FATE 实战

FATE 实战

FATE 实战：纵向联邦学习

数据集预处理

数据集：California 房价预测数据集（内置在 sklearn 库），样本数为 20640，共有 8 个特征

import pandas as pd
from sklearn.datasets import fetch_california_housing

# 数据预处理
california_dataset = fetch_california_housing()
california = pd.DataFrame(california_dataset.data, columns=california_dataset.feature_names)
california = (california-california.mean())/(california.std())
california.head()

col_names = california.columns.values.tolist()
columns = {}
for idx, n in enumerate(col_names):
	columns[n] = "x%d"%idx
california = california.rename(columns=columns)
california['y'] = california_dataset.target
california['idx'] = range(california.shape[0])
idx = california['idx']
california.drop(labels=['idx'], axis=1, inplace = True)
california.insert(0, 'idx', idx)

纵向数据切分

切分策略：

  • 前 15000 个数据作为训练数据，后 5640 个数据作为测试数据
  • 训练数据集切分：随机抽取 12000 条数据和前 4 个特征作为机构 A 的本地数据，随机抽取 13000 条数据和后 4 个特征以及标签作为机构 B 的本地数据
  • 测试数据集切分：随机抽取 3600 条数据和前 4 个特征作为机构 A 的本地测试数据，随机抽取 3800 条数据和后 4 个特征以及标签作为机构 B 的本地测试数据
# 切分训练数据
train = california.iloc[:15000]
df1 = train.sample(12000)
df2 = train.sample(13000)
housing_1_train = df1[["idx", "x0", "x1", "x2", "x3"]]
housing_1_train.to_csv('data/house/housing_1_train.csv', index=False, header=True)
housing_2_train = df2[["idx", "y", "x4", "x5", "x6", "x7"]]
housing_2_train.to_csv('data/house/housing_2_train.csv', index=False, header=True)

# 切分测试数据
eval = california.iloc[15000:]
df1 = eval.sample(3600)
df2 = eval.sample(3800)
housing_1_eval = df1[["idx", "x0", "x1", "x2", "x3"]]
housing_1_eval.to_csv('data/house/housing_1_eval.csv', index=True, header=True)
housing_2_eval = df2[["idx", "y", "x4", "x5", "x6", "x7"]]
housing_2_eval.to_csv('data/house/housing_2_eval.csv', index=True, header=True)

通过 dsl 和 conf 运行训练和预测任务

准备工作

  1. 启动容器，进行目录挂载
docker run -it --name standalone_fate -p 8080:8080 -v ${本机地址}:${Docker内地址} federatedai/standalone_fate:1.8.0

# 例如
docker run -it --name standalone_fate -p 8080:8080 -v D:\Dropbox\学习计划\FATE\data\house:/workspace federatedai/standalone_fate:1.8.0
  1. 进入容器
docker exec -it standalone_fate /bin/bash
  1. 在容器内进入挂载地址
cd /workspace

数据上传

我们需要在本机的挂载目录下新建如下四个文件，文件会自动同步到 Docker：

  • upload_train_host_conf.json：上传训练数据至机构 1
{
    "file": "housing_1_train.csv",
    "table_name": "homo_housing_1_train",
    "namespace": "homo_host_housing_train",
    "head": 1,
    "partition": 16,
    "work_mode": 0,
    "backend": 0
}
  • upload_train_guest_conf.json：上传训练数据至机构 2
{
    "file": "housing_2_train.csv",
    "table_name": "homo_housing_2_train",
    "namespace": "homo_guest_housing_train",
    "head": 1,
    "partition": 16,
    "work_mode": 0,
    "backend": 0
}
  • upload_eval_host_conf.json：上传测试数据至机构 1
{
    "file": "housing_1_eval.csv",
    "table_name": "homo_housing_1_eval",
    "namespace": "homo_host_housing_eval",
    "head": 1,
    "partition": 16,
    "work_mode": 0,
    "backend": 0
}
  • upload_eval_guest_conf.json：上传测试数据至机构 2`
{
    "file": "housing_2_eval.csv",
    "table_name": "homo_housing_2_eval",
    "namespace": "homo_guest_housing_eval",
    "head": 1,
    "partition": 16,
    "work_mode": 0,
    "backend": 0
}
  1. 上传数据：在 bash 界面右键可粘贴命令
flow init --ip 127.0.0.1 --port 9380

flow data upload -c upload_train_host_conf.json
flow data upload -c upload_train_guest_conf.json
flow data upload -c upload_eval_host_conf.json
flow data upload -c upload_eval_guest_conf.json

