Embedding的两种实现方式torch代码

详细请看：

https://mp.weixin.qq.com/s?__biz=Mzk0MzIzODM5MA==&mid=2247484496&idx=1&sn=d55e638447726b2a527c347ed8735e92&chksm=c337b9c9f44030df48cff5bd6d2e674fc8f466412957367ed32676b28fbdd5883bfef743dbd2#rd

以DNN为例全部代码：

from sklearn.model_selection import train_test_split
from tqdm import tqdm
from itertools import combinations
import torch
from torch import nn, optim
from torch.utils.data import Dataset, DataLoader, TensorDataset
# import torch.utils.data as Data
from torchkeras import summary

from tools import *

''' 用于 spare field embedding '''
def sparseFeature(feat, vocabulary_size, embed_dim):
    return {
    'spare': feat, 'vocabulary_size': vocabulary_size, 'embed_dim': embed_dim}

''' 用于 dense field embedding '''
def denseFeature(feat):
    return {
    'dense': feat}

class NN_emb_offsets(nn.Module):
    def __init__(self, spare_feature_columns, dense_feature_columns, feature_fields,
                 emb_dim=10, dnn_hidden_units=(64, 32), dropout=0.5):
        super(NN_emb_offsets, self).__init__()
        self.spare_feature_columns = spare_feature_columns
        self.dense_feature_columns = dense_feature_columns
        self.spare_field_num = len(spare_feature_columns)
        self.dense_field_num = len(dense_feature_columns)

        self.offsets = np.array((0, *np.cumsum(feature_fields)[:-1]), dtype=np.long)
        # self.offsets = nn.Parameter(torch.tensor([0] + feature_fields[:-1]).cumsum(0), requires_grad=False)

        # Embedding layer
        self.embedding = nn.Embedding(sum(feature_fields) + 1, embedding_dim=emb_dim)
        torch.nn.init.xavier_uniform_(self.embedding.weight.data)

        # DNN layer
        self.dnn_hidden_units = dnn_hidden_units
        dnn_layers = []
        input_dim = self.dense_field_num + self.spare_field_num * emb_dim
        for hidden in dnn_hidden_units:
            dnn_layers.append(nn.Linear(input_dim, hidden))
            dnn_layers.append(nn.BatchNorm1d(hidden))
            dnn_layers.append(nn.ReLU())
            dnn_layers.append(nn.Dropout(p=dropout))
            input_dim = hidden
        dnn_layers.append(nn.Linear(input_dim, 1))
        self.DNN = nn.Sequential(*dnn_layers)

    def forward(self, inputs):

        dense_inputs, sparse_inputs = inputs[:, :13], inputs[:, 13:]

        sparse_inputs = sparse_inputs.long()

        sparse_inputs = sparse_inputs + sparse_inputs.new_tensor(self.offsets).unsqueeze(0)

        sparse_embed = self.embedding(sparse_inputs)  # (None, field_num, emb_dim)

        sparse_embed = sparse_embed.view(-1, 26 * 8)

        x = torch.cat([dense_inputs, sparse_embed], dim=1)  # (batchsize, 26*embed_dim + 13)

        dnn_out = self.DNN(x)  # (None, 1)

        return torch.sigmoid(dnn_out).squeeze(dim=-1)  # (batchsize,)

class NN_emb(nn.Module):
    def __init__(self, spare_feature_columns, dense_feature_columns, hidden_units=(128, 64, 32), emb_dim=10, droup_out=0.):
        super(NN_emb, self).__init__()
        self.spare_feature_columns = spare_feature_columns
        self.dense_feature_columns = dense_feature_columns
        self.spare_field_num = len(spare_feature_columns)
        self.dense_field_num = len(dense_feature_columns)

