72 lines
2.8 KiB
Python
72 lines
2.8 KiB
Python
from tensorflow.keras.models import Model
|
||
from tensorflow.keras.layers import Flatten, Concatenate
|
||
from features import FeatureEncoder
|
||
from layers import DNN, CosinSimilarity, PredictLayer
|
||
|
||
|
||
def process_feature(user_feature_columns, item_feature_columns, feature_encode):
|
||
"""
|
||
根据FeatureEncoder获取所有输入的Input层或者Embedding层，然后根据自己
|
||
实际场景的业务数据，对不同的特征进行处理.
|
||
user_feature_columns = [
|
||
SparseFeat('user_id', feature_max_index_dict['user_id'], embedding_dim=4),
|
||
SparseFeat('gender', feature_max_index_dict['gender'], embedding_dim=4),
|
||
SparseFeat('occupation', feature_max_index_dict['occupation'], embedding_dim=4),
|
||
SparseFeat('zip', feature_max_index_dict['zip'], embedding_dim=4),
|
||
VarLenSparseFeat(SparseFeat('hist_movie_id', feature_max_idx['movie_id'], embedding_dim,
|
||
embedding_name="movie_id"), SEQ_LEN, 'mean', 'hist_len'),
|
||
]
|
||
item_feature_columns = [SparseFeat('movie_id', feature_max_index_dict['movie_id'], embedding_dim=4)]
|
||
"""
|
||
group_embedding_dict = feature_encode.sparse_feature_dict
|
||
|
||
user_emb_name = [fc.embedding_name for fc in user_feature_columns]
|
||
item_emb_name = [fc.embedding_name for fc in item_feature_columns]
|
||
|
||
user_dnn_input = [v for k, v in group_embedding_dict['default_group'].items()
|
||
if k in user_emb_name]
|
||
item_dnn_input = [v for k, v in group_embedding_dict['default_group'].items()
|
||
if k in item_emb_name]
|
||
|
||
return user_dnn_input, item_dnn_input
|
||
|
||
|
||
def DSSM(user_feature_columns, item_feature_columns, dnn_units=[64, 32],
|
||
temp=10, task='binary'):
|
||
# 构建所有特征的Input层和Embedding层
|
||
feature_encode = FeatureEncoder(user_feature_columns + item_feature_columns)
|
||
feature_input_layers_list = list(feature_encode.feature_input_layer_dict.values())
|
||
|
||
# 特征处理
|
||
user_dnn_input, item_dnn_input = process_feature(user_feature_columns,\
|
||
item_feature_columns, feature_encode)
|
||
|
||
# 构建模型的核心层
|
||
if len(user_dnn_input) >= 2:
|
||
user_dnn_input = Concatenate(axis=1)(user_dnn_input)
|
||
else:
|
||
user_dnn_input = user_dnn_input[0]
|
||
if len(item_dnn_input) >= 2:
|
||
item_dnn_input = Concatenate(axis=1)(item_dnn_input)
|
||
else:
|
||
item_dnn_input = item_dnn_input[0]
|
||
|
||
user_dnn_input = Flatten()(user_dnn_input)
|
||
item_dnn_input = Flatten()(item_dnn_input)
|
||
|
||
user_dnn_out = DNN(dnn_units)(user_dnn_input)
|
||
item_dnn_out = DNN(dnn_units)(item_dnn_input)
|
||
|
||
# 计算相似度
|
||
scores = CosinSimilarity(temp)([user_dnn_out, item_dnn_out]) # (B,1)
|
||
|
||
# 确定拟合目标
|
||
output = PredictLayer()(scores)
|
||
|
||
# 根据输入输出构建模型
|
||
model = Model(feature_input_layers_list, output)
|
||
# model.summary()
|
||
|
||
return model
|
||
|