def input_from_feature_columns(features,
feature_columns,
l2_reg,
init_std,
seed,
prefix='',
seq_mask_zero=True,
support_dense=True,
support_group=False,
embedding_matrix_dict=None):
sparse_feature_columns = list(
filter(lambda x: isinstance(x, SparseFeat),
feature_columns)) if feature_columns else []
varlen_sparse_feature_columns = list(
filter(lambda x: isinstance(x, VarLenSparseFeat),
feature_columns)) if feature_columns else []
if embedding_matrix_dict is None:
embedding_matrix_dict = create_embedding_matrix(
feature_columns,
l2_reg,
init_std,
seed,
prefix=prefix,
seq_mask_zero=seq_mask_zero)
group_sparse_embedding_dict = embedding_lookup(embedding_matrix_dict,
features,
sparse_feature_columns)
dense_value_list = get_dense_input(features, feature_columns)
if not support_dense and len(dense_value_list) > 0:
raise ValueError("DenseFeat is not supported in dnn_feature_columns")
sequence_embed_dict = varlen_embedding_lookup(
embedding_matrix_dict, features, varlen_sparse_feature_columns)
group_varlen_sparse_embedding_dict = get_varlen_pooling_list(
sequence_embed_dict, features, varlen_sparse_feature_columns)
group_embedding_dict = mergeDict(group_sparse_embedding_dict,
group_varlen_sparse_embedding_dict)
if not support_group:
group_embedding_dict = list(
chain.from_iterable(group_embedding_dict.values()))
return group_embedding_dict, dense_value_list
def MIND(dnn_feature_columns,
history_feature_list,
target_song_size,
k_max=2,
dnn_use_bn=False,
user_hidden_unit=64,
dnn_activation='relu',
l2_reg_dnn=0,
l2_reg_embedding=1e-6,
dnn_dropout=0,
init_std=0.0001,
seed=1024):
"""
:param dnn_feature_columns: An iterable containing all the features used by deep part of the model.
:param history_feature_list: list,to indicate sequence sparse field
:param target_song_size: int, the total size of the recall songs
:param k_max: int, the max size of user interest embedding
:param dnn_use_bn: bool. Whether use BatchNormalization before activation or not in deep net
:param user_hidden_unit: int. user dnn hidden layer size
:param dnn_activation: Activation function to use in deep net
:param l2_reg_dnn: L2 regularizer strength applied to DNN
:param l2_reg_embedding: float. L2 regularizer strength applied to embedding vector
:param dnn_dropout: float in [0,1), the probability we will drop out a given DNN coordinate.
:param init_std: float,to use as the initialize std of embedding vector
:param seed: integer ,to use as random seed.
:return:
"""
features = build_input_features(dnn_feature_columns)
sparse_feature_columns = list(
filter(lambda x: isinstance(x, SparseFeat),
dnn_feature_columns)) if dnn_feature_columns else []
dense_feature_columns = list(
filter(lambda x: isinstance(x, DenseFeat),
dnn_feature_columns)) if dnn_feature_columns else []
varlen_sparse_feature_columns = list(
filter(lambda x: isinstance(x, VarLenSparseFeat),
dnn_feature_columns)) if dnn_feature_columns else []
history_feature_columns = []
sparse_varlen_feature_columns = []
history_fc_names = list(map(lambda x: "hist_" + x, history_feature_list))
for fc in varlen_sparse_feature_columns:
feature_name = fc.name
if feature_name in history_fc_names:
history_feature_columns.append(fc)
else:
sparse_varlen_feature_columns.append(fc)
hist_len = features['hist_len']
inputs_list = list(features.values())
embedding_dict = create_embedding_matrix(dnn_feature_columns,
l2_reg_embedding,
init_std,