        # Embedding
        self.embedding_layer = nn.ModuleDict({
    'embed_layer{}'.format(i): nn.Embedding(feat['vocabulary_size'], feat['embed_dim'])
                                              for i, feat in enumerate(self.spare_feature_columns)})

        for i in range(self.spare_field_num):
            torch.nn.init.xavier_uniform_(self.embedding_layer['embed_layer{}'.format(i)].weight.data)


        fc_layers = []
        input_dim = self.dense_field_num + self.spare_field_num * emb_dim
        for fc_dim in hidden_units:
            fc_layers.append(nn.Linear(input_dim, fc_dim))
            fc_layers.append(nn.BatchNorm1d(fc_dim))
            fc_layers.append(nn.ReLU())
            fc_layers.append(nn.Dropout(p=droup_out))
            input_dim = fc_dim
        fc_layers.append(nn.Linear(input_dim, 1))
        self.DNN = nn.Sequential(*fc_layers)

    def forward(self, inputs):
        # dense_inputs: 数值特征，13维
        # sparse_inputs： 类别特征，26维 28?
        dense_inputs, sparse_inputs = inputs[:, :13], inputs[:, 13:]

        # embedding
        sparse_inputs = sparse_inputs.long()  # 转成long类型才能作为nn.embedding的输入
        sparse_embed = torch.cat([self.embedding_layer['embed_layer{}'.format(i)](sparse_inputs[:, i])
                                  for i in range(sparse_inputs.shape[1])], dim=1)  # (batchsize, 26*embed_dim)

        x = torch.cat([dense_inputs, sparse_embed], dim=1)  # (batchsize, 26*embed_dim + 13)

        # DNN Layer
        dnn_out = self.DNN(x)  # (batchsize, 1)

        return torch.sigmoid(dnn_out).squeeze(dim=-1)  # (batchsize,)

if __name__ == '__main__':

    data = pd.read_csv('./data/criteo_sampled_data_test.csv')

    stat_pnrate_pd(data=data, labname='label', message='criteo_sampled_data_test')

    # I1-I13：总共 13 列数值型特征
    # C1-C26：共有 26 列类别型特征
    dense_cols = ['I' + str(i) for i in range(1, 14)]
    sparse_cols = ['C' + str(i) for i in range(1, 27)]

    data_X = data[dense_cols + sparse_cols]
    data_y = data['label']

    tmp_X, test_X, tmp_y, test_y = train_test_split(data_X, data_y, test_size=0.01, random_state=42, stratify=data_y)
    train_X, val_X, train_y, val_y = train_test_split(tmp_X, tmp_y, test_size=0.01, random_state=42, stratify=tmp_y)

    train_set = TensorDataset(torch.tensor(train_X.values).float(), torch.tensor(train_y.values).float())
    val_set = TensorDataset(torch.tensor(val_X.values).float(), torch.tensor(val_y.values).float())

    train_loader = DataLoader(dataset=train_set, batch_size=2048, shuffle=True)
    val_loader = DataLoader(dataset=val_set, batch_size=2048, shuffle=False)


    emb_dim = 8
    dense_feature_columns = [denseFeature(feat) for feat in dense_cols]
    # spare_feature_columns = [sparseFeature(feat, data_X[feat].nunique(), emb_dim) for feat in sparse_cols]
    spare_feature_columns = [sparseFeature(feat, data_X[feat].max() + 1, emb_dim) for feat in sparse_cols]
    print(len(dense_feature_columns), dense_feature_columns)
    print(len(spare_feature_columns), spare_feature_columns)

    feature_fields = data_X[sparse_cols].max().values + 1
    feature_fields = feature_fields.astype(np.int)
    print(len(feature_fields), feature_fields)

    # net = NN_emb(spare_feature_columns=spare_feature_columns, dense_feature_columns=dense_feature_columns, emb_dim=emb_dim)
    net = NN_emb_offsets(spare_feature_columns=spare_feature_columns, dense_feature_columns=dense_feature_columns, feature_fields=feature_fields, emb_dim=emb_dim)

    loss_function = nn.BCELoss()
    optimizer = optim.Adam(net.parameters(), lr=0.001)
    fit(net, train_loader=train_loader, val_loader=val_loader,
        epochs=2, loss_function=loss_function, optimizer=optimizer, metric_name=['accuracy', 'auc'])

tf、torch实现embedding
https://wangguisen.blog.csdn.net/article/details/122697991