Merge pull request #114 from overlordmax/pr_6192341

add fibinet

README.md

@@ -59,7 +59,8 @@
     |   排序   |                  [xDeepFM](models/rank/xdeepfm/model.py)                  |    ✓    |    x    |     ✓     |     x     | [KDD 2018][xDeepFM: Combining Explicit and Implicit Feature Interactions for Recommender Systems](https://dl.acm.org/doi/pdf/10.1145/3219819.3220023)                                                       |
     |   排序   |                      [DIN](models/rank/din/model.py)                      |    ✓    |    x    |     ✓     |     x     | [KDD 2018][Deep Interest Network for Click-Through Rate Prediction](https://dl.acm.org/doi/pdf/10.1145/3219819.3219823)                                                                                     |
     |   排序   |                [Wide&Deep](models/rank/wide_deep/model.py)                |    ✓    |    x    |     ✓     |     x     | [DLRS 2016][Wide & Deep Learning for Recommender Systems](https://dl.acm.org/doi/pdf/10.1145/2988450.2988454)                                                                                               |
-    |   排序   |                    [FGCNN](models/rank/fgcnn/model.py)                    |    ✓    |    ✓    |     ✓     |     ✓     | [WWW 2019][Feature Generation by Convolutional Neural Network for Click-Through Rate Prediction](https://arxiv.org/pdf/1904.04447.pdf)                                                                      |
+    |   排序   |                    [FGCNN](models/rank/fgcnn/model.py)                    |    ✓    |    ✓    |     ✓     |     ✓     | [WWW 2019][Feature Generation by Convolutional Neural Network for Click-Through Rate Prediction](https://arxiv.org/pdf/1904.04447.pdf)  
+    |   排序   |                    [Fibinet](models/rank/fibinet/model.py)                    |    ✓    |    ✓    |     ✓     |     ✓     | [RecSys19][FiBiNET: Combining Feature Importance and Bilinear feature Interaction for Click-Through Rate Prediction]( https://arxiv.org/pdf/1905.09433.pdf) |
     |  多任务  |                  [ESMM](models/multitask/esmm/model.py)                   |    ✓    |    ✓    |     ✓     |     ✓     | [SIGIR 2018][Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate](https://arxiv.org/abs/1804.07931)                                                              |
     |  多任务  |                  [MMOE](models/multitask/mmoe/model.py)                   |    ✓    |    ✓    |     ✓     |     ✓     | [KDD 2018][Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts](https://dl.acm.org/doi/abs/10.1145/3219819.3220007)                                                       |
     |  多任务  |           [ShareBottom](models/multitask/share-bottom/model.py)           |    ✓    |    ✓    |     ✓     |     ✓     | [1998][Multitask learning](http://reports-archive.adm.cs.cmu.edu/anon/1997/CMU-CS-97-203.pdf)                                                                                                               |

