deepctr_torch/models/multitask/ple.py

# -*- coding:utf-8 -*-
"""
Author:
    zanshuxun, zanshuxun@aliyun.com

Reference:
    [1] Tang H, Liu J, Zhao M, et al. Progressive layered extraction (ple): A novel multi-task learning (mtl) model for personalized recommendations[C]//Fourteenth ACM Conference on Recommender Systems. 2020.(https://dl.acm.org/doi/10.1145/3383313.3412236)
"""
import torch
import torch.nn as nn

from ..basemodel import BaseModel
from ...inputs import combined_dnn_input
from ...layers import DNN, PredictionLayer


class PLE(BaseModel):
    """Instantiates the multi level of Customized Gate Control of Progressive Layered Extraction architecture.

    :param dnn_feature_columns: An iterable containing all the features used by deep part of the model.
    :param shared_expert_num: integer, number of task-shared experts.
    :param specific_expert_num: integer, number of task-specific experts.
    :param num_levels: integer, number of CGC levels.
    :param expert_dnn_hidden_units: list, list of positive integer or empty list, the layer number and units in each layer of expert DNN.
    :param gate_dnn_hidden_units: list, list of positive integer or empty list, the layer number and units in each layer of gate DNN.
    :param tower_dnn_hidden_units: list, list of positive integer or empty list, the layer number and units in each layer of task-specific DNN.
    :param l2_reg_linear: float, L2 regularizer strength applied to linear part.
    :param l2_reg_embedding: float, L2 regularizer strength applied to embedding vector.
    :param l2_reg_dnn: float, L2 regularizer strength applied to DNN.
    :param init_std: float, to use as the initialize std of embedding vector.
    :param seed: integer, to use as random seed.
    :param dnn_dropout: float in [0,1), the probability we will drop out a given DNN coordinate.
    :param dnn_activation: Activation function to use in DNN.
    :param dnn_use_bn: bool, Whether use BatchNormalization before activation or not in DNN.
    :param task_types: list of str, indicating the loss of each tasks, ``"binary"`` for  binary logloss, ``"regression"`` for regression loss. e.g. ['binary', 'regression']
    :param task_names: list of str, indicating the predict target of each tasks.
    :param device: str, ``"cpu"`` or ``"cuda:0"``.
    :param gpus: list of int or torch.device for multiple gpus. If None, run on `device`. `gpus[0]` should be the same gpu with `device`.

    :return: A PyTorch model instance.
    """

    def __init__(self, dnn_feature_columns, shared_expert_num=1, specific_expert_num=1, num_levels=2,
                 expert_dnn_hidden_units=(256, 128), gate_dnn_hidden_units=(64,), tower_dnn_hidden_units=(64,),
                 l2_reg_linear=0.00001, l2_reg_embedding=0.00001, l2_reg_dnn=0, init_std=0.0001, seed=1024,
                 dnn_dropout=0, dnn_activation='relu', dnn_use_bn=False, task_types=('binary', 'binary'),
                 task_names=('ctr', 'ctcvr'), device='cpu', gpus=None):
        super(PLE, self).__init__(linear_feature_columns=[], dnn_feature_columns=dnn_feature_columns,
                                  l2_reg_linear=l2_reg_linear, l2_reg_embedding=l2_reg_embedding, init_std=init_std,
                                  seed=seed, device=device, gpus=gpus)
        self.num_tasks = len(task_names)
        if self.num_tasks <= 1:
            raise ValueError("num_tasks must be greater than 1!")
        if len(dnn_feature_columns) == 0:
            raise ValueError("dnn_feature_columns is null!")
        if len(task_types) != self.num_tasks:
            raise ValueError("num_tasks must be equal to the length of task_types")

        for task_type in task_types:
            if task_type not in ['binary', 'regression']:
                raise ValueError("task must be binary or regression, {} is illegal".format(task_type))

        self.specific_expert_num = specific_expert_num
        self.shared_expert_num = shared_expert_num
        self.num_levels = num_levels
        self.task_names = task_names
        self.input_dim = self.compute_input_dim(dnn_feature_columns)
        self.expert_dnn_hidden_units = expert_dnn_hidden_units
        self.gate_dnn_hidden_units = gate_dnn_hidden_units
        self.tower_dnn_hidden_units = tower_dnn_hidden_units

