def get_mov_combined_features():
MOV_DICT_SIZE = paddle.dataset.movielens.max_movie_id() + 1
mov_id = layers.data(name='movie_id', shape=[1], dtype='int64')
mov_emb = layers.embedding(input=mov_id,
dtype='float32',
size=[MOV_DICT_SIZE, 32],
param_attr='movie_table',
is_sparse=IS_SPARSE)
mov_fc = layers.fc(input=mov_emb, size=32)
CATEGORY_DICT_SIZE = len(paddle.dataset.movielens.movie_categories())
category_id = layers.data(name='category_id',
shape=[1],
dtype='int64',
lod_level=1)
mov_categories_emb = layers.embedding(input=category_id,
size=[CATEGORY_DICT_SIZE, 32],
is_sparse=IS_SPARSE)
mov_categories_hidden = layers.sequence_pool(input=mov_categories_emb,
pool_type="sum")
MOV_TITLE_DICT_SIZE = len(paddle.dataset.movielens.get_movie_title_dict())
mov_title_id = layers.data(name='movie_title',
shape=[1],
dtype='int64',
lod_level=1)
mov_title_emb = layers.embedding(input=mov_title_id,
size=[MOV_TITLE_DICT_SIZE, 32],
is_sparse=IS_SPARSE)
# 电影标题名称(title)是一个序列的整数，整数代表的是这个词在索引序列中的下标。
# 这个序列会被送入 sequence_conv_pool 层，这个层会在时间维度上使用卷积和池化。
# 因为如此，所以输出会是固定长度，尽管输入的序列长度各不相同。
mov_title_conv = nets.sequence_conv_pool(input=mov_title_emb,
num_filters=32,
filter_size=3,
act="tanh",
pool_type="sum")
concat_embed = layers.concat(
input=[mov_fc, mov_categories_hidden, mov_title_conv], axis=1)
mov_combined_features = layers.fc(input=concat_embed, size=200, act="tanh")
return mov_combined_features
def textcnn(token_ids, vocab_size, num_classes, emb_dim, num_filters,
mlp_hid_dim):
"""TextCNN模型前向过程实现
Args:
token_ids: 包含不同长度样本的lod tensor，形状为[-1,1]
vocab_size: 词典大小
num_classes: 类别数量
emb_dim: 词向量维度
num_filters: 每种尺寸的卷积核数量
mlm_hid_dim: MLP的隐层维度
Returns:
prediction: 预测结果，各个类别的概率分布
"""
emb = layers.embedding( # 得到输入样本的词向量表示
input=token_ids, size=[vocab_size, emb_dim])
res_size3 = nets.sequence_conv_pool( # 尺寸为3的卷积层&池化操作
input=emb,
num_filters=num_filters,
filter_size=3,
act="tanh",
pool_type="max")
res_size4 = nets.sequence_conv_pool( # 尺寸为4的卷积层&池化操作
input=emb,
num_filters=num_filters,
filter_size=4,
act="tanh",
pool_type="max")
res_size5 = nets.sequence_conv_pool( # 尺寸为5的卷积层&池化操作
input=emb,
num_filters=num_filters,
filter_size=5,
act="tanh",
pool_type="max")
hidden = layers.fc( # 特征向量到MLP隐层的映射
input=[res_size3, res_size4, res_size4],
size=mlp_hid_dim)
prediction = fluid.layers.fc( # MLP隐层到类别
input=hidden, size=num_classes, act="softmax")
return prediction
def get_mov_combined_features():
MOV_DICT_SIZE = paddle.dataset.movielens.max_movie_id() + 1
mov_id = layers.data(name='movie_id', shape=[1], dtype='int64')
mov_emb = layers.embedding(input=mov_id,
dtype='float32',
size=[MOV_DICT_SIZE, 32],
param_attr='movie_table',
is_sparse=IS_SPARSE)
mov_fc = layers.fc(input=mov_emb, size=32)
CATEGORY_DICT_SIZE = len(paddle.dataset.movielens.movie_categories())
category_id = layers.data(name='category_id',
shape=[1],
dtype='int64',
lod_level=1)
mov_categories_emb = layers.embedding(input=category_id,
size=[CATEGORY_DICT_SIZE, 32],
is_sparse=IS_SPARSE)
mov_categories_hidden = layers.sequence_pool(input=mov_categories_emb,
pool_type="sum")
MOV_TITLE_DICT_SIZE = len(paddle.dataset.movielens.get_movie_title_dict())
mov_title_id = layers.data(name='movie_title',
shape=[1],
dtype='int64',
lod_level=1)
mov_title_emb = layers.embedding(input=mov_title_id,
size=[MOV_TITLE_DICT_SIZE, 32],
is_sparse=IS_SPARSE)
mov_title_conv = nets.sequence_conv_pool(input=mov_title_emb,
num_filters=32,
filter_size=3,
act="tanh",
pool_type="sum")
concat_embed = layers.concat(
input=[mov_fc, mov_categories_hidden, mov_title_conv], axis=1)