models/rank/fibinet/README.md

+# fibinet
+
+ 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
+	├── download.sh
+	├── 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
+
+[《FiBiNET: Combining Feature Importance and Bilinear feature Interaction for Click-Through Rate Prediction》]( https://arxiv.org/pdf/1905.09433.pdf)是新浪微博机器学习团队发表在RecSys19上的一篇论文，文章指出当前的许多通过特征组合进行CTR预估的工作主要使用特征向量的内积或哈达玛积来计算交叉特征，这种方法忽略了特征本身的重要程度。提出通过使用Squeeze-Excitation network (SENET) 结构动态学习特征的重要性以及使用一个双线性函数来更好的建模交叉特征。
+
+本项目在paddlepaddle上实现FibiNET的网络结构，并在开源数据集Criteo上验证模型效果。
+
+## 数据下载及预处理
+
+数据地址：[Criteo]( https://fleet.bj.bcebos.com/ctr_data.tar.gz)
+
+（1）将原始训练集按9:1划分为训练集和验证集
+
+（2）数值特征（连续特征）进行归一化处理
+
+执行run.sh生成训练集和测试集
+
+```
+sh run.sh
+```
+
+## 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.fibinet
+```
+
+## 模型效果
+
+在样例数据上测试模型
+
+训练：
+
+```
+Running SingleStartup.
+W0623 12:03:35.130075   509 device_context.cc:237] Please NOTE: device: 0, CUDA Capability: 70, Driver API Version: 9.2, Runtime API Version: 9.0
+W0623 12:03:35.134771   509 device_context.cc:245] device: 0, cuDNN Version: 7.3.
+Running SingleRunner.
+batch: 100, AUC: [0.6449976], BATCH_AUC: [0.69029814]
+batch: 200, AUC: [0.6769844], BATCH_AUC: [0.70255003]
+batch: 300, AUC: [0.67131597], BATCH_AUC: [0.68954499]
+batch: 400, AUC: [0.68129822], BATCH_AUC: [0.70892718]
+batch: 500, AUC: [0.68242937], BATCH_AUC: [0.69269376]
+batch: 600, AUC: [0.68741928], BATCH_AUC: [0.72034578]
+...
+batch: 1400, AUC: [0.84607023], BATCH_AUC: [0.93358024]
+batch: 1500, AUC: [0.84796116], BATCH_AUC: [0.95302841]
+batch: 1600, AUC: [0.84949111], BATCH_AUC: [0.92868531]
+batch: 1700, AUC: [0.85113661], BATCH_AUC: [0.95452616]
+batch: 1800, AUC: [0.85260467], BATCH_AUC: [0.92847032]
+epoch 3 done, use time: 1618.1106688976288
+```
+
+预测
+
+```
+load persistables from increment_model/3
+batch: 20, AUC: [0.85304064], BATCH_AUC: [0.94178556]
+batch: 40, AUC: [0.85304544], BATCH_AUC: [0.95207907]
+batch: 60, AUC: [0.85303907], BATCH_AUC: [0.94782551]
+batch: 80, AUC: [0.85298773], BATCH_AUC: [0.93987691]
+...
+batch: 1780, AUC: [0.866046], BATCH_AUC: [0.96424594]
+batch: 1800, AUC: [0.86633785], BATCH_AUC: [0.96900967]
+batch: 1820, AUC: [0.86662365], BATCH_AUC: [0.96759972]
+```
+

models/rank/fibinet/__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/fibinet/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.fibinet"
+
+# 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 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26"
+  dense_slots: "dense_var:13"
+- name: dataset_infer # name
+  batch_size: 2
+  type: DataLoader # or QueueDataset
+  data_path: "{workspace}/data/sample_data/train"
+  sparse_slots: "click 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26"
+  dense_slots: "dense_var:13"
+
+# 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: 27
+  sparse_feature_number: 1000001
+  sparse_feature_dim: 9
+  dense_input_dim: 13
+  bilinear_type: 'all'
+  reduction_ratio: 3
+  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: 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: 100
+  phases: [phase1]
+
+- name: single_gpu_train
+  class: train
+  # num of epochs
+  epochs: 4
+  # 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: 100
+  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: 8
+
+- name: phase2
+  model: "{workspace}/model.py" # user-defined model
+  dataset_name: dataset_infer # select dataset by name
+  thread_num: 8

models/rank/fibinet/data/download.sh

+wget --no-check-certificate https://fleet.bj.bcebos.com/ctr_data.tar.gz
+tar -zxvf ctr_data.tar.gz
+mv ./raw_data ./train_data_full
+mkdir train_data && cd train_data
+cp ../train_data_full/part-0 ../train_data_full/part-1 ./ && cd ..
+mv ./test_data ./test_data_full
+mkdir test_data && cd test_data
+cp ../test_data_full/part-220 ./  && cd ..
+echo "Complete data download."
+echo "Full Train data stored in ./train_data_full "
+echo "Full Test data stored in ./test_data_full "
+echo "Rapid Verification train data stored in ./train_data "
+echo "Rapid Verification test data stored in ./test_data "

models/rank/fibinet/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
+
+cont_min_ = [0, -3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
+cont_max_ = [20, 600, 100, 50, 64000, 500, 100, 50, 500, 10, 10, 10, 50]
+cont_diff_ = [20, 603, 100, 50, 64000, 500, 100, 50, 500, 10, 10, 10, 50]
+hash_dim_ = 1000001
+continuous_range_ = range(1, 14)
+categorical_range_ = range(14, 40)
+
+
+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.rstrip('\n').split('\t')
+            dense_feature = []
+            sparse_feature = []
+            for idx in continuous_range_:
+                if features[idx] == "":
+                    dense_feature.append(0.0)
+                else:
+                    dense_feature.append(
+                        (float(features[idx]) - cont_min_[idx - 1]) /
+                        cont_diff_[idx - 1])
+            for idx in categorical_range_:
+                sparse_feature.append(
+                    [hash(str(idx) + features[idx]) % hash_dim_])
+            label = [int(features[0])]
+            process_line = dense_feature, sparse_feature, label
+            feature_name = ["dense_feature"]
+            for idx in categorical_range_:
+                feature_name.append("C" + str(idx - 13))
+            feature_name.append("label")
+            s = "click:" + str(label[0])
+            for i in dense_feature:
+                s += " dense_feature:" + str(i)
+            for i in range(1, 1 + len(categorical_range_)):
+                s += " " + str(i) + ":" + str(sparse_feature[i - 1][0])
+            print(s.strip())
+            yield None
+
+        return reader
+
+
+d = CriteoDataset()
+d.run_from_stdin()