seed,
prefix="")
history_emb_list = embedding_lookup(embedding_dict,
features,
history_feature_columns,
history_fc_names,
history_fc_names,
to_list=True)
history_emb = concat_func(history_emb_list, mask=False)
target_emb_list = embedding_lookup(embedding_dict,
features,
sparse_feature_columns, ['item'],
history_feature_list,
to_list=True)
target_emb_tmp = concat_func(target_emb_list, mask=False)
target_emb_size = target_emb_tmp.get_shape()[-1].value
target_emb = tf.keras.layers.Lambda(
shape_target, arguments={'target_emb_size':
target_emb_size})(target_emb_tmp)
dnn_input_emb_list = embedding_lookup(embedding_dict,
features,
sparse_feature_columns,
mask_feat_list=history_feature_list,
to_list=True)
sequence_embed_dict = varlen_embedding_lookup(
embedding_dict, features, sparse_varlen_feature_columns)
sequence_embed_list = get_varlen_pooling_list(
sequence_embed_dict,
features,
sparse_varlen_feature_columns,
to_list=True)
dnn_input_emb_list += sequence_embed_list
deep_input_emb = concat_func(dnn_input_emb_list)
user_other_feature = Flatten()(deep_input_emb)
max_len = history_emb.get_shape()[1].value
high_capsule = CapsuleLayer(input_units=target_emb_size,
out_units=target_emb_size,
max_len=max_len,
k_max=k_max)((history_emb, hist_len))
other_feature_tile = tf.keras.layers.Lambda(
tile_user_otherfeat, arguments={'k_max': k_max})(user_other_feature)
user_deep_input = Concatenate()(
[NoMask()(other_feature_tile), high_capsule])
user_embeddings = DNN((user_hidden_unit, target_emb_size),
dnn_activation,
l2_reg_dnn,
dnn_dropout,
dnn_use_bn,
seed,
name="user_embedding")(user_deep_input)
k_user = tf.cast(tf.maximum(
1.,
tf.minimum(tf.cast(k_max, dtype="float32"),
tf.log1p(tf.cast(hist_len, dtype="float32")) / tf.log(2.))),
dtype="int64") # [B,1] forword/Cast_2
user_embedding_final = DotProductAttentionLayer(
shape=[target_emb_size, target_emb_size])(
(user_embeddings, target_emb), seq_length=k_user, max_len=k_max)
output = SampledSoftmaxLayer(
target_song_size=target_song_size,
target_emb_size=target_emb_size)(inputs=(user_embedding_final,
features['item']))
model = Model(inputs=inputs_list, outputs=output)
return model
def SDM(user_feature_columns,
item_feature_columns,
history_feature_list,
num_sampled=5,
units=64,
rnn_layers=2,
dropout_rate=0.2,
rnn_num_res=1,
num_head=4,
l2_reg_embedding=1e-6,
dnn_activation='tanh',
init_std=0.0001,
seed=1024):
"""Instantiates the Sequential Deep Matching Model architecture.
:param user_feature_columns: An iterable containing user's features used by the model.
:param item_feature_columns: An iterable containing item's features used by the model.
:param history_feature_list: list,to indicate short and prefer sequence sparse field
:param num_sampled: int, the number of classes to randomly sample per batch.
:param units: int, dimension for each output layer
:param rnn_layers: int, layer number of rnn
:param dropout_rate: float in [0,1), the probability we will drop out a given DNN coordinate.
:param rnn_num_res: int. The number of residual layers in rnn layers
:param num_head: int int, the number of attention head
:param l2_reg_embedding: float. L2 regularizer strength applied to embedding vector
:param dnn_activation: Activation function to use in deep net
:param init_std: float,to use as the initialize std of embedding vector
:param seed: integer ,to use as random seed.
:return: A Keras model instance.