        def multi_module_list(num_level, num_tasks, expert_num, inputs_dim_level0, inputs_dim_not_level0, hidden_units):
            return nn.ModuleList(
                [nn.ModuleList([nn.ModuleList([DNN(inputs_dim_level0 if level_num == 0 else inputs_dim_not_level0,
                                                   hidden_units, activation=dnn_activation,
                                                   l2_reg=l2_reg_dnn, dropout_rate=dnn_dropout, use_bn=dnn_use_bn,
                                                   init_std=init_std, device=device) for _ in
                                               range(expert_num)])
                                for _ in range(num_tasks)]) for level_num in range(num_level)])

        # 1. experts
        # task-specific experts
        self.specific_experts = multi_module_list(self.num_levels, self.num_tasks, self.specific_expert_num,
                                                  self.input_dim, expert_dnn_hidden_units[-1], expert_dnn_hidden_units)

        # shared experts
        self.shared_experts = multi_module_list(self.num_levels, 1, self.specific_expert_num,
                                                self.input_dim, expert_dnn_hidden_units[-1], expert_dnn_hidden_units)

        # 2. gates
        # gates for task-specific experts
        specific_gate_output_dim = self.specific_expert_num + self.shared_expert_num
        if len(gate_dnn_hidden_units) > 0:
            self.specific_gate_dnn = multi_module_list(self.num_levels, self.num_tasks, 1,
                                                       self.input_dim, expert_dnn_hidden_units[-1],
                                                       gate_dnn_hidden_units)
            self.add_regularization_weight(
                filter(lambda x: 'weight' in x[0] and 'bn' not in x[0], self.specific_gate_dnn.named_parameters()),
                l2=l2_reg_dnn)
        self.specific_gate_dnn_final_layer = nn.ModuleList(
            [nn.ModuleList([nn.Linear(
                gate_dnn_hidden_units[-1] if len(gate_dnn_hidden_units) > 0 else self.input_dim if level_num == 0 else
                expert_dnn_hidden_units[-1], specific_gate_output_dim, bias=False)
                for _ in range(self.num_tasks)]) for level_num in range(self.num_levels)])

        # gates for shared experts
        shared_gate_output_dim = self.num_tasks * self.specific_expert_num + self.shared_expert_num
        if len(gate_dnn_hidden_units) > 0:
            self.shared_gate_dnn = nn.ModuleList([DNN(self.input_dim if level_num == 0 else expert_dnn_hidden_units[-1],
                                                      gate_dnn_hidden_units, activation=dnn_activation,
                                                      l2_reg=l2_reg_dnn, dropout_rate=dnn_dropout, use_bn=dnn_use_bn,
                                                      init_std=init_std, device=device) for level_num in
                                                  range(self.num_levels)])
            self.add_regularization_weight(
                filter(lambda x: 'weight' in x[0] and 'bn' not in x[0], self.shared_gate_dnn.named_parameters()),
                l2=l2_reg_dnn)
        self.shared_gate_dnn_final_layer = nn.ModuleList(
            [nn.Linear(
                gate_dnn_hidden_units[-1] if len(gate_dnn_hidden_units) > 0 else self.input_dim if level_num == 0 else
                expert_dnn_hidden_units[-1], shared_gate_output_dim, bias=False)
                for level_num in range(self.num_levels)])

        # 3. tower dnn (task-specific)
        if len(tower_dnn_hidden_units) > 0:
            self.tower_dnn = nn.ModuleList(
                [DNN(expert_dnn_hidden_units[-1], tower_dnn_hidden_units, activation=dnn_activation,
                     l2_reg=l2_reg_dnn, dropout_rate=dnn_dropout, use_bn=dnn_use_bn,
                     init_std=init_std, device=device) for _ in range(self.num_tasks)])
            self.add_regularization_weight(
                filter(lambda x: 'weight' in x[0] and 'bn' not in x[0], self.tower_dnn.named_parameters()),
                l2=l2_reg_dnn)
        self.tower_dnn_final_layer = nn.ModuleList([nn.Linear(
            tower_dnn_hidden_units[-1] if len(tower_dnn_hidden_units) > 0 else expert_dnn_hidden_units[-1], 1,
            bias=False)
            for _ in range(self.num_tasks)])

        self.out = nn.ModuleList([PredictionLayer(task) for task in task_types])

        regularization_modules = [self.specific_experts, self.shared_experts, self.specific_gate_dnn_final_layer,
                                  self.shared_gate_dnn_final_layer, self.tower_dnn_final_layer]
        for module in regularization_modules:
            self.add_regularization_weight(
                filter(lambda x: 'weight' in x[0] and 'bn' not in x[0], module.named_parameters()), l2=l2_reg_dnn)
        self.to(device)

