC: [0.42705314 0.43781095], BATCH_AUC: [0.45878136 0.4874552 ]
+batch: 22, AUC: [0.45176471 0.46011281], BATCH_AUC: [0.48046875 0.45878136]
+batch: 24, AUC: [0.48375    0.48910256], BATCH_AUC: [0.56630824 0.59856631]
+epoch 0 done, use time: 0.21532440185546875
+PaddleRec Finish
+```
+
+预测
+
+```
+PaddleRec: Runner single_cpu_infer Begin
+Executor Mode: infer
+processor_register begin
+Running SingleInstance.
+Running SingleNetwork.
+QueueDataset can not support PY3, change to DataLoader
+QueueDataset can not support PY3, change to DataLoader
+Running SingleInferStartup.
+Running SingleInferRunner.
+load persistables from increment_model/0
+batch: 20, AUC: [0.49121353], BATCH_AUC: [0.66176471]
+batch: 40, AUC: [0.51156463], BATCH_AUC: [0.55197133]
+Infer phase2 of 0 done, use time: 0.3941819667816162
+PaddleRec Finish
+```
+

models/rank/flen/__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.

models/rank/flen/config.yaml

+# 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.
+
+# workspace
+workspace: "paddlerec.models.rank.flen"
+
+# list of dataset
+dataset:
+- name: dataloader_train # name of dataset to distinguish different datasets
+  batch_size: 2
+  type: DataLoader # or QueueDataset 
+  data_path: "{workspace}/data/sample_data/train"
+  sparse_slots: "click user_0 user_1 user_2 user_3 user_4 user_5 user_6 user_7 user_8 user_9 user_10 user_11 item_0 item_1 item_2 contex_0 contex_1 contex_2 contex_3 contex_4 contex_5"
+  dense_slots: ""
+- name: dataset_infer # name
+  batch_size: 2
+  type: DataLoader # or QueueDataset
+  data_path: "{workspace}/data/sample_data/train"
+  sparse_slots: "click user_0 user_1 user_2 user_3 user_4 user_5 user_6 user_7 user_8 user_9 user_10 user_11 item_0 item_1 item_2 contex_0 contex_1 contex_2 contex_3 contex_4 contex_5"
+  dense_slots: ""
+
+# hyper parameters of user-defined network
+hyper_parameters:
+  # optimizer config
+  optimizer:
+    class: Adam
+    learning_rate: 0.001
+    strategy: async
+  # user-defined <key, value> pairs
+  sparse_inputs_slots: 21
+  sparse_feature_number: 100
+  sparse_feature_dim: 8
+  dense_input_dim: 1
+  dropout_rate: 0.5
+
+# select runner by name
+mode: [single_cpu_train, single_cpu_infer]
+# config of each runner.
+# runner is a kind of paddle training class, which wraps the train/infer process.
+runner:
+- name: single_cpu_train
+  class: train
+  # num of epochs
+  epochs: 1
+  # device to run training or infer
+  device: cpu
+  save_checkpoint_interval: 1 # save model interval of epochs
+  save_inference_interval: 4 # save inference
+  save_checkpoint_path: "increment_model" # 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: 2
+  phases: [phase1]
+
+- name: single_gpu_train
+  class: train
+  # num of epochs
+  epochs: 1
+  # device to run training or infer
+  device: gpu
+  save_checkpoint_interval: 1 # save model interval of epochs
+  save_inference_interval: 4 # save inference
+  save_checkpoint_path: "increment_model" # 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: 2
+  phases: [phase1]
+
+- name: single_cpu_infer
+  class: infer
+  # device to run training or infer
+  device: cpu
+  init_model_path: "increment_model" # load model path
+  phases: [phase2]
+
+- name: single_gpu_infer
+  class: infer
+  # device to run training or infer
+  device: gpu
+  init_model_path: "increment_model" # load model path
+  phases: [phase2]
+
+# runner will run all the phase in each epoch
+phase:
+- name: phase1
+  model: "{workspace}/model.py" # user-defined model
+  dataset_name: dataloader_train # select dataset by name
+  thread_num: 2
+
+- name: phase2
+  model: "{workspace}/model.py" # user-defined model
+  dataset_name: dataset_infer # select dataset by name
+  thread_num: 2
+  