# FIXME(dzh) : need tanh operator
mov_combined_features = layers.fc(input=concat_embed, size=200, act="tanh")
return mov_combined_features
def get_mov_combined_features():
MOV_DICT_SIZE = paddle.dataset.movielens.max_movie_id() + 1
mov_id = layers.data(name='movie_id', shape=[1], dtype='int64')
mov_emb = layers.embedding(
input=mov_id,
dtype='float32',
size=[MOV_DICT_SIZE, 32],
param_attr='movie_table',
is_sparse=IS_SPARSE)
mov_fc = layers.fc(input=mov_emb, size=32)
CATEGORY_DICT_SIZE = len(paddle.dataset.movielens.movie_categories())
category_id = layers.data(
name='category_id', shape=[1], dtype='int64', lod_level=1)
mov_categories_emb = layers.embedding(
input=category_id, size=[CATEGORY_DICT_SIZE, 32], is_sparse=IS_SPARSE)
mov_categories_hidden = layers.sequence_pool(
input=mov_categories_emb, pool_type="sum")
MOV_TITLE_DICT_SIZE = len(paddle.dataset.movielens.get_movie_title_dict())
mov_title_id = layers.data(
name='movie_title', shape=[1], dtype='int64', lod_level=1)
mov_title_emb = layers.embedding(
input=mov_title_id, size=[MOV_TITLE_DICT_SIZE, 32], is_sparse=IS_SPARSE)
mov_title_conv = nets.sequence_conv_pool(
input=mov_title_emb,
num_filters=32,
filter_size=3,
act="tanh",
pool_type="sum")
concat_embed = layers.concat(
input=[mov_fc, mov_categories_hidden, mov_title_conv], axis=1)
# FIXME(dzh) : need tanh operator
mov_combined_features = layers.fc(input=concat_embed, size=200, act="tanh")
return mov_combined_features
def model():
"""model"""
user_phone_brand_id = layers.data(name='user_phone_brand', shape=[1], dtype='int64')
user_gender_id = layers.data(name='user_gender', shape=[1], dtype='int64')
user_age_id = layers.data(name='user_age', shape=[1], dtype='int64')
user_status_id = layers.data(name='user_status', shape=[1], dtype="int64")
user_trade_id = fluid.layers.data(name='user_trade', shape=[1], dtype='int64')
user_cater_id = fluid.layers.data(name='user_cater', shape=[1], dtype='int64')
user_income_id = fluid.layers.data(name='user_income', shape=[1], dtype='int64')
user_city_id = fluid.layers.data(name='user_city', shape=[1], dtype='int64')
user_click_id = fluid.layers.data(name='user_click', shape=[1], dtype='int64')
user_b_click_id = fluid.layers.data(name='user_b_click', shape=[1], dtype='int64')
user_c_click_id = fluid.layers.data(name='user_c_click', shape=[1], dtype='int64')
user_d_click_id = fluid.layers.data(name='user_d_click', shape=[1], dtype='int64')
week_id = layers.data(name='week', shape=[1], dtype="int64")
hour_id = layers.data(name='hour', shape=[1], dtype='int64')
content_b_c_d_id = layers.data(name='content_b_c_d', shape=[1], dtype='int64')
content_tags_id = layers.data(name='content_tags', shape=[1], dtype='int64', lod_level=1)
content_subtags_id = layers.data(name='content_subtags', shape=[1], dtype='int64', lod_level=1)
user_content_tag_click_id = layers.data(name='user_content_tag_click', shape=[1], dtype='int64')
user_content_subtag_click_id = layers.data(name='user_content_subtag_click', shape=[1], dtype='int64')
content_pctr_discrete_id = layers.data(name='content_pctr_discrete', shape=[1], dtype='int64')
# dnn_score_discrete_id = layers.data(name='dnn_score_discrete', shape=[1], dtype='int64')
content_pctr = layers.data(name='content_pctr', shape=[1], dtype='float32')
# dnn_score = layers.data(name='dnn_score', shape=[1], dtype='float32')
# content_emb = layers.data(name='content_emb', shape=[64], dtype='float32')
# user_emb = layers.data(name='user_emb', shape=[64], dtype='float32')
user_click_tags_id = layers.data(
name='user_click_tags_id', shape=[1], dtype='int64', lod_level=1)
