[feat] backbone

examples/modular/match/dssm_taobao_local_backbone.config

@@ -0,0 +1,319 @@
+train_input_path: "data/taobao_data_recall_train/*.parquet"
+eval_input_path: "data/taobao_data_recall_eval/*.parquet"
+model_dir: "experiments/dssm_taobao_backbone"
+
+train_config {
+    sparse_optimizer {
+        adagrad_optimizer {
+            lr: 0.001
+        }
+        constant_learning_rate {
+        }
+    }
+    dense_optimizer {
+        adam_optimizer {
+            lr: 0.001
+        }
+        constant_learning_rate {
+        }
+    }
+    num_epochs: 8
+}
+
+eval_config {
+}
+
+data_config {
+    batch_size: 2048
+    dataset_type: ParquetDataset
+    fg_mode: FG_DAG
+    label_fields: "clk"
+    num_workers: 8
+    negative_sampler {
+        input_path: "data/taobao_ad_feature_gl"
+        num_sample: 4096
+        attr_fields: "adgroup_id"
+        attr_fields: "cate_id"
+        attr_fields: "campaign_id"
+        attr_fields: "customer"
+        attr_fields: "brand"
+        attr_fields: "price"
+        item_id_field: "adgroup_id"
+        attr_delimiter: "\x02"
+    }
+}
+
+feature_configs {
+    id_feature {
+        feature_name: "user_id"
+        expression: "user:user_id"
+        num_buckets: 1141730
+        embedding_dim: 16
+    }
+}
+feature_configs {
+    id_feature {
+        feature_name: "cms_segid"
+        expression: "user:cms_segid"
+        num_buckets: 98
+        embedding_dim: 16
+    }
+}
+feature_configs {
+    id_feature {
+        feature_name: "cms_group_id"
+        expression: "user:cms_group_id"
+        num_buckets: 14
+        embedding_dim: 16
+    }
+}
+feature_configs {
+    id_feature {
+        feature_name: "final_gender_code"
+        expression: "user:final_gender_code"
+        num_buckets: 3
+        embedding_dim: 16
+    }
+}
+feature_configs {
+    id_feature {
+        feature_name: "age_level"
+        expression: "user:age_level"
+        num_buckets: 8
+        embedding_dim: 16
+    }
+}
+feature_configs {
+    id_feature {
+        feature_name: "pvalue_level"
+        expression: "user:pvalue_level"
+        num_buckets: 5
+        embedding_dim: 16
+    }
+}
+feature_configs {
+    id_feature {
+        feature_name: "shopping_level"
+        expression: "user:shopping_level"
+        num_buckets: 5
+        embedding_dim: 16
+    }
+}
+feature_configs {
+    id_feature {
+        feature_name: "occupation"
+        expression: "user:occupation"
+        num_buckets: 3
+        embedding_dim: 16
+    }
+}
+feature_configs {
+    id_feature {
+        feature_name: "new_user_class_level"
+        expression: "user:new_user_class_level"
+        num_buckets: 6
+        embedding_dim: 16
+    }
+}
+feature_configs {
+    id_feature {
+        feature_name: "adgroup_id"
+        expression: "item:adgroup_id"
+        num_buckets: 846812
+        embedding_dim: 16
+    }
+}
+feature_configs {
+    id_feature {
+        feature_name: "cate_id"
+        expression: "item:cate_id"
+        num_buckets: 12961
+        embedding_dim: 16
+    }
+}
+feature_configs {
+    id_feature {
+        feature_name: "campaign_id"
+        expression: "item:campaign_id"
+        num_buckets: 423438
+        embedding_dim: 16
+    }
+}
+feature_configs {
+    id_feature {
+        feature_name: "customer"
+        expression: "item:customer"
+        num_buckets: 255877
+        embedding_dim: 16
+    }
+}
+feature_configs {
+    id_feature {
+        feature_name: "brand"
+        expression: "item:brand"
+        num_buckets: 461498
+        embedding_dim: 16
+    }
+}
+feature_configs {
+    raw_feature {
+        feature_name: "price"
+        expression: "item:price"
+        boundaries: [1.1, 2.2, 3.6, 5.2, 7.39, 9.5, 10.5, 12.9, 15, 17.37, 19, 20, 23.8, 25.8, 28, 29.8, 31.5, 34, 36, 38, 39, 40, 45, 48, 49, 51.6, 55.2, 58, 59, 63.8, 68, 69, 72, 78, 79, 85, 88, 90, 97.5, 98, 99, 100, 108, 115, 118, 124, 128, 129, 138, 139, 148, 155, 158, 164, 168, 171.8, 179, 188, 195, 198, 199, 216, 228, 238, 248, 258, 268, 278, 288, 298, 299, 316, 330, 352, 368, 388, 398, 399, 439, 478, 499, 536, 580, 599, 660, 699, 780, 859, 970, 1080, 1280, 1480, 1776, 2188, 2798, 3680, 5160, 8720]
+        embedding_dim: 16
