2.0支持纵向的lstm吗？

[FancyXun]

我看fate对torch的nn 有些封装，包括Sequential这一类，同时也看到了lstm的模型，但怎么使用呢？lstm的输出有是个tuple，没法直接add 进Sequentia吧？

[FancyXun]

bottom_model=Sequential(
nn.Linear(10, 10),
nn.LSTM(input_size=10, hidden_size=10, batch_first=True),
nn.Linear(10, 10),
),

这种定义肯定有问题吧，LSTM的输出是个tuple

[talkingwallace]

请问一下，你看的是哪个教程呢？

[FancyXun]

多谢答复，我使用的是fate 2.0里面自带的nn例子，这是我基于nn的例子写了一个，你看下有什么问题呢。

#  Copyright 2019 The FATE 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 argparse
from fate_client.pipeline.utils import test_utils
from fate_client.pipeline import FateFlowPipeline
from fate_client.pipeline.components.fate.nn.torch import nn, optim
from fate_client.pipeline.components.fate.nn.torch.base import Sequential
from fate_client.pipeline.components.fate.hetero_nn import HeteroNN, get_config_of_default_runner
from fate_client.pipeline.components.fate.psi import PSI
from fate_client.pipeline.components.fate.reader import Reader
from fate_client.pipeline.components.fate.nn.algo_params import TrainingArguments
from fate_client.pipeline.components.fate import Evaluation
from fate_client.pipeline.components.fate.nn.algo_params import FedPassArgument



class LSTMModel(Sequential):
    def __init__(self, input_size, hidden_size, num_layers, num_classes ):
        super(LSTMModel, self).__init__()
        self.lstm = nn.LSTM(input_size=input_size, hidden_size=hidden_size, num_layers=num_layers, batch_first=True)
        self.fc = nn.Linear(hidden_size, num_classes)

    def forward(self, x):
        # 设置初始状态
        # h0 = Variable(torch.zeros(num_layers, batch_size, hidden_size))
        # c0 = Variable(torch.zeros(num_layers, batch_size, hidden_size))

        # 前向传播
        out, _ = self.lstm(x)
        out = self.fc(out[:, -1, :]) 
        return out






def main(config="../../config.yaml", namespace=""):
    # obtain config
    if isinstance(config, str):
        config = test_utils.load_job_config(config)
    parties = config.parties
    guest = parties.guest[0]
    host = parties.host[0]
    arbiter = parties.arbiter[0]

    pipeline = FateFlowPipeline().set_parties(guest=guest, host=host, arbiter=arbiter)

    reader_0 = Reader("reader_0", runtime_parties=dict(guest=guest, host=host))
    reader_0.guest.task_parameters(
        namespace=f"experiment{namespace}",
        name="breast_hetero_guest"
    )
    reader_0.hosts[0].task_parameters(
        namespace=f"experiment{namespace}",
        name="breast_hetero_host"
    )
    psi_0 = PSI("psi_0", input_data=reader_0.outputs["output_data"])

    training_args = TrainingArguments(
            num_train_epochs=1,
            per_device_train_batch_size=16,
            logging_strategy='epoch'
        )

    guest_conf = get_config_of_default_runner(
        bottom_model= LSTMModel(10, 10, 3, 10),
        top_model=Sequential(
            nn.Linear(10, 1),
            nn.Sigmoid()
        ),
        training_args=training_args,
        optimizer=optim.Adam(lr=0.01),
        loss=nn.BCELoss()
    )

    host_conf = get_config_of_default_runner(
        bottom_model=nn.Linear(20, 20),
        optimizer=optim.Adam(lr=0.01),
        training_args=training_args,
        agglayer_arg=FedPassArgument(
            layer_type='linear',
            in_channels_or_features=20,
            hidden_features=20,
            out_channels_or_features=10,
            passport_mode='single',
            passport_distribute='gaussian'
        )
    )

    hetero_nn_0 = HeteroNN(
        'hetero_nn_0',
        train_data=psi_0.outputs['output_data']
    )

    hetero_nn_0.guest.task_parameters(runner_conf=guest_conf)
    hetero_nn_0.hosts[0].task_parameters(runner_conf=host_conf)

    hetero_nn_1 = HeteroNN(
        'hetero_nn_1',
        test_data=psi_0.outputs['output_data'],
        predict_model_input=hetero_nn_0.outputs['train_model_output']
    )

    evaluation_0 = Evaluation(
        'eval_0',
        runtime_parties=dict(guest=guest),
        metrics=['auc'],
        input_data=[hetero_nn_1.outputs['predict_data_output'], hetero_nn_0.outputs['train_data_output']]
    )

    pipeline.add_tasks([reader_0, psi_0, hetero_nn_0, hetero_nn_1, evaluation_0])
    pipeline.compile()
    pipeline.fit()

    result_summary = pipeline.get_task_info("eval_0").get_output_metric()[0]["data"]
    print(f"result_summary: {result_summary}")


if __name__ == "__main__":

    parser = argparse.ArgumentParser("PIPELINE DEMO")
    parser.add_argument("--config", type=str, default="../config.yaml",
                        help="config file")
    parser.add_argument("--namespace", type=str, default="",
                        help="namespace for data stored in FATE")
    args = parser.parse_args()
    main(config=args.config, namespace=args.namespace)

fate_client能编译通过，但是在fate server端回报错：

lstm的输出是一个tuple，不是一个tensor，我还特点在自定义的model的forward阶段处理了下，这个模型使用torch自带也能预测出结果。不知道fate是不是对torch的一些中间接口改了，导致模型发送到server阶段后，重新加载，丢失了一些信息。 或者有没有一个可以用纵向的lstm用作分类的例子呢？lstm作为bottle model，或者top model都可以。参考官方自带的nn例子总有bug，不知道是不是我使用的不对，还请多指教。@talkingwallace

[talkingwallace]

这个是一个基于pipeline的例子，可以参考下ml直接运行的例子呢