    # a single cgc Layer
    def cgc_net(self, inputs, level_num):
        # inputs: [task1, task2, ... taskn, shared task]

        # 1. experts
        # task-specific experts
        specific_expert_outputs = []
        for i in range(self.num_tasks):
            for j in range(self.specific_expert_num):
                specific_expert_output = self.specific_experts[level_num][i][j](inputs[i])
                specific_expert_outputs.append(specific_expert_output)

        # shared experts
        shared_expert_outputs = []
        for k in range(self.shared_expert_num):
            shared_expert_output = self.shared_experts[level_num][0][k](inputs[-1])
            shared_expert_outputs.append(shared_expert_output)

        # 2. gates
        # gates for task-specific experts
        cgc_outs = []
        for i in range(self.num_tasks):
            # concat task-specific expert and task-shared expert
            cur_experts_outputs = specific_expert_outputs[
                                  i * self.specific_expert_num:(i + 1) * self.specific_expert_num] + shared_expert_outputs
            cur_experts_outputs = torch.stack(cur_experts_outputs, 1)

            # gate dnn
            if len(self.gate_dnn_hidden_units) > 0:
                gate_dnn_out = self.specific_gate_dnn[level_num][i][0](inputs[i])
                gate_dnn_out = self.specific_gate_dnn_final_layer[level_num][i](gate_dnn_out)
            else:
                gate_dnn_out = self.specific_gate_dnn_final_layer[level_num][i](inputs[i])
            gate_mul_expert = torch.matmul(gate_dnn_out.softmax(1).unsqueeze(1), cur_experts_outputs)  # (bs, 1, dim)
            cgc_outs.append(gate_mul_expert.squeeze())

        # gates for shared experts
        cur_experts_outputs = specific_expert_outputs + shared_expert_outputs
        cur_experts_outputs = torch.stack(cur_experts_outputs, 1)

        if len(self.gate_dnn_hidden_units) > 0:
            gate_dnn_out = self.shared_gate_dnn[level_num](inputs[-1])
            gate_dnn_out = self.shared_gate_dnn_final_layer[level_num](gate_dnn_out)
        else:
            gate_dnn_out = self.shared_gate_dnn_final_layer[level_num](inputs[-1])
        gate_mul_expert = torch.matmul(gate_dnn_out.softmax(1).unsqueeze(1), cur_experts_outputs)  # (bs, 1, dim)
        cgc_outs.append(gate_mul_expert.squeeze())

        return cgc_outs

    def forward(self, X):
        sparse_embedding_list, dense_value_list = self.input_from_feature_columns(X, self.dnn_feature_columns,
                                                                                  self.embedding_dict)
        dnn_input = combined_dnn_input(sparse_embedding_list, dense_value_list)

        # repeat `dnn_input` for several times to generate cgc input
        ple_inputs = [dnn_input] * (self.num_tasks + 1)  # [task1, task2, ... taskn, shared task]
        ple_outputs = []
        for i in range(self.num_levels):
            ple_outputs = self.cgc_net(inputs=ple_inputs, level_num=i)
            ple_inputs = ple_outputs

        # tower dnn (task-specific)
        task_outs = []
        for i in range(self.num_tasks):
            if len(self.tower_dnn_hidden_units) > 0:
                tower_dnn_out = self.tower_dnn[i](ple_outputs[i])
                tower_dnn_logit = self.tower_dnn_final_layer[i](tower_dnn_out)
            else:
                tower_dnn_logit = self.tower_dnn_final_layer[i](ple_outputs[i])
            output = self.out[i](tower_dnn_logit)
            task_outs.append(output)
        task_outs = torch.cat(task_outs, -1)
        return task_outs