+    }
+}
+feature_configs {
+    id_feature {
+        feature_name: "pid"
+        expression: "context:pid"
+        hash_bucket_size: 20
+        embedding_dim: 16
+    }
+}
+feature_configs {
+    sequence_feature {
+        sequence_name: "click_50_seq"
+        sequence_length: 100
+        sequence_delim: "|"
+        features {
+            id_feature {
+                feature_name: "adgroup_id"
+                expression: "item:adgroup_id"
+                num_buckets: 846812
+                embedding_dim: 16
+            }
+        }
+        features {
+            id_feature {
+                feature_name: "cate_id"
+                expression: "item:cate_id"
+                num_buckets: 12961
+                embedding_dim: 16
+            }
+        }
+        features {
+            id_feature {
+                feature_name: "brand"
+                expression: "item:brand"
+                num_buckets: 461498
+                embedding_dim: 16
+            }
+        }
+    }
+}
+
+model_config {
+    feature_groups {
+        group_name: "user"
+        feature_names: "user_id"
+        feature_names: "cms_segid"
+        feature_names: "cms_group_id"
+        feature_names: "final_gender_code"
+        feature_names: "age_level"
+        feature_names: "pvalue_level"
+        feature_names: "shopping_level"
+        feature_names: "occupation"
+        feature_names: "new_user_class_level"
+        feature_names: "pid"
+        sequence_groups {
+            group_name: "click_50_seq"
+            feature_names: "click_50_seq__adgroup_id"
+            feature_names: "click_50_seq__cate_id"
+            feature_names: "click_50_seq__brand"
+        }
+        sequence_encoders {
+            pooling_encoder: {
+                input: "click_50_seq"
+                pooling_type: "mean"
+            }
+        }
+        group_type: DEEP
+    }
+    feature_groups {
+        group_name: "item"
+        feature_names: "adgroup_id"
+        feature_names: "cate_id"
+        feature_names: "campaign_id"
+        feature_names: "customer"
+        feature_names: "brand"
+        feature_names: "price"
+        group_type: DEEP
+    }
+
+    # 使用组件化的match_backbone配置
+    match_backbone {
+        backbone {
+            # 主要的backbone block定义
+            blocks {
+                name: "user"
+                inputs {
+                    feature_group_name: "user"
+                }
+                input_layer {
+                }
+            }
+            blocks {
+                name: "item"
+                inputs {
+                    feature_group_name: "item"
+                }
+                input_layer {
+                }
+            }
+            # 用户塔MLP
+            blocks {
+                name: "user_tower"
+                inputs {
+                    block_name: "user"
+                }
+                module {
+                    class_name: "MLP"
+                    mlp {
+                        hidden_units: [512, 256, 128]
+                        activation: "nn.ReLU"
+                        use_bn: false
+                        dropout_ratio: [0.0, 0.0, 0.0]
+                    }
+                }
+            }
+            # 物品塔MLP
+            blocks {
+                name: "item_tower"
+                inputs {
+                    block_name: "item"
+                }
+                module {
+                    class_name: "MLP"
+                    mlp {
+                        hidden_units: [512, 256, 128]
+                        activation: "nn.ReLU"
+                        use_bn: false
+                        dropout_ratio: [0.0, 0.0, 0.0]
+                    }
+                }
+            }
+            # 输出blocks配置 - 指定用户塔和物品塔的输出
+            output_blocks: "user_tower"
+            output_blocks: "item_tower"
+        }
+        model_params {
+            # 可以在这里配置一些通用参数
+            # 具体的output_dim、similarity等参数会通过代码默认值处理
+        }
+    }
+
+    metrics {
+        recall_at_k {
+            top_k: 1
+        }
+    }
+    metrics {
+        recall_at_k {
+            top_k: 5
+        }
+    }
+    losses {
+        softmax_cross_entropy {}
+    }
+}