85 lines
3.5 KiB
Python
85 lines
3.5 KiB
Python
from audioop import bias
|
||
from itertools import chain
|
||
import tensorflow as tf
|
||
from tensorflow.keras.models import Model
|
||
from tensorflow.keras.layers import Flatten, Concatenate, Dense, Reshape
|
||
from features import FeatureEncoder
|
||
from layers import DNN, InnerProduct, OuterProduct
|
||
from layers.core import PredictLayer
|
||
|
||
|
||
def process_feature(linear_feature_columns, dnn_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.
|
||
"""
|
||
linear_input_sparse_dict = feature_encode.linear_sparse_feature_dict
|
||
group_embedding_dict = feature_encode.sparse_feature_dict
|
||
|
||
linear_input_sparse_list = list(linear_input_sparse_dict['default_group'].values())
|
||
linear_dense_dict = feature_encode.dense_feature_dict
|
||
linear_dense_list = list(linear_dense_dict.values())
|
||
|
||
dnn_emb_name = [fc.embedding_name for fc in dnn_feature_columns]
|
||
embedding_list = [v for k, v in group_embedding_dict['pnn'].items()
|
||
if k in dnn_emb_name]
|
||
|
||
linear_sparse_inputs = Flatten()(Concatenate(axis=1)(linear_input_sparse_list))
|
||
linear_dense_list.append(linear_sparse_inputs)
|
||
linear_inputs = []
|
||
if len(linear_dense_list) == 2:
|
||
linear_inputs = Concatenate(axis=-1)(linear_dense_list)
|
||
else:
|
||
linear_inputs = linear_dense_list[0]
|
||
|
||
return linear_inputs, embedding_list
|
||
|
||
|
||
def PNN(linear_feature_columns, dnn_feature_columns, hidden_units=(32, 16, 1),
|
||
use_inner=True, inner_units=512, outer_units=512, use_outter=True,
|
||
use_mat_dot=False, activation='relu', use_bias=True, dp_rate=0.2, use_bn=True,
|
||
use_dp=True, task='binary'):
|
||
|
||
feature_columns = list(set(linear_feature_columns + dnn_feature_columns))
|
||
feature_encode = FeatureEncoder(feature_columns,
|
||
linear_sparse_feature=linear_feature_columns)
|
||
feature_input_layers_list = list(feature_encode.feature_input_layer_dict.values())
|
||
|
||
# 特征处理
|
||
linear_inputs, embedding_list = \
|
||
process_feature(linear_feature_columns, dnn_feature_columns, feature_encode)
|
||
|
||
# 这里本来应该对原始的embedding做矩阵内积的，但是考虑到参数量比较多，下面直接将embedding矩阵展开
|
||
field_size = len(embedding_list)
|
||
embedding_dim = embedding_list[0].shape[-1]
|
||
embedding_matrix = Concatenate(axis=1)(embedding_list)
|
||
# (None, field_size * embedding_size)
|
||
dnn_linear_input = Reshape([field_size * embedding_dim])(embedding_matrix)
|
||
|
||
dnn_interaction_list = []
|
||
if use_inner:
|
||
dnn_inputs = InnerProduct(use_mat_dot, inner_units)(embedding_list)
|
||
dnn_interaction_list.append(dnn_inputs)
|
||
if use_outter:
|
||
dnn_inputs = OuterProduct(use_mat_dot, outer_units)(embedding_list)
|
||
dnn_interaction_list.append(dnn_inputs)
|
||
|
||
if len(dnn_interaction_list) == 2:
|
||
dnn_interaction_input = Concatenate(axis=-1)(dnn_interaction_list)
|
||
else:
|
||
dnn_interaction_input = dnn_interaction_list[0]
|
||
|
||
if linear_inputs is not None:
|
||
dnn_input_list = [linear_inputs, dnn_linear_input, dnn_interaction_input]
|
||
else:
|
||
dnn_input_list = [dnn_linear_input, dnn_interaction_input]
|
||
|
||
dnn_inputs = Flatten()(Concatenate(axis=-1)(dnn_input_list))
|
||
|
||
logits = DNN(hidden_units, activation, use_bias, use_dp, dp_rate, dp_rate,
|
||
use_bn, get_logits=True)(dnn_inputs)
|
||
|
||
output = PredictLayer(task)(logits)
|
||
model = Model(feature_input_layers_list, output)
|
||
return model
|