"""
if len(item_feature_columns) > 1:
raise ValueError("Now MIND only support 1 item feature like item_id")
item_feature_column = item_feature_columns[0]
item_feature_name = item_feature_column.name
item_vocabulary_size = item_feature_columns[0].vocabulary_size
features = build_input_features(user_feature_columns)
user_inputs_list = list(features.values())
sparse_feature_columns = list(
filter(lambda x: isinstance(x, SparseFeat),
user_feature_columns)) if user_feature_columns else []
dense_feature_columns = list(
filter(lambda x: isinstance(x, DenseFeat),
user_feature_columns)) if user_feature_columns else []
if len(dense_feature_columns) != 0:
raise ValueError("Now SDM don't support dense feature")
varlen_sparse_feature_columns = list(
filter(lambda x: isinstance(x, VarLenSparseFeat),
user_feature_columns)) if user_feature_columns else []
sparse_varlen_feature_columns = []
prefer_history_columns = []
short_history_columns = []
prefer_fc_names = list(map(lambda x: "prefer_" + x, history_feature_list))
short_fc_names = list(map(lambda x: "short_" + x, history_feature_list))
for fc in varlen_sparse_feature_columns:
feature_name = fc.name
if feature_name in prefer_fc_names:
prefer_history_columns.append(fc)
elif feature_name in short_fc_names:
short_history_columns.append(fc)
else:
sparse_varlen_feature_columns.append(fc)
embedding_matrix_dict = create_embedding_matrix(user_feature_columns +
item_feature_columns,
l2_reg_embedding,
init_std,
seed,
prefix="")
item_features = build_input_features(item_feature_columns)
item_inputs_list = list(item_features.values())
prefer_emb_list = embedding_lookup(embedding_matrix_dict,
features,
prefer_history_columns,
prefer_fc_names,
prefer_fc_names,
to_list=True) # L^u
short_emb_list = embedding_lookup(embedding_matrix_dict,
features,
short_history_columns,
short_fc_names,
short_fc_names,
to_list=True) # S^u
# dense_value_list = get_dense_input(features, dense_feature_columns)
user_emb_list = embedding_lookup(embedding_matrix_dict,
features,
sparse_feature_columns,
to_list=True)
sequence_embed_dict = varlen_embedding_lookup(
embedding_matrix_dict, features, sparse_varlen_feature_columns)
sequence_embed_list = get_varlen_pooling_list(
sequence_embed_dict,
features,
sparse_varlen_feature_columns,
to_list=True)
user_emb_list += sequence_embed_list # e^u
# if len(user_emb_list) > 0 or len(dense_value_list) > 0:
# user_emb_feature = combined_dnn_input(user_emb_list, dense_value_list)
user_emb = concat_func(user_emb_list)
user_emb_output = Dense(units,
activation=dnn_activation,
name="user_emb_output")(user_emb)
prefer_sess_length = features['prefer_sess_length']
prefer_att_outputs = []
for i, prefer_emb in enumerate(prefer_emb_list):
prefer_attention_output = AttentionSequencePoolingLayer(
dropout_rate=0)([user_emb_output, prefer_emb, prefer_sess_length])
prefer_att_outputs.append(prefer_attention_output)
prefer_att_concat = concat_func(prefer_att_outputs)
prefer_output = Dense(units,
activation=dnn_activation,
name="prefer_output")(prefer_att_concat)
short_sess_length = features['short_sess_length']
short_emb_concat = concat_func(short_emb_list)
short_emb_input = Dense(units,
activation=dnn_activation,
name="short_emb_input")(short_emb_concat)
short_rnn_output = DynamicMultiRNN(
num_units=units,
return_sequence=True,
num_layers=rnn_layers,
num_residual_layers=rnn_num_res,
dropout_rate=dropout_rate)([short_emb_input, short_sess_length])
short_att_output = SelfMultiHeadAttention(
num_units=units,
head_num=num_head,
dropout_rate=dropout_rate,
future_binding=True,
use_layer_norm=True)([short_rnn_output, short_sess_length