模型训练

与横向联邦学习相比较，纵向联邦学习需要进行样本对齐，即在不泄露双方数据的前提下，求取出双方用户的交集，从而确定模型训练的训练数据集。

  1. 在本机的挂载目录下新建 hetero_linr_train_job_dsl.json，内容模板在 data/projects/fate/examples/dsl/v2/hetero_linear_regression/test_hetero_linr_train_job_dsl.json，增加 evaluation_0 组件

    1. reader_0：数据读取组件，支持图像数据

    2. data_transform_0：数据 IO 组件

    3. intersection_0：样本对齐组件

    4. hetero_linr_0：纵向线性回归模型组件

    5. evaluation_0：模型评估组件

{
    "components": {
        "reader_0": {
            "module": "Reader",
            "output": {
                "data": ["data"]
            }
        },
        "data_transform_0": {
            "module": "DataTransform",
            "input": {
                "data": {
                    "data": ["reader_0.data"]
                }
            },
            "output": {
                "data": ["data"],
                "model": ["model"]
            }
        },
        "intersection_0": {
            "module": "Intersection",
            "input": {
                "data": {
                    "data": ["data_transform_0.data"]
                }
            },
            "output": {
                "data": ["data"]
            }
        },
        "hetero_linr_0": {
            "module": "HeteroLinR",
            "input": {
                "data": {
                    "train_data": ["intersection_0.data"]
                }
            },
            "output": {
                "data": ["data"],
                "model": ["model"]
            }
        },
        "evaluation_0": {
            "module": "Evaluation",
            "input": {
                "data": {
                    "data": ["hetero_linr_0.data"]
                }
            },
            "output": {
                "data": ["data"]
            }
        }
    }
}
  1. 在本机的挂载目录下新建 hetero_linr_train_job_conf.json，内容模板在 data/projects/fate/examples/dsl/v2/hetero_linear_regression/test_hetero_linr_train_job_conf.json，需要修改
    1. 修改 party ID 为对应 ID
    2. 指定数据对应上传数据时的设置
    3. 修改 label_name
    4. 设置模型参数
    5. 设置 evaluation_0 相关参数
{
    "dsl_version": 2,
    "initiator": {
        "role": "guest",
        "party_id": 10000
    },
    "role": {
        "arbiter": [10000],
        "host": [10000],
        "guest": [10000]
    },
    "component_parameters": {
        "common": {
            "hetero_linr_0": {
                "penalty": "L2",
                "tol": 0.001,
                "alpha": 0.01,
                "optimizer": "sgd",
                "batch_size": -1,
                "learning_rate": 0.15,
                "init_param": {
                    "init_method": "zeros"
                },
                "max_iter": 20,
                "early_stop": "weight_diff",
                "decay": 0.0,
                "decay_sqrt": false,
                "floating_point_precision": 23
            },
            "evaluation_0": {
                "eval_type": "regression",
                "pos_label": 1
            }
        },
        "role": {
            "host": {
                "0": {
                    "reader_0": {
                        "table": {
                            "name": "homo_housing_1_train",
                            "namespace": "homo_host_housing_train"
                        }
                    },
                    "data_transform_0": {
                        "with_label": false
                    }
                }
            },
            "guest": {
                "0": {
                    "reader_0": {
                        "table": {
                            "name": "homo_housing_2_train",
                            "namespace": "homo_guest_housing_train"
                        }
                    },
                    "data_transform_0": {
                        "with_label": true,
                        "label_name": "y",
                        "label_type": "float",
                        "output_format": "dense"
                    }
                }
            }
        }
    }
}
  1. 提交任务：flow job submit -c ${conf_path} -d ${dsl_path}
flow job submit -c hetero_linr_train_job_conf.json -d hetero_linr_train_job_dsl.json
  1. 在 arbiter 节点中查看训练过程中 loss 的变化

image-20230523204952121

  1. 在 guest 节点中查看模型在训练数据上的效果

image-20230523205049002

模型评估

在模型评估阶段，官方提供的配置文件示例在

  • data/projects/fate/examples/dsl/v1/hetero_linear_regression/test_hetero_linr_validate_job_dsl.json
  • data/projects/fate/examples/dsl/v1/hetero_linear_regression/test_hetero_linr_validate_job_conf.json