user_click_subtags_id = layers.data(
name='user_click_subtags_id', shape=[1], dtype='int64', lod_level=1)
candidate_title_word = layers.data(name='candidate_title', shape=[1], dtype='int64', lod_level=1)
candidate_subtitle_word = layers.data(name='candidate_subtitle', shape=[1], dtype='int64', lod_level=1)
candidate_title_len_id = layers.data(name='candidate_title_len', shape=[1], dtype='int64')
candidate_subtitle_len_id = layers.data(name='candidate_subtitle_len', shape=[1], dtype='int64')
click_title_list = layers.data(name='click_title_list', shape=[1], dtype='int64', lod_level=2)
click_subtitle_list = layers.data(name='click_subtitle_list', shape=[1], dtype='int64', lod_level=2)
click_title_len_list = layers.data(name='click_title_len_list', shape=[1], dtype='int64', lod_level=1)
click_subtitle_len_list = layers.data(name='click_subtitle_len_list', shape=[1], dtype='int64', lod_level=1)
label = layers.data(name='label', shape=[1], dtype='int64')
# dnn_score_discrete_id.name, dnn_score.name, content_emb.name,user_emb.name,
load_list = [user_phone_brand_id, user_gender_id, user_age_id,
user_status_id, user_trade_id, user_cater_id, user_income_id,
user_city_id, user_click_id, user_b_click_id, user_c_click_id,
user_d_click_id, week_id, hour_id, content_b_c_d_id,
content_tags_id, content_subtags_id, user_content_tag_click_id,
user_content_subtag_click_id, content_pctr_discrete_id,
content_pctr,
user_click_tags_id, user_click_subtags_id, candidate_title_word,
candidate_subtitle_word, candidate_title_len_id, candidate_subtitle_len_id,
click_title_list, click_subtitle_list,
click_title_len_list, click_subtitle_len_list,
label]
feed_order = [x.name for x in load_list]
user_phone_brand_emb = layers.embedding(
input=user_phone_brand_id, dtype='float32',
size=[7, EMB_LEN], param_attr='user_phone_brand_emb', is_sparse=True)
user_gender_emb = layers.embedding(
input=user_gender_id, dtype='float32',
size=[3, EMB_LEN], param_attr='user_gender_emb', is_sparse=True)
user_age_emb = layers.embedding(
input=user_age_id, dtype='float32',
size=[8, EMB_LEN], param_attr='user_age_emb', is_sparse=True)
user_status_emb = layers.embedding(
input=user_status_id, dtype='float32',
size=[3, EMB_LEN], is_sparse=True, param_attr='user_status_emb')
user_trade_emb = layers.embedding(
input=user_trade_id, dtype='float32',
size=[24, EMB_LEN], is_sparse=True, param_attr='user_trade_emb')
user_cater_emb = layers.embedding(
input=user_cater_id, dtype='float32',
size=[4, EMB_LEN], is_sparse=True, param_attr='user_cater_emb')
user_income_emb = layers.embedding(
input=user_income_id, dtype='float32',
size=[6, EMB_LEN], is_sparse=True, param_attr='user_income_emb')
user_city_emb = layers.embedding(
input=user_city_id, dtype='float32',
size=[4000, EMB_LEN], is_sparse=True, param_attr='user_city_emb')
user_click_emb = layers.embedding(
input=user_click_id, dtype='float32',
size=[6, EMB_LEN], is_sparse=True, param_attr='user_click_emb')
user_b_click_emb = layers.embedding(
input=user_b_click_id, dtype='float32',
size=[6, EMB_LEN], is_sparse=True, param_attr='user_b_click_emb')
user_c_click_emb = layers.embedding(
input=user_c_click_id, dtype='float32',
size=[6, EMB_LEN], is_sparse=True, param_attr='user_c_click_emb')
user_d_click_emb = layers.embedding(
input=user_d_click_id, dtype='float32',
size=[6, EMB_LEN], is_sparse=True, param_attr='user_d_click_emb')
week_emb = layers.embedding(
input=week_id, dtype='float32',
size=[8, EMB_LEN], is_sparse=True, param_attr='week_emb')
hour_emb = layers.embedding(
input=hour_id, dtype='float32',
