FateWithShellmiaoComment/Fate/python/federatedml/nn/backend/torch/interactive.py

import torch as t
from torch.nn import ReLU, Linear, LazyLinear, Tanh, Sigmoid, Dropout, Sequential
from federatedml.nn.backend.torch.base import FateTorchLayer


class InteractiveLayer(t.nn.Module, FateTorchLayer):

    r"""A :class: InteractiveLayer.

           An interface for InteractiveLayer. In interactive layer, the forward method is:
           out = activation( Linear(guest_input) + Linear(host_0_input) + Linear(host_1_input) ..)

           Args:
                out_dim: int, the output dimension of InteractiveLayer
                host_num: int, specify the number of host party, default is 1, need to modify this parameter
                               when running multi-party modeling
                guest_dim: int or None, the input dimension of guest features, if None, will use LazyLinear layer
                           that automatically infers the input dimension
                host_dim: int, or None:
                           int: the input dimension of all host features
                           None: automatically infer the input dimension of all host features
                activation: str, support relu, tanh, sigmoid
                dropout: float in 0-1, if None, dropout is disabled
                guest_bias: bias for guest linear layer
                host_bias: bias for host linear layers
                need_guest: if false, will ignore the input of guest bottom model
           """

    def __init__(
        self,
        out_dim,
        guest_dim=None,
        host_num=1,
        host_dim=None,
        activation='relu',
        dropout=None,
        guest_bias=True,
        host_bias=True,
        need_guest=True,
    ):

        t.nn.Module.__init__(self)
        FateTorchLayer.__init__(self)

        self.activation = None
        if activation is not None:
            if activation.lower() == 'relu':
                self.activation = ReLU()
            elif activation.lower() == 'tanh':
                self.activation = Tanh()
            elif activation.lower() == 'sigmoid':
                self.activation = Sigmoid()
            else:
                raise ValueError(
                    'activation not support {}, avail: relu, tanh, sigmoid'.format(activation))

        self.dropout = None
        if dropout is not None:
            assert isinstance(dropout, float), 'dropout must be a float'
            self.dropout = Dropout(p=dropout)

        assert isinstance(out_dim, int), 'out_dim must be an int >= 0'

        self.param_dict['out_dim'] = out_dim
        self.param_dict['activation'] = activation
        self.param_dict['dropout'] = dropout
        self.param_dict['need_guest'] = need_guest

        assert isinstance(
            host_num, int) and host_num >= 1, 'host number is an int >= 1'
        self.param_dict['host_num'] = host_num

        if guest_dim is not None:
            assert isinstance(guest_dim, int)
        if host_dim is not None:
            assert isinstance(host_dim, int)

        self.guest_bias = guest_bias

        self.param_dict['guest_dim'] = guest_dim
        self.param_dict['host_dim'] = host_dim
        self.param_dict['guest_bias'] = guest_bias
        self.param_dict['host_bias'] = host_bias

        if need_guest:
            if guest_dim is None:
                self.guest_model = LazyLinear(out_dim, guest_bias)
            else:
                self.guest_model = Linear(guest_dim, out_dim, guest_bias)
        else:
            self.guest_model = None

        self.out_dim = out_dim
        self.host_dim = host_dim
        self.host_bias = host_bias
        self.host_model = None
        self.need_guest = need_guest

        self.host_model = t.nn.ModuleList()
        for i in range(host_num):
            self.host_model.append(self.make_host_model())

        if self.dropout is not None:
            self.act_seq = Sequential(
                self.activation,
                self.dropout
            )
        else:
            self.act_seq = Sequential(
                self.activation
            )

    def lazy_to_linear(self, guest_dim=None, host_dims=None):

        if isinstance(
                self.guest_model,
                t.nn.LazyLinear) and guest_dim is not None:
            self.guest_model = t.nn.Linear(
                guest_dim, self.out_dim, bias=self.guest_bias)

        if isinstance(
                self.host_model[0],
                t.nn.LazyLinear) and host_dims is not None:
            new_model_list = t.nn.ModuleList()
            for dim in host_dims:
                new_model_list.append(
                    t.nn.Linear(
                        dim,
                        self.out_dim,
                        bias=self.host_bias))
            self.host_model = new_model_list

    def make_host_model(self):
        if self.host_dim is None:
            return LazyLinear(self.out_dim, self.host_bias)
        else:
            return Linear(self.host_dim, self.out_dim, self.host_bias)

    def forward(self, x_guest, x_host):

        if self.need_guest:
            g_out = self.guest_model(x_guest)
        else:
            g_out = 0

        h_out = None
        if isinstance(x_host, list):
            for m, data in zip(self.host_model, x_host):
                out_ = m(data)
                if h_out is None:
                    h_out = out_
                else:
                    h_out += out_
        else:
            h_out = self.host_model[0](x_host)

        return self.activation(g_out + h_out)