models/rank/flen/data/get_slot_data.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle.fluid.incubate.data_generator as dg
+
+
+class CriteoDataset(dg.MultiSlotDataGenerator):
+    """
+    DacDataset: inheritance MultiSlotDataGeneratior, Implement data reading
+    Help document: http://wiki.baidu.com/pages/viewpage.action?pageId=728820675
+    """
+
+    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.strip().split(',')
+
+            label = [int(features[0])]
+
+            s = "click:" + str(label[0])
+            for i, elem in enumerate(features[1:13]):
+                s += " user_" + str(i) + ":" + str(elem)
+            for i, elem in enumerate(features[13:16]):
+                s += " item_" + str(i) + ":" + str(elem)
+            for i, elem in enumerate(features[16:]):
+                s += " contex_" + str(i) + ":" + str(elem)
+            print(s.strip())
+            yield None
+
+        return reader
+
+
+d = CriteoDataset()
+d.run_from_stdin()

models/rank/flen/data/run.sh

+mkdir train
+
+for i in `ls ./train_data`
+do
+    cat train_data/$i | python get_slot_data.py > train/$i
+done

models/rank/flen/model.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 paddle.fluid as fluid
+import itertools
+from paddlerec.core.utils import envs
+from paddlerec.core.model import ModelBase
+
+
+class Model(ModelBase):
+    def __init__(self, config):
+        ModelBase.__init__(self, config)
+
+    def _init_hyper_parameters(self):
+        self.is_distributed = True if envs.get_fleet_mode().upper(
+        ) == "PSLIB" else False
+        self.sparse_feature_number = envs.get_global_env(
+            "hyper_parameters.sparse_feature_number")
+        self.sparse_feature_dim = envs.get_global_env(
+            "hyper_parameters.sparse_feature_dim")
+        self.learning_rate = envs.get_global_env(
+            "hyper_parameters.optimizer.learning_rate")
+
+    def _FieldWiseBiInteraction(self, inputs):
+        # MF module
+        field_wise_embeds_list = inputs
+
+        field_wise_vectors = [
+            fluid.layers.reduce_sum(
+                field_i_vectors, dim=1, keep_dim=True)
+            for field_i_vectors in field_wise_embeds_list
+        ]
+        num_fields = len(field_wise_vectors)
+
+        h_mf_list = []
+        for emb_left, emb_right in itertools.combinations(field_wise_vectors,
+                                                          2):
+            embeddings_prod = fluid.layers.elementwise_mul(emb_left, emb_right)
+
+            field_weighted_embedding = fluid.layers.fc(
+                input=embeddings_prod,
+                size=self.sparse_feature_dim,
+                param_attr=fluid.initializer.ConstantInitializer(value=1),
+                name='kernel_mf')
+            h_mf_list.append(field_weighted_embedding)
+        h_mf = fluid.layers.concat(h_mf_list, axis=1)
+        h_mf = fluid.layers.reshape(
+            x=h_mf, shape=[-1, num_fields, self.sparse_feature_dim])
+        h_mf = fluid.layers.reduce_sum(h_mf, dim=1)
+
+        square_of_sum_list = [
+            fluid.layers.square(
+                fluid.layers.reduce_sum(
+                    field_i_vectors, dim=1, keep_dim=True))
+            for field_i_vectors in field_wise_embeds_list
+        ]
+
+        sum_of_square_list = [
+            fluid.layers.reduce_sum(
+                fluid.layers.elementwise_mul(field_i_vectors, field_i_vectors),
+                dim=1,
+                keep_dim=True) for field_i_vectors in field_wise_embeds_list
+        ]
+
+        field_fm_list = []
+        for square_of_sum, sum_of_square in zip(square_of_sum_list,
+                                                sum_of_square_list):
+            field_fm = fluid.layers.reshape(
+                fluid.layers.elementwise_sub(square_of_sum, sum_of_square),
+                shape=[-1, self.sparse_feature_dim])
+            field_fm = fluid.layers.fc(