size=[24, EMB_LEN], is_sparse=True, param_attr='hour_emb')
content_b_c_d_emb = layers.embedding(
input=content_b_c_d_id, dtype='float32',
size=[3, EMB_LEN], is_sparse=True, param_attr='content_b_c_d_emb')
content_tags_emb = layers.embedding(
input=content_tags_id, size=[11, EMB_LEN], dtype='float32', is_sparse=True,
param_attr=fluid.ParamAttr(
name="content_tags_emb", learning_rate=0.5, regularizer=fluid.regularizer.L2Decay(1.0))
)
content_tags_emb_avg = fluid.layers.sequence_pool(input=content_tags_emb, pool_type='average')
content_subtags_emb = layers.embedding(
input=content_subtags_id, size=[65, EMB_LEN], dtype='float32', is_sparse=True,
param_attr=fluid.ParamAttr(
name="content_subtags_emb", learning_rate=0.5,
regularizer=fluid.regularizer.L2Decay(1.0))
)
content_subtags_emb_avg = fluid.layers.sequence_pool(
input=content_subtags_emb, pool_type='average')
user_content_tag_click_emb = layers.embedding(
input=user_content_tag_click_id, dtype='float32',
size=[11 * 6, EMB_LEN], is_sparse=True, param_attr='user_content_tag_click_emb')
user_content_subtag_click_emb = layers.embedding(
input=user_content_subtag_click_id, dtype='float32',
size=[65 * 6, EMB_LEN], is_sparse=True, param_attr='user_content_subtag_click_emb')
content_pctr_discrete_emb = layers.embedding(
input=content_pctr_discrete_id, dtype='float32',
size=[55, EMB_LEN], is_sparse=True, param_attr='content_pctr_discrete_emb')
# dnn_score_discrete_emb = layers.embedding(
# input=dnn_score_discrete_id, dtype='float32',
# size=[21, EMB_LEN], is_sparse=True, param_attr='dnn_score_discrete_emb')
user_click_tags_id_emb = layers.embedding(
input=user_click_tags_id, size=[11 * 6, EMB_LEN], dtype='float32', is_sparse=True,
param_attr="user_content_tag_click_emb")
user_click_tags_id_emb_avg = fluid.layers.sequence_pool(
input=user_click_tags_id_emb, pool_type='average')
user_click_subtags_id_emb = layers.embedding(
input=user_click_subtags_id, size=[65 * 6, EMB_LEN], dtype='float32', is_sparse=True,
param_attr="user_content_subtag_click_emb")
user_click_subtags_id_emb_avg = fluid.layers.sequence_pool(
input=user_click_subtags_id_emb, pool_type='average')
# 候选内容feature生成
cand_title_emb = layers.embedding(input=candidate_title_word, size=[19962, EMB_LEN], dtype='float32',
is_sparse=False, param_attr='word_embedding')
cand_title_conv_pool = nets.sequence_conv_pool(
input=cand_title_emb, num_filters=NUM_FILTERS, filter_size=3,
act="relu", pool_type="average", param_attr='title_emb_conv', bias_attr='title_emb_conv_b')
cand_subtitle_emb = layers.embedding(input=candidate_subtitle_word, size=[19962, EMB_LEN], dtype='float32',
is_sparse=False, param_attr='word_embedding')
cand_subtitle_conv_pool = nets.sequence_conv_pool(
input=cand_subtitle_emb, num_filters=NUM_FILTERS, filter_size=3,
act="relu", pool_type="average", param_attr='subtitle_emb_conv', bias_attr='subtitle_emb_conv_b')
cand_title_len_emb = layers.embedding(input=candidate_title_len_id, size=[100, EMB_LEN], dtype='float32',
is_sparse=True, param_attr='title_len_emb')
cand_subtitle_len_emb = layers.embedding(input=candidate_subtitle_len_id, size=[100, EMB_LEN], dtype='float32',
is_sparse=True, param_attr='subtitle_len_emb')
cand_title_inf = layers.concat(
input=[cand_title_conv_pool, cand_subtitle_conv_pool,
cand_title_len_emb, cand_subtitle_len_emb], axis=-1)
cand_title_feature = layers.fc(
input=cand_title_inf, size=32, act="relu", param_attr='title_feature_list') #共享参数
# 用户历史点击内容feature生成
click_title_emb = layers.embedding(input=click_title_list, size=[19962, EMB_LEN], dtype='float32',
is_sparse=False, param_attr='word_embedding')
