Add pdn model

.git_bin_path

@@ -38,7 +38,7 @@
 {"leaf_name": "data/test/movielens_1m", "leaf_file": ["data/test/movielens_1m/ml_test_data", "data/test/movielens_1m/ml_train_data"]}
 {"leaf_name": "data/test/mt_ckpt", "leaf_file": ["data/test/mt_ckpt/model.ckpt-100.data-00000-of-00001", "data/test/mt_ckpt/model.ckpt-100.index", "data/test/mt_ckpt/model.ckpt-100.meta"]}
 {"leaf_name": "data/test/rtp", "leaf_file": ["data/test/rtp/taobao_fg_pred.out", "data/test/rtp/taobao_test_bucketize_feature.txt", "data/test/rtp/taobao_test_feature.txt", "data/test/rtp/taobao_test_input.txt", "data/test/rtp/taobao_train_bucketize_feature.txt", "data/test/rtp/taobao_train_feature.txt", "data/test/rtp/taobao_train_input.txt", "data/test/rtp/taobao_valid.csv", "data/test/rtp/taobao_valid_feature.txt"]}
-{"leaf_name": "data/test/tb_data", "leaf_file": ["data/test/tb_data/taobao_ad_feature_gl", "data/test/tb_data/taobao_clk_edge_gl", "data/test/tb_data/taobao_multi_seq_test_data", "data/test/tb_data/taobao_multi_seq_train_data", "data/test/tb_data/taobao_noclk_edge_gl", "data/test/tb_data/taobao_test_data", "data/test/tb_data/taobao_test_data_compress.gz", "data/test/tb_data/taobao_test_data_for_expr", "data/test/tb_data/taobao_test_data_kd", "data/test/tb_data/taobao_test_data_remap_label", "data/test/tb_data/taobao_train_data", "data/test/tb_data/taobao_train_data_for_expr", "data/test/tb_data/taobao_train_data_kd", "data/test/tb_data/taobao_train_data_remap_label", "data/test/tb_data/taobao_user_profile_gl"]}
+{"leaf_name": "data/test/tb_data", "leaf_file": ["data/test/tb_data/taobao_ad_feature_gl", "data/test/tb_data/taobao_clk_edge_gl", "data/test/tb_data/taobao_multi_seq_test_data", "data/test/tb_data/taobao_multi_seq_train_data", "data/test/tb_data/taobao_noclk_edge_gl", "data/test/tb_data/taobao_pdn_fake_test_data", "data/test/tb_data/taobao_pdn_fake_train_data", "data/test/tb_data/taobao_test_data", "data/test/tb_data/taobao_test_data_compress.gz", "data/test/tb_data/taobao_test_data_for_expr", "data/test/tb_data/taobao_test_data_kd", "data/test/tb_data/taobao_test_data_remap_label", "data/test/tb_data/taobao_train_data", "data/test/tb_data/taobao_train_data_for_expr", "data/test/tb_data/taobao_train_data_kd", "data/test/tb_data/taobao_train_data_remap_label", "data/test/tb_data/taobao_user_profile_gl"]}
 {"leaf_name": "data/test/tb_data/hard_negative_sampler_edge", "leaf_file": ["data/test/tb_data/hard_negative_sampler_edge/taobao_noclk_edge_gl.csv"]}
 {"leaf_name": "data/test/tb_data/hard_negative_sampler_item", "leaf_file": ["data/test/tb_data/hard_negative_sampler_item/taobao_ad_feature_gl.csv"]}
 {"leaf_name": "data/test/tb_data/hard_negative_sampler_user", "leaf_file": ["data/test/tb_data/hard_negative_sampler_user/taobao_user_profile_gl.csv"]}