models/rank/fibinet/data/run.sh

+sh download.sh
+
+mkdir slot_train_data_full
+for i in `ls ./train_data_full`
+do
+    cat train_data_full/$i | python get_slot_data.py > slot_train_data_full/$i
+done
+
+mkdir slot_test_data_full
+for i in `ls ./test_data_full`
+do
+    cat test_data_full/$i | python get_slot_data.py > slot_test_data_full/$i
+done
+
+mkdir slot_train_data
+for i in `ls ./train_data`
+do
+    cat train_data/$i | python get_slot_data.py > slot_train_data/$i
+done
+
+mkdir slot_test_data
+for i in `ls ./test_data`
+do
+    cat test_data/$i | python get_slot_data.py > slot_test_data/$i
+done

models/rank/fibinet/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 _SENETLayer(self, inputs, filed_size, reduction_ratio=3):
+        reduction_size = max(1, filed_size // reduction_ratio)
+        Z = fluid.layers.reduce_mean(inputs, dim=-1)
+
+        A_1 = fluid.layers.fc(
+            input=Z,
+            size=reduction_size,
+            param_attr=fluid.initializer.Xavier(uniform=False),
+            act='relu',
+            name='W_1')
+
+        A_2 = fluid.layers.fc(
+            input=A_1,
+            size=filed_size,
+            param_attr=fluid.initializer.Xavier(uniform=False),
+            act='relu',
+            name='W_2')
+
+        V = fluid.layers.elementwise_mul(
+            inputs, y=fluid.layers.unsqueeze(
+                input=A_2, axes=[2]))
+
+        return fluid.layers.split(V, num_or_sections=filed_size, dim=1)
+
+    def _BilinearInteraction(self,
+                             inputs,
+                             filed_size,
+                             embedding_size,
+                             bilinear_type="interaction"):
+        if bilinear_type == "all":
+            p = [
+                fluid.layers.elementwise_mul(
+                    fluid.layers.fc(
+                        input=v_i,
+                        size=embedding_size,
+                        param_attr=fluid.initializer.Xavier(uniform=False),
+                        act=None,
+                        name=None),
+                    fluid.layers.squeeze(
+                        input=v_j, axes=[1]))
+                for v_i, v_j in itertools.combinations(inputs, 2)
+            ]
+        else:
+            raise NotImplementedError
+
+        return fluid.layers.concat(input=p, axis=1)
+
+    def _DNNLayer(self, inputs, dropout_rate=0.5):
+        deep_input = inputs
+        for i, hidden_unit in enumerate([400, 400, 400]):
+            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 net(self, input, is_infer=False):
+        self.sparse_inputs = self._sparse_data_var[1:]
+        self.dense_input = self._dense_data_var[0]
+        self.label_input = self._sparse_data_var[0]
+
+        emb = []
+        for data in self.sparse_inputs:
+            feat_emb = fluid.embedding(
+                input=data,
+                size=[self.sparse_feature_number, self.sparse_feature_dim],
+                param_attr=fluid.ParamAttr(
+                    name='dis_emb',
+                    learning_rate=5,
+                    initializer=fluid.initializer.Xavier(
+                        fan_in=self.sparse_feature_dim,
+                        fan_out=self.sparse_feature_dim)),
+                is_sparse=True)
+            emb.append(feat_emb)
+        concat_emb = fluid.layers.concat(emb, axis=1)
+
+        filed_size = len(self.sparse_inputs)
+        bilinear_type = envs.get_global_env("hyper_parameters.bilinear_type")
+        reduction_ratio = envs.get_global_env(
+            "hyper_parameters.reduction_ratio")
+        dropout_rate = envs.get_global_env("hyper_parameters.dropout_rate")
+
+        senet_output = self._SENETLayer(concat_emb, filed_size,
+                                        reduction_ratio)
+        senet_bilinear_out = self._BilinearInteraction(
+            senet_output, filed_size, self.sparse_feature_dim, bilinear_type)
+
+        concat_emb = fluid.layers.split(
+            concat_emb, num_or_sections=filed_size, dim=1)
+        bilinear_out = self._BilinearInteraction(
+            concat_emb, filed_size, self.sparse_feature_dim, bilinear_type)
+        dnn_input = fluid.layers.concat(
+            input=[senet_bilinear_out, bilinear_out, self.dense_input], axis=1)
+        dnn_output = self._DNNLayer(dnn_input, dropout_rate)
+
+        y_pred = fluid.layers.fc(
+            input=dnn_output,
+            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