click_title_drnn = fluid.layers.DynamicRNN()
with click_title_drnn.block():
title_emb = click_title_drnn.step_input(click_title_emb)
click_title_conv_pool = nets.sequence_conv_pool(
input=title_emb, num_filters=NUM_FILTERS, filter_size=3,
act="relu", pool_type="average", param_attr='title_emb_conv', bias_attr='title_emb_conv_b')
click_title_drnn.output(click_title_conv_pool)
click_title_conv_pool_list = click_title_drnn()
click_subtitle_emb = layers.embedding(input=click_subtitle_list, size=[19962, EMB_LEN], dtype='float32',
is_sparse=False, param_attr='word_embedding')
click_subtitle_drnn = fluid.layers.DynamicRNN()
with click_subtitle_drnn.block():
subtitle_emb = click_subtitle_drnn.step_input(click_subtitle_emb)
click_subtitle_conv_pool = nets.sequence_conv_pool(
input=subtitle_emb, num_filters=NUM_FILTERS, filter_size=3,
act="relu", pool_type="average", param_attr='subtitle_emb_conv', bias_attr='subtitle_emb_conv_b')
click_subtitle_drnn.output(click_subtitle_conv_pool)
click_subtitle_conv_pool_list = click_subtitle_drnn()
click_title_len_emb_list = layers.embedding(input=click_title_len_list, size=[100, EMB_LEN], dtype='float32',
is_sparse=True, param_attr='title_len_emb')
click_subtitle_len_emb_list = layers.embedding(input=click_subtitle_len_list, size=[100, EMB_LEN], dtype='float32',
is_sparse=True, param_attr='subtitle_len_emb')
click_title_inf_list = layers.concat(
input=[click_title_conv_pool_list, click_subtitle_conv_pool_list,
click_title_len_emb_list, click_subtitle_len_emb_list], axis=-1)
click_title_feature_list = layers.fc(
input=click_title_inf_list, size=32, act="relu", param_attr='title_feature_list') #共享参数
user_click_title_feature = layers.sequence_pool(input=click_title_feature_list, pool_type="average")
user_emb_feature = layers.concat(
input=[user_phone_brand_emb, user_gender_emb, user_age_emb, user_status_emb, user_trade_emb,
user_cater_emb, user_income_emb, user_city_emb,
user_click_emb, user_b_click_emb, user_c_click_emb, user_d_click_emb], axis=1)
content_emb_feature = layers.concat(
input=[content_b_c_d_emb, content_tags_emb_avg, content_subtags_emb_avg,
content_pctr_discrete_emb, cand_title_feature], axis=1)
cross_emb_feature = layers.concat(
input=[user_content_tag_click_emb, user_content_subtag_click_emb,
user_click_tags_id_emb_avg, user_click_subtags_id_emb_avg,
user_click_title_feature], axis=1)
env_emb_feature = layers.concat(
input=[week_emb, hour_emb], axis=1)
combined_features = layers.concat(input=[
user_emb_feature, content_emb_feature, cross_emb_feature, env_emb_feature], axis=1)
fc1 = layers.fc(input=combined_features, size=200, act='relu', param_attr='fc1', bias_attr='fc1_b')
fc2 = layers.fc(input=fc1, size=200, act="relu", param_attr='fc2', bias_attr='fc2_b')
fc3 = layers.fc(input=fc2, size=200, act="relu", param_attr='fc3', bias_attr='fc3_b')
content_pctr_discrete_id_one_hot = layers.one_hot(
content_pctr_discrete_id, 55, allow_out_of_range=False)
final_layer = layers.concat(input=[fc3, content_pctr, content_pctr_discrete_id_one_hot], axis=1)
predict = layers.fc(
input=final_layer, size=2, act="softmax",
param_attr='final_predict', bias_attr='final_predict_b')
auc = fluid.layers.auc(
input=predict, label=label, num_thresholds=2 ** 12)
cost = layers.cross_entropy(input=predict, label=label)
avg_cost = layers.reduce_mean(cost)
loader = fluid.io.DataLoader.from_generator(
feed_list=load_list, capacity=256, use_double_buffer=True, iterable=True)
return {'predict': predict, 'avg_cost': avg_cost, 'feed_order': feed_order, 'loader': loader, 'auc': auc}