与训练阶段的操作类似，我们需要对配置文件进行修改：

  1. 在本机的挂载目录下新建 hetero_linr_validate_job_dsl.json，增加 evaluation_0 组件
"evaluation_0": {
    "module": "Evaluation",
    "input": {
        "data": {
            "data": ["hetero_linr_0.data"]
        }
    }
}
  1. 在本机的挂载目录下新建 hetero_linr_validate_job_conf.json，内容修改
    1. 修改 party ID
    2. 修改数据源，修改 label_name
    3. 设置 evaluation_0 组件的参数
    4. 设置模型参数
{
    "dsl_version": 2,
    "initiator": {
        "role": "guest",
        "party_id": 10000
    },
    "role": {
        "arbiter": [
            10000
        ],
        "host": [
            10000
        ],
        "guest": [
            10000
        ]
    },
    "component_parameters": {
        "common": {
            "hetero_linr_0": {
                "penalty": "L2",
                "tol": 0.001,
                "alpha": 0.01,
                "optimizer": "sgd",
                "batch_size": -1,
                "learning_rate": 0.15,
                "init_param": {
                    "init_method": "zeros"
                },
                "max_iter": 20,
                "early_stop": "weight_diff",
                "decay": 0.0,
                "decay_sqrt": false,
                "callback_param": {
                    "callbacks": [
                        "EarlyStopping",
                        "PerformanceEvaluate"
                    ],
                    "validation_freqs": 1,
                    "early_stopping_rounds": 5,
                    "metrics": [
                        "mean_absolute_error",
                        "root_mean_squared_error"
                    ],
                    "use_first_metric_only": false,
                    "save_freq": 1
                }
            },
            "evaluation_0": {
                "eval_type": "regression",
                "pos_label": 1
            }
        },
        "role": {
            "host": {
                "0": {
                    "data_transform_0": {
                        "with_label": false
                    },
                    "reader_0": {
                        "table": {
                            "name": "homo_housing_1_train",
                            "namespace": "homo_host_housing_train"
                        }
                    },
                    "reader_1": {
                        "table": {
                            "name": "homo_housing_1_eval",
                            "namespace": "homo_host_housing_eval"
                        }
                    },
                    "data_transform_1": {
                        "with_label": false
                    }
                }
            },
            "guest": {
                "0": {
                    "data_transform_0": {
                        "with_label": true,
                        "label_name": "y",
                        "label_type": "float",
                        "output_format": "dense"
                    },
                    "reader_0": {
                        "table": {
                            "name": "homo_housing_2_train",
                            "namespace": "homo_guest_housing_train"
                        }
                    },
                    "reader_1": {
                        "table": {
                            "name": "homo_housing_2_eval",
                            "namespace": "homo_guest_housing_eval"
                        }
                    },
                    "data_transform_1": {
                        "with_label": true,
                        "label_name": "y",
                        "label_type": "float",
                        "output_format": "dense"
                    }
                }
            }
        }
    }
}
  1. 提交任务：flow job submit -c ${conf_path} -d ${dsl_path}
flow job submit -c hetero_linr_validate_job_conf.json -d hetero_linr_validate_job_dsl.json
  1. 查看 DAG 图
  1. 查看模型效果

image-20230523210606921

参考资料

  1. 【联邦学习 FATE 框架实战】（四）用 FATE 从零开始实现纵向线性回归_HarrisonWu42 的博客 - CSDN 博客
posted @ 2023-05-23 21:08  Lockegogo  阅读(85)  评论(0)  编辑  收藏  举报