.git_bin_url

@@ -38,7 +38,7 @@
 {"leaf_path": "data/test/movielens_1m", "sig": "99badbeec64f2fcabe0dfa1d2bfd8fb5", "remote_path": "data/git_oss_sample_data/data_test_movielens_1m_99badbeec64f2fcabe0dfa1d2bfd8fb5"}
 {"leaf_path": "data/test/mt_ckpt", "sig": "803499f48e2df5e51ce5606e9649c6d4", "remote_path": "data/git_oss_sample_data/data_test_mt_ckpt_803499f48e2df5e51ce5606e9649c6d4"}
 {"leaf_path": "data/test/rtp", "sig": "76cda60582617ddbb7cd5a49eb68a4b9", "remote_path": "data/git_oss_sample_data/data_test_rtp_76cda60582617ddbb7cd5a49eb68a4b9"}
-{"leaf_path": "data/test/tb_data", "sig": "f1279ca42de1734be321e88f85775d5f", "remote_path": "data/git_oss_sample_data/data_test_tb_data_f1279ca42de1734be321e88f85775d5f"}
+{"leaf_path": "data/test/tb_data", "sig": "b1579db090d72b3b70b59ba3c7692701", "remote_path": "data/git_oss_sample_data/data_test_tb_data_b1579db090d72b3b70b59ba3c7692701"}
 {"leaf_path": "data/test/tb_data/hard_negative_sampler_edge", "sig": "48f994681d719a2546ec4003fcbc638c", "remote_path": "data/git_oss_sample_data/data_test_tb_data_hard_negative_sampler_edge_48f994681d719a2546ec4003fcbc638c"}
 {"leaf_path": "data/test/tb_data/hard_negative_sampler_item", "sig": "f23a9eb9457c14a8e57b455804b1f013", "remote_path": "data/git_oss_sample_data/data_test_tb_data_hard_negative_sampler_item_f23a9eb9457c14a8e57b455804b1f013"}
 {"leaf_path": "data/test/tb_data/hard_negative_sampler_user", "sig": "23514156eae5a4250ac1d0a118883430", "remote_path": "data/git_oss_sample_data/data_test_tb_data_hard_negative_sampler_user_23514156eae5a4250ac1d0a118883430"}

.gitattributes

@@ -5,6 +5,7 @@ data/test/inference/fg_export_multi/variables/variables.index filter=lfs diff=lf
 data/test/inference/tb_multitower_export/assets/pipeline.config filter=lfs diff=lfs merge=lfs -text
 data/test/latest_ckpt_test/model.ckpt-500.meta filter=lfs diff=lfs merge=lfs -text
 data/test/tb_data/taobao_test_data filter=lfs diff=lfs merge=lfs -text
+data/test/tb_data/taobao_pdn_fake_test_data filter=lfs diff=lfs merge=lfs -text
 data/test/tb_data/taobao_multi_seq_test_data filter=lfs diff=lfs merge=lfs -text
 data/test/test.csv filter=lfs diff=lfs merge=lfs -text
 data/test/inference/tb_multitower_placeholder_rename_export/assets/pipeline.config filter=lfs diff=lfs merge=lfs -text
@@ -16,6 +17,7 @@ data/test/criteo_sample.tfrecord filter=lfs diff=lfs merge=lfs -text
 data/test/rtp/taobao_valid.csv filter=lfs diff=lfs merge=lfs -text
 data/test/rtp/taobao_train_feature.txt filter=lfs diff=lfs merge=lfs -text
 data/test/tb_data/taobao_train_data filter=lfs diff=lfs merge=lfs -text
+data/test/tb_data/taobao_pdn_fake_train_data filter=lfs diff=lfs merge=lfs -text
 data/test/tb_data/taobao_multi_seq_train_data filter=lfs diff=lfs merge=lfs -text
 data/test/inference/fg_export_single/variables/variables.index filter=lfs diff=lfs merge=lfs -text
 data/test/inference/lookup_data_test80.csv filter=lfs diff=lfs merge=lfs -text

README.md

@@ -60,7 +60,7 @@ Running Platform:
 
