PyTorch 自定义特征嵌入模块 FeatureEmbedder 实现解析

PyTorch FeatureEmbedder 类实现解析

概述

这段 Python 代码定义了一个名为 FeatureEmbedder 的类，它继承自 PyTorch 的 nn.Module。这个类是一个自定义的神经网络模块，用于从电子健康记录（EHR）中嵌入不同类型的特征。

代码实现

import math
import torch
from torch import nn
from utils import get_extended_attention_mask

class FeatureEmbedder(nn.Module):
    def __init__(self, args):
        super().__init__()
        self.embeddings = {}
        self.feature_keys = args.feature_keys
        self.dx_embeddings = nn.Embedding(args.vocab_sizes['dx_ints']+1, args.hidden_size, padding_idx=args.vocab_sizes['dx_ints'])
        self.proc_embeddings = nn.Embedding(args.vocab_sizes['proc_ints']+1, args.hidden_size, padding_idx=args.vocab_sizes['proc_ints'])
        self.visit_embeddings = nn.Embedding(1, args.hidden_size)
        self.layernorm = nn.LayerNorm(args.hidden_size)
        self.dropout = nn.Dropout(args.hidden_dropout_prob)

    def forward(self, features):
        batch_size = features[self.feature_keys[0]].shape[0]
        embeddings = {}
        masks = {}
        embeddings['dx_ints'] = self.dx_embeddings(features['dx_ints'])
        embeddings[] = .proc_embeddings(features[])
        device = features[].device
        embeddings[] = .visit_embeddings(torch.tensor([]).to(device))
        embeddings[] = embeddings[].unsqueeze().expand(batch_size,-,-)
        masks[] = torch.ones(batch_size,).to(device)
         name, embedding  embeddings.items():
            embeddings[name] = .layernorm(embedding)
            embeddings[name] = .dropout(embeddings[name])
         embeddings, masks