]) # [batch_size, time, num_units]
short_output = UserAttention(num_units=units, activation=dnn_activation, use_res=True, dropout_rate=dropout_rate) \
([user_emb_output, short_att_output, short_sess_length])
gate_input = concat_func([prefer_output, short_output, user_emb_output])
gate = Dense(units, activation='sigmoid')(gate_input)
gate_output = Lambda(
lambda x: tf.multiply(x[0], x[1]) + tf.multiply(1 - x[0], x[2]))(
[gate, short_output, prefer_output])
gate_output_reshape = Lambda(lambda x: tf.squeeze(x, 1))(gate_output)
item_index = EmbeddingIndex(list(range(item_vocabulary_size)))(
item_features[item_feature_name])
item_embedding_matrix = embedding_matrix_dict[item_feature_name]
item_embedding_weight = NoMask()(item_embedding_matrix(item_index))
pooling_item_embedding_weight = PoolingLayer()([item_embedding_weight])
output = SampledSoftmaxLayer(num_sampled=num_sampled)([
pooling_item_embedding_weight, gate_output_reshape,
item_features[item_feature_name]
])
model = Model(inputs=user_inputs_list + item_inputs_list, outputs=output)
model.__setattr__("user_input", user_inputs_list)
model.__setattr__("user_embedding", gate_output_reshape)
model.__setattr__("item_input", item_inputs_list)
model.__setattr__(
"item_embedding",
get_item_embedding(pooling_item_embedding_weight,
item_features[item_feature_name]))
return model
def YoutubeDNN(
user_feature_columns,
item_feature_columns,
num_sampled=5,
user_dnn_hidden_units=(64, 16),
dnn_activation='relu',
dnn_use_bn=False,
l2_reg_dnn=0,
l2_reg_embedding=1e-6,
dnn_dropout=0,
init_std=0.0001,
seed=1024,
):
"""Instantiates the YoutubeDNN Model architecture.
:param user_feature_columns: An iterable containing user's features used by the model.
:param item_feature_columns: An iterable containing item's features used by the model.
:param num_sampled: int, the number of classes to randomly sample per batch.
:param user_dnn_hidden_units: list,list of positive integer or empty list, the layer number and units in each layer of user tower
:param dnn_activation: Activation function to use in deep net
:param dnn_use_bn: bool. Whether use BatchNormalization before activation or not in deep net
:param l2_reg_dnn: float. L2 regularizer strength applied to DNN
:param l2_reg_embedding: float. L2 regularizer strength applied to embedding vector
:param dnn_dropout: float in [0,1), the probability we will drop out a given DNN coordinate.
:param init_std: float,to use as the initialize std of embedding vector
:param seed: integer ,to use as random seed.
:return: A Keras model instance.
"""
if len(item_feature_columns) > 1:
raise ValueError(
"Now YoutubeNN only support 1 item feature like item_id")
item_feature_name = item_feature_columns[0].name
embedding_matrix_dict = create_embedding_matrix(user_feature_columns +
item_feature_columns,
l2_reg_embedding,
init_std,
seed,
prefix="")
user_features = build_input_features(user_feature_columns)
user_inputs_list = list(user_features.values())
user_sparse_embedding_list, user_dense_value_list = input_from_feature_columns(
user_features,
user_feature_columns,
l2_reg_embedding,
init_std,
seed,
embedding_matrix_dict=embedding_matrix_dict)
user_dnn_input = combined_dnn_input(user_sparse_embedding_list,
user_dense_value_list)
item_features = build_input_features(item_feature_columns)
item_inputs_list = list(item_features.values())
user_dnn_out = DNN(
user_dnn_hidden_units,
dnn_activation,
l2_reg_dnn,
dnn_dropout,
dnn_use_bn,
seed,
)(user_dnn_input)
item_embedding = embedding_matrix_dict[item_feature_name]
output = SampledSoftmaxLayer(item_embedding, num_sampled=num_sampled)(
inputs=(user_dnn_out, item_features[item_feature_name]))