 ### A variety of models
 
-- [DSSM](docs/source/models/dssm.md) / [MIND](docs/source/models/mind.md) / [DropoutNet](docs/source/models/dropoutnet.md) / [CoMetricLearningI2I](docs/source/models/co_metric_learning_i2i.md)
+- [DSSM](docs/source/models/dssm.md) / [MIND](docs/source/models/mind.md) / [DropoutNet](docs/source/models/dropoutnet.md) / [CoMetricLearningI2I](docs/source/models/co_metric_learning_i2i.md) / [PDN](docs/source/models/pdn.md)
 - [W&D](docs/source/models/wide_and_deep.md) / [DeepFM](docs/source/models/deepfm.md) / [MultiTower](docs/source/models/multi_tower.md) / [DCN](docs/source/models/dcn.md) / [FiBiNet](docs/source/models/fibinet.md) / [MaskNet](docs/source/models/masknet.md)
 - [DIN](docs/source/models/din.md) / [BST](docs/source/models/bst.md)
 - [MMoE](docs/source/models/mmoe.md) / [ESMM](docs/source/models/esmm.md) / [DBMTL](docs/source/models/dbmtl.md) / [PLE](docs/source/models/ple.md)

docs/source/intro.md

@@ -46,7 +46,7 @@ EasyRec implements state of the art machine learning models used in common recom
 
 ### A variety of models
 
-- [DSSM](models/dssm.md) / [MIND](models/mind.md) / [DropoutNet](models/dropoutnet.md) / [CoMetricLearningI2I](models/co_metric_learning_i2i.md)
+- [DSSM](models/dssm.md) / [MIND](models/mind.md) / [DropoutNet](models/dropoutnet.md) / [CoMetricLearningI2I](models/co_metric_learning_i2i.md) / [PDN](models/pdn.md)
 - [W&D](models/wide_and_deep.md) / [DeepFM](models/deepfm.md) / [MultiTower](models/multi_tower.md) / [DCN](models/dcn.md) / [DIN](models/din.md) / [BST](models/bst.md)
 - [MMoE](models/mmoe.md) / [ESMM](models/esmm.md) / [DBMTL](models/dbmtl.md) / [PLE](models/ple.md)
 - [CMBF](models/cmbf.md) / [UNITER](models/uniter.md)

