使用dag发起homo_nn 和 hetero_nn的训练流程

[lynne0012006]

问题描述1：可以使用dag发起homo_nn 和 hetero_nn的训练流程吗？如果能发起，可不可以传一些dag案例作为参考？ 问题描述2：下面dag发起test_nn_binary流程，运行到homo_nn_0组件时就报错“list indices must be integers or slices, not str”，应该怎么修改dag文件呢？

附1：报错内容： File "/data/projects/fate/fate/python/fate/components/components/nn/loader.py", line 165, in from_dict 166 module_name=data_dict["module_name"], 167 TypeError: list indices must be integers or slices, not str

附2：dag内容 dag: stage: train party_tasks: host_9999: parties:

• role: host party_id:
  • '9999' tasks: reader_0: parameters: name: breast_homo_host namespace: experiment guest_10000: parties:
• role: guest party_id:
  • '10000' tasks: reader_0: parameters: name: breast_homo_guest namespace: experiment parties:
    • role: guest party_id:
    • '10000'
    • role: host party_id:
    • '9999' tasks: reader_0: stage: default parties:
• role: guest party_id:
  • '10000'
• role: host party_id:
  • '9999' component_ref: reader homo_nn_1: inputs: data: train_data: task_output_artifact: output_artifact_key: output_data producer_task: reader_0 model: warm_start_model: task_output_artifact: output_artifact_key: output_model producer_task: homo_nn_0 component_ref: homo_nn dependent_tasks:
• homo_nn_0
• reader_0 parameters: {} homo_nn_0: inputs: data: train_data: task_output_artifact: output_artifact_key: output_data producer_task: reader_0 model: {} component_ref: homo_nn dependent_tasks:
• reader_0 parameters: runner_conf: training_args_conf: num_train_epochs: 5 per_device_train_batch_size: 64 loss_conf: BCELoss optimizer_conf: lr: 0.01 type: Adam model_conf:
  • layers:
    • in_features: 30 out_features: 16 type: Linear
    • type: ReLU
    • in_features: 16 out_features: 1 type: Linear
    • type: Sigmoid type: Sequential task_type: binary algo: fedavg evaluation_0: stage: default inputs: data: input_datas: task_output_artifact: output_artifact_key: output_data producer_task: homo_nn_1 component_ref: evaluation dependent_tasks:
• homo_nn_1 parameters: metrics:
  • auc

[talkingwallace]

不建议使用dag来定义流程了，可以参考下pipeline的例子