+                input=field_fm,
+                size=self.sparse_feature_dim,
+                param_attr=fluid.initializer.ConstantInitializer(value=0.5),
+                name='kernel_fm')
+            field_fm_list.append(field_fm)
+
+        h_fm = fluid.layers.concat(field_fm_list, axis=1)
+        h_fm = fluid.layers.reshape(
+            x=h_fm, shape=[-1, num_fields, self.sparse_feature_dim])
+        h_fm = fluid.layers.reduce_sum(h_fm, dim=1)
+
+        return fluid.layers.elementwise_add(h_mf, h_fm)
+
+    def _DNNLayer(self, inputs, dropout_rate=0.2):
+        deep_input = inputs
+        for i, hidden_unit in enumerate([64, 32]):
+            fc_out = fluid.layers.fc(
+                input=deep_input,
+                size=hidden_unit,
+                param_attr=fluid.initializer.Xavier(uniform=False),
+                act='relu',
+                name='d_' + str(i))
+            fc_out = fluid.layers.dropout(fc_out, dropout_prob=dropout_rate)
+            deep_input = fc_out
+
+        return deep_input
+
+    def embeddingLayer(self, inputs):
+        emb_list = []
+        in_len = len(inputs)
+        for data in inputs:
+            feat_emb = fluid.embedding(
+                input=data,
+                size=[self.sparse_feature_number, self.sparse_feature_dim],
+                param_attr=fluid.ParamAttr(
+                    name='item_emb',
+                    learning_rate=5,
+                    initializer=fluid.initializer.Xavier(
+                        fan_in=self.sparse_feature_dim,
+                        fan_out=self.sparse_feature_dim)),
+                is_sparse=True)
+            emb_list.append(feat_emb)
+        concat_emb = fluid.layers.concat(emb_list, axis=1)
+        field_emb = fluid.layers.reshape(
+            x=concat_emb, shape=[-1, in_len, self.sparse_feature_dim])
+
+        return field_emb
+
+    def net(self, input, is_infer=False):
+        self.user_inputs = self._sparse_data_var[1:13]
+        self.item_inputs = self._sparse_data_var[13:16]
+        self.contex_inputs = self._sparse_data_var[16:]
+        self.label_input = self._sparse_data_var[0]
+
+        dropout_rate = envs.get_global_env("hyper_parameters.dropout_rate")
+
+        field_wise_embeds_list = []
+        for inputs in [self.user_inputs, self.item_inputs, self.contex_inputs]:
+            field_emb = self.embeddingLayer(inputs)
+            field_wise_embeds_list.append(field_emb)
+
+        dnn_input = fluid.layers.concat(
+            [
+                fluid.layers.flatten(
+                    x=field_i_vectors, axis=1)
+                for field_i_vectors in field_wise_embeds_list
+            ],
+            axis=1)
+
+        #mlp part
+        dnn_output = self._DNNLayer(dnn_input, dropout_rate)
+
+        #field-weighted embedding
+        fm_mf_out = self._FieldWiseBiInteraction(field_wise_embeds_list)
+        logits = fluid.layers.concat([fm_mf_out, dnn_output], axis=1)
+
+        y_pred = fluid.layers.fc(
+            input=logits,
+            size=1,
+            param_attr=fluid.initializer.Xavier(uniform=False),
+            act='sigmoid',
+            name='logit')
+
+        self.predict = y_pred
+        auc, batch_auc, _ = fluid.layers.auc(input=self.predict,
+                                             label=self.label_input,
+                                             num_thresholds=2**12,
+                                             slide_steps=20)
+
+        if is_infer:
+            self._infer_results["AUC"] = auc
+            self._infer_results["BATCH_AUC"] = batch_auc
+            return
+
+        self._metrics["AUC"] = auc
+        self._metrics["BATCH_AUC"] = batch_auc
+        cost = fluid.layers.log_loss(
+            input=self.predict,
+            label=fluid.layers.cast(
+                x=self.label_input, dtype='float32'))
+        avg_cost = fluid.layers.reduce_mean(cost)
+        self._cost = avg_cost