docs/source/models/pdn.md

@@ -0,0 +1,114 @@
+# PDN
+
+### 简介
+
+在推荐系统的召回阶段，工业界常多采用 基于Item的协同过滤 (Item-based CF) 和 基于embedding的检索 (EBR)。然而，Item-based CF很难满足个性化，而EBR则很难满足多样性。
+
+论文中提出了一种新颖的匹配架构——基于路径的深度网络（Path-based Deep Network, PDN）, 它可以结合个性化和多样性来提高匹配性能。
+
+具体来说，PDN由两个模块组成：Trigger Net和Similarity Net（SimNet，与下文一致）。 PDN 利用 Trigger Net 来捕获用户对其每个交互item的兴趣，并利用 SimNet 根据这些item的特征和共现信息来评估每个交互item与目标item之间的相似度。
+
+用户和目标item之间的最终相关性是通过明确考虑用户的不同兴趣来计算的，即聚合相关两跳路径的相关权重（路径的一跳对应于user - trigger item交互，另一跳对应于trigger item - target item相关性）。
+
+![pdn.png](../../images/models/pdn.jpg)
+（图片来自参考论文）
+
+### 配置说明
+
+#### PDN
+
+```protobuf
+model_config {
+  model_class: "PDN"
+  feature_groups: {
+    group_name: 'u2i_seq'
+    feature_names: 'event_type_seq'
+  }
+  feature_groups: {
+    group_name: 'i_seq'
+    feature_names: 'tag_category_list'
+  }
+  feature_groups: {
+    group_name: 'i2i_seq'
+    feature_names: 'i2i_rnk'
+  }
+  feature_groups: {
+    group_name: 'user'
+    feature_names: 'user_id'
+    feature_names: 'cms_segid'
+    wide_deep:DEEP
+  }
+  feature_groups: {
+    group_name: "item"
+    feature_names: 'adgroup_id'
+    feature_names: 'cate_id'
+    wide_deep:DEEP
+  }
+  feature_groups {
+    group_name: 'bias'
+    feature_names: 'position'
+  }
+
+  pdn {
+    user_dnn {
+      hidden_units: [128, 64, 32]
+    }
+    item_dnn {
+      hidden_units: [128, 64, 32]
+    }
+    u2i_dnn {
+      hidden_units: [64, 32]
+    }
+    i2i_dnn {
+      hidden_units: [128, 64, 32]
+    }
+    trigger_dnn {
+      hidden_units: [64, 32, 1]
+    }
+    sim_dnn {
+      hidden_units: [64, 32, 1]
+    }
+    bias_dnn {
+      hidden_units: [32, 32, 1]
+    }
+
+    l2_regularization: 1e-6
+  }
+  embedding_regularization: 5e-6
+}
+```
+
+- model_class: 'PDN', 不需要修改
+
+- feature_groups: 需要六个feature_group: u2i_seq, i_seq, i2i_seq, user, item 和 bias.
+
+  **group name不能变, 且u2i_seq, i_seq, i2i_seq的输入均为Sequence Feature.**
+
+  特征说明可以参考下图
+  ![pdn_1.png](../../images/models/pdn_1.png)
+
+  - u2i_seq: u用户对交互过的trigger物品的行为侧信息的序列，如 event_type, click_time等
+  - i_seq: 用户交互过的trigger物品的相关特征的序列，如 category, brand等
+  - i2i_seq: trigger物品到target物品的统计特征和共现特征，如 是否同品牌, swing_score等
+  - user: 用户侧相关特征，如 age, sex等
+  - item: 物品侧相关特征，如 category, brand等
+  - bias: 导致选择性偏差的特征（位置信息等），如 position等
+
+- pdn: pdn模型相关的参数
+
+  - user_dnn：构造DNN编码user特征信息转换为向量
+  - item_dnn：构造DNN编码item特征信息转换为向量
+  - u2i_dnn：构造DNN编码user to trigger item的行为特征信息转换为向量
+  - i2i_dnn：构造DNN编码trigger item to target item特征转换为向量
+  - trigger_dnn：构造TriggerNet DNN计算user对每个交互item的喜爱程度来捕获user的多峰兴趣
+  - sim_dnn：构造SimNet DNN基于item信息和共现信息计算trigger item与target item的相似度
+  - bias_dnn (可选)：基于会导致选择性偏差的特征（位置信息等）训练一个浅层DNN
+  - embedding_regularization: 对embedding部分加regularization，防止overfit
+
+### 示例Config
+
+[PDN_demo.config](../../../samples/model_config/pdn_on_taobao.config)
+
+### 参考论文
+
+[PDN](https://arxiv.org/abs/2105.08246)