models/rank/readme.md

@@ -37,35 +37,43 @@
 | xDeepFM | xDeepFM | [xDeepFM: Combining Explicit and Implicit Feature Interactions for Recommender Systems](https://dl.acm.org/doi/pdf/10.1145/3219819.3220023)(2018) |
 | DIN | Deep Interest Network | [Deep Interest Network for Click-Through Rate Prediction](https://dl.acm.org/doi/pdf/10.1145/3219819.3219823)(2018) |
 | FGCNN | Feature Generation by CNN | [Feature Generation by Convolutional Neural Network for Click-Through Rate Prediction](https://arxiv.org/pdf/1904.04447.pdf)(2019) |
+| FIBINET | Combining Feature Importance and Bilinear feature Interaction | [《FiBiNET: Combining Feature Importance and Bilinear feature Interaction for Click-Through Rate Prediction》]( https://arxiv.org/pdf/1905.09433.pdf)(2019) |
 
 下面是每个模型的简介（注：图片引用自链接中的论文）
 
 [wide&deep](https://dl.acm.org/doi/pdf/10.1145/2988450.2988454):
+
 <p align="center">
 <img align="center" src="../../doc/imgs/wide&deep.png">
 <p>
-
 [DeepFM](https://arxiv.org/pdf/1703.04247.pdf):
+
 <p align="center">
 <img align="center" src="../../doc/imgs/deepfm.png">
 <p>
-
 [XDeepFM](https://dl.acm.org/doi/pdf/10.1145/3219819.3220023):
+
 <p align="center">
 <img align="center" src="../../doc/imgs/xdeepfm.png">
 <p>
-
 [DCN](https://dl.acm.org/doi/pdf/10.1145/3124749.3124754):
+
 <p align="center">
 <img align="center" src="../../doc/imgs/dcn.png">
 <p>
-
 [DIN](https://dl.acm.org/doi/pdf/10.1145/3219819.3219823):
+
 <p align="center">
 <img align="center" src="../../doc/imgs/din.png">
 <p>
+[FIBINET](https://arxiv.org/pdf/1905.09433.pdf):
+
+<p align="center">
+<img align="center" src="../../doc/imgs/fibinet.png">
+<p>
 
 ## 使用教程(快速开始)
+
 使用样例数据快速开始，参考[训练](###训练) & [预测](###预测)
 ## 使用教程(复现论文)
 为了方便使用者能够快速的跑通每一个模型，我们在每个模型下都提供了样例数据，并且调整了batch_size等超参以便在样例数据上更加友好的显示训练&测试日志。如果需要复现readme中的效果请按照如下表格调整batch_size等超参，并使用提供的脚本下载对应数据集以及数据预处理。
@@ -77,6 +85,7 @@
 |       DIN        |       32       |       10       |        100       |
 |       Wide&Deep  |       40       |       1       |        40       |
 |       xDeepFM        |       100       |       1       |        10       |
+| Fibinet | 1000 | 8 | 4 |
 ### 数据处理
 参考每个模型目录数据下载&预处理脚本
 
@@ -116,6 +125,7 @@ python -m paddlerec.run -m ./config.yaml # 以DNN为例
 |       Criteo        |       xDeepFM       |       0.48657        |       --          |       --          |       --          |
 |       Census-income Data        |       Wide&Deep       |       0.76195         |       0.90577          |       --          |       --          |
 |       Amazon Product        |       DIN       |       0.47005        |       0.86379         |       --          |       --          |
+| Criteo | Fibinet | -- | 0.86662 | -- | -- |
 
 
 ## 分布式