model = Model(inputs=user_inputs_list + item_inputs_list, outputs=output)
model.__setattr__("user_input", user_inputs_list)
model.__setattr__("user_embedding", user_dnn_out)
model.__setattr__("item_input", item_inputs_list)
model.__setattr__(
"item_embedding",
get_item_embedding(item_embedding, item_features[item_feature_name]))
return model
def _model_fn(features, labels, mode, config):
train_flag = (mode == tf.estimator.ModeKeys.TRAIN)
with variable_scope(DNN_SCOPE_NAME):
sparse_feature_columns = []
dense_feature_columns = []
varlen_sparse_feature_columns = []
for feat in dnn_feature_columns:
new_feat_name = list(feat.parse_example_spec.keys())[0]
if new_feat_name in ['hist_price_id', 'hist_des_id']:
varlen_sparse_feature_columns.append(
VarLenSparseFeat(SparseFeat(new_feat_name,
vocabulary_size=100,
embedding_dim=32,
use_hash=False),
maxlen=3))
elif is_embedding(feat):
sparse_feature_columns.append(
SparseFeat(new_feat_name,
vocabulary_size=feat[0]._num_buckets + 1,
embedding_dim=feat.dimension))
else:
dense_feature_columns.append(DenseFeat(new_feat_name))
history_feature_columns = []
sparse_varlen_feature_columns = []
history_fc_names = list(
map(lambda x: "hist_" + x, history_feature_list))
for fc in varlen_sparse_feature_columns:
feature_name = fc.name
if feature_name in history_fc_names:
history_feature_columns.append(fc)
else:
sparse_varlen_feature_columns.append(fc)
my_feature_columns = sparse_feature_columns + dense_feature_columns + varlen_sparse_feature_columns
embedding_dict = create_embedding_matrix(my_feature_columns,
l2_reg_embedding,
seed,
prefix="")
query_emb_list = embedding_lookup(embedding_dict,
features,
sparse_feature_columns,
history_feature_list,
history_feature_list,
to_list=True)
print('query_emb_list', query_emb_list)
print('embedding_dict', embedding_dict)
print('haha')
print('history_feature_columns', history_feature_columns)
print('haha')
keys_emb_list = embedding_lookup(embedding_dict,
features,
history_feature_columns,
history_fc_names,
history_fc_names,
to_list=True)
print('keys_emb_list', keys_emb_list)
dnn_input_emb_list = embedding_lookup(
embedding_dict,
features,
sparse_feature_columns,
mask_feat_list=history_feature_list,
to_list=True)
print('dnn_input_emb_list', dnn_input_emb_list)
dense_value_list = get_dense_input(features, dense_feature_columns)
sequence_embed_dict = varlen_embedding_lookup(
embedding_dict, features, sparse_varlen_feature_columns)
sequence_embed_list = get_varlen_pooling_list(
sequence_embed_dict,
features,
sparse_varlen_feature_columns,
to_list=True)
dnn_input_emb_list += sequence_embed_list
keys_emb = concat_func(keys_emb_list, mask=True)
deep_input_emb = concat_func(dnn_input_emb_list)
query_emb = concat_func(query_emb_list, mask=True)
hist = AttentionSequencePoolingLayer(
att_hidden_size,
att_activation,
weight_normalization=att_weight_normalization,
supports_masking=True)([query_emb, keys_emb])
deep_input_emb = tf.keras.layers.Concatenate()(
[NoMask()(deep_input_emb), hist])
deep_input_emb = tf.keras.layers.Flatten()(deep_input_emb)
dnn_input = combined_dnn_input([deep_input_emb], dense_value_list)
output = DNN(dnn_hidden_units,
dnn_activation,
l2_reg_dnn,
dnn_dropout,
dnn_use_bn,
seed=seed)(dnn_input)
final_logit = tf.keras.layers.Dense(
1,
use_bias=False,
kernel_initializer=tf.keras.initializers.glorot_normal(seed))(
output)
# logits_list.append(final_logit)
# logits = add_func(logits_list)
# print(labels)
# tf.summary.histogram(final_logit + '/final_logit', final_logit)
return deepctr_model_fn(features,
mode,
final_logit,
labels,
task,
linear_optimizer,
dnn_optimizer,