easy_rec/python/model/pdn.py

@@ -0,0 +1,203 @@
+# -*- encoding:utf-8 -*-
+# Copyright (c) Alibaba, Inc. and its affiliates.
+
+import tensorflow as tf
+
+from easy_rec.python.layers import dnn
+from easy_rec.python.model.match_model import MatchModel
+from easy_rec.python.protos.simi_pb2 import Similarity
+
+if tf.__version__ >= '2.0':
+  tf = tf.compat.v1
+losses = tf.losses
+metrics = tf.metrics
+
+
+class PDN(MatchModel):
+
+  def __init__(self,
+               model_config,
+               feature_configs,
+               features,
+               labels=None,
+               is_training=False):
+    super(PDN, self).__init__(model_config, feature_configs, features, labels,
+                              is_training)
+    assert self._model_config.WhichOneof('model') == 'pdn', \
+        'invalid model config: %s' % self._model_config.WhichOneof('model')
+    self._model_config = self._model_config.pdn
+
+    self._user_features, _ = self._input_layer(self._feature_dict, 'user')
+    self._item_features, _ = self._input_layer(self._feature_dict, 'item')
+
+    if self._input_layer.has_group('bias'):
+      self._bias_features, _ = self._input_layer(self._feature_dict, 'bias')
+    else:
+      self._bias_features = None
+
+    self._u2i_seq, self._seq_len = self._get_seq_features('u2i_seq')
+    self._i_seq, _ = self._get_seq_features('i_seq')
+    self._i2i_seq, _ = self._get_seq_features('i2i_seq')
+
+  def build_predict_graph(self):
+    trigger_out = self._build_trigger_net()
+    sim_out = self._build_similarity_net()
+    logits = tf.multiply(sim_out, trigger_out)
+
+    seq_mask = tf.to_float(
+        tf.sequence_mask(self._seq_len,
+                         tf.shape(sim_out)[1]))
+    logits = tf.reduce_sum(logits * seq_mask[:, :, None], axis=1)
+
+    direct_logits = self._build_direct_net()
+    if direct_logits is not None:
+      logits += direct_logits
+
+    bias_logits = self._build_bias_net()
+    if bias_logits is not None:
+      logits += bias_logits
+
+    logits = tf.squeeze(logits, axis=1)
+    probs = 1 - tf.exp(-logits)  # map [0, inf) to [0, 1)
+
+    self._prediction_dict['probs'] = probs
+    self._prediction_dict['logits'] = tf.log(
+        tf.clip_by_value(probs, 1e-8, 1 - 1e-8))
+    return self._prediction_dict
+
+  def _get_seq_features(self, name):
+    seqs, _, _ = self._input_layer(self._feature_dict, name, is_combine=False)
+    seq_len = seqs[0][1]
+    seq = tf.concat([x[0] for x in seqs], axis=2)
+    return seq, seq_len
+
+  def _build_trigger_net(self):
+    user_dnn_layer = dnn.DNN(self._model_config.user_dnn, self._l2_reg,
+                             'user_dnn', self._is_training)
+    user_fea = user_dnn_layer(self._user_features)
+
+    trigger_seq = tf.concat([self._u2i_seq, self._i_seq], axis=2)
+    u2i_dnn_layer = dnn.DNN(self._model_config.u2i_dnn, self._l2_reg, 'u2i_dnn',
+                            self._is_training)
+    trigger_seq_fea = u2i_dnn_layer(trigger_seq)
+
+    trigger_merge_fea = trigger_seq_fea + user_fea[:, None, :]
+    trigger_dnn_layer = dnn.DNN(
+        self._model_config.trigger_dnn,
+        self._l2_reg,
+        'trigger_dnn',
+        self._is_training,
+        last_layer_no_activation=True,
+        last_layer_no_batch_norm=True)
+
+    # output: N x seq_len x d, d is usually set to 1
+    trigger_out = trigger_dnn_layer(trigger_merge_fea)
+    # exp(x): map (-inf, inf) to (0, inf)
+    trigger_out = tf.exp(trigger_out)
+
+    self._prediction_dict['trigger_out'] = tf.reduce_join(
+        tf.reduce_join(
+            tf.as_string(trigger_out, precision=4, shortest=True),
+            axis=2,
+            separator=','),
+        axis=1,
+        separator=';')
+    return trigger_out
+
+  def _build_similarity_net(self):
+    item_dnn_layer = dnn.DNN(self._model_config.item_dnn, self._l2_reg,
+                             'item_dnn', self._is_training)
+    item_fea = item_dnn_layer(self._item_features)
+
+    sim_side_dnn_layer = dnn.DNN(self._model_config.i2i_dnn, self._l2_reg,
+                                 'i2i_dnn', self._is_training)
+    sim_seq_fea = sim_side_dnn_layer(self._i_seq)
+
+    sim_seq_cross = sim_seq_fea * item_fea[:, None, :]
+
+    item_fea_tile = tf.tile(item_fea[:, None, :],
+                            [1, tf.shape(sim_seq_fea)[1], 1])
+
+    sim_seq_concat = tf.concat(
+        [sim_seq_cross, sim_seq_cross, self._i2i_seq, item_fea_tile], axis=2)
+    sim_dnn_layer = dnn.DNN(
+        self._model_config.sim_dnn,
+        self._l2_reg,
+        'sim_dnn',
+        self._is_training,
+        last_layer_no_activation=True,
+        last_layer_no_batch_norm=True)
+    # output: N x seq_len x 1
+    sim_out = sim_dnn_layer(sim_seq_concat)
+    # exp(x): map (-inf, inf) to (0, inf)
+    sim_out = tf.exp(sim_out)
+
+    self._prediction_dict['sim_out'] = tf.reduce_join(
+        tf.reduce_join(
+            tf.as_string(sim_out, precision=4, shortest=True),
+            axis=2,
+            separator=','),
+        axis=1,
+        separator=';')
+    return sim_out
+
+  def _build_direct_net(self):
+    if self._model_config.HasField('direct_user_dnn') and \
+       self._model_config.HasField('direct_item_dnn'):
+      direct_user_layer = dnn.DNN(
+          self._model_config.direct_user_dnn,
+          'direct_user_dnn',
+          self._is_training,
+          last_layer_no_activation=True,
+          last_layer_no_batch_norm=True)
+      direct_user_out = direct_user_layer(self._user_features)
+      direct_item_layer = dnn.DNN(
+          self._model_config.direct_item_dnn,
+          'direct_item_dnn',
+          self._is_training,
+          last_layer_no_activation=True,
+          last_layer_no_batch_norm=True)
+      direct_item_out = direct_item_layer(self._item_features)
+
+      if self._model_config.simi_func == Similarity.COSINE:
+        direct_user_out = self.norm(direct_user_out)
+        direct_item_out = self.norm(direct_item_out)
+
+      self._prediction_dict['direct_user_embedding'] = direct_user_out
+      self._prediction_dict['direct_item_embedding'] = direct_item_out
+      direct_logits = tf.reduce_sum(direct_user_out * direct_item_out, axis=1)
+
+      if self._model_config.scale_simi:
+        sim_w = tf.get_variable(
+            'direct_net/sim_w',
+            dtype=tf.float32,
+            shape=(1),
+            initializer=tf.ones_initializer())
+        sim_b = tf.get_variable(
+            'direct_net/sim_b',
+            dtype=tf.float32,
+            shape=(1),
+            initializer=tf.zeros_initializer())
+        direct_logits = direct_logits * tf.abs(sim_w) + sim_b
+
+      return tf.nn.softplus(direct_logits)
+    else:
+      return None
+
+  def _build_bias_net(self):
+    if self._model_config.HasField('bias_dnn'):
+      assert self._bias_features is not None, 'bias group must be defined'
+      bias_dnn_layer = dnn.DNN(
+          self._model_config.bias_dnn,
+          self._l2_reg,
+          'bias_dnn',
+          self._is_training,
+          last_layer_no_activation=True,
+          last_layer_no_batch_norm=True)
+      bias_logits = bias_dnn_layer(self._bias_features)
+      return tf.nn.softplus(bias_logits)
+    else:
+      return None
+
+  def get_outputs(self):
+    return ['logits', 'probs', 'trigger_out', 'sim_out']

easy_rec/python/protos/easy_rec_model.proto

@@ -26,6 +26,7 @@ import "easy_rec/python/protos/rocket_launching.proto";
 import "easy_rec/python/protos/variational_dropout.proto";
 import "easy_rec/python/protos/multi_tower_recall.proto";
 import "easy_rec/python/protos/tower.proto";
+import "easy_rec/python/protos/pdn.proto";
 
 // for input performance test
 message DummyModel {
@@ -82,6 +83,7 @@ message EasyRecModel {
         MIND mind = 202;
         DropoutNet dropoutnet = 203;
         CoMetricLearningI2I metric_learning = 204;
+        PDN pdn = 205;
 
         MMoE mmoe = 301;
         ESMM esmm = 302;