training_chief_hooks=training_chief_hooks)
def DSSM(user_feature_columns,
item_feature_columns,
user_dnn_hidden_units=(64, 32),
item_dnn_hidden_units=(64, 32),
dnn_activation='tanh',
dnn_use_bn=False,
l2_reg_dnn=0,
l2_reg_embedding=1e-6,
dnn_dropout=0,
init_std=0.0001,
seed=1024,
metric='cos'):
embedding_matrix_dict = create_embedding_matrix(user_feature_columns +
item_feature_columns,
l2_reg_embedding,
init_std,
seed,
seq_mask_zero=True)
user_features = build_input_features(user_feature_columns)
user_inputs_list = list(user_features.values())
user_sparse_embedding_list, user_dense_value_list = input_from_feature_columns(
user_features,
user_feature_columns,
l2_reg_embedding,
init_std,
seed,
embedding_matrix_dict=embedding_matrix_dict)
user_dnn_input = combined_dnn_input(user_sparse_embedding_list,
user_dense_value_list)
item_features = build_input_features(item_feature_columns)
item_inputs_list = list(item_features.values())
item_sparse_embedding_list, item_dense_value_list = input_from_feature_columns(
item_features,
item_feature_columns,
l2_reg_embedding,
init_std,
seed,
embedding_matrix_dict=embedding_matrix_dict)
item_dnn_input = combined_dnn_input(item_sparse_embedding_list,
item_dense_value_list)
user_dnn_out = DNN(
user_dnn_hidden_units,
dnn_activation,
l2_reg_dnn,
dnn_dropout,
dnn_use_bn,
seed,
)(user_dnn_input)
item_dnn_out = DNN(item_dnn_hidden_units, dnn_activation, l2_reg_dnn,
dnn_dropout, dnn_use_bn, seed)(item_dnn_input)
score = Similarity(type=metric)([user_dnn_out, item_dnn_out])
output = PredictionLayer("binary", False)(score)
model = Model(inputs=user_inputs_list + item_inputs_list, outputs=output)
plot_model(model, to_file='dnn.png', show_shapes=True)
print("go")
model.__setattr__("user_input", user_inputs_list)
model.__setattr__("item_input", item_inputs_list)
model.__setattr__("user_embedding", user_dnn_out)
model.__setattr__("item_embedding", item_dnn_out)
return model
def DSSM(user_feature_columns,
item_feature_columns,
user_dnn_hidden_units=(64, 32),
item_dnn_hidden_units=(64, 32),
dnn_activation='tanh',
dnn_use_bn=False,
l2_reg_dnn=0,
l2_reg_embedding=1e-6,
dnn_dropout=0,
init_std=0.0001,
seed=1024,
metric='cos'):
"""Instantiates the Deep Structured Semantic Model architecture.
:param user_feature_columns: An iterable containing user's features used by the model.
:param item_feature_columns: An iterable containing item's features used by the model.
:param user_dnn_hidden_units: list,list of positive integer or empty list, the layer number and units in each layer of user tower
:param item_dnn_hidden_units: list,list of positive integer or empty list, the layer number and units in each layer of item tower
:param dnn_activation: Activation function to use in deep net
:param dnn_use_bn: bool. Whether use BatchNormalization before activation or not in deep net
:param l2_reg_dnn: float. L2 regularizer strength applied to DNN
:param l2_reg_embedding: float. L2 regularizer strength applied to embedding vector
:param dnn_dropout: float in [0,1), the probability we will drop out a given DNN coordinate.
:param init_std: float,to use as the initialize std of embedding vector
:param seed: integer ,to use as random seed.
:param metric: str, ``"cos"`` for cosine or ``"ip"`` for inner product
:return: A Keras model instance.
"""
embedding_matrix_dict = create_embedding_matrix(user_feature_columns +
item_feature_columns,
l2_reg_embedding,
init_std,
seed,
seq_mask_zero=True)
user_features = build_input_features(user_feature_columns)
user_inputs_list = list(user_features.values())
user_sparse_embedding_list, user_dense_value_list = input_from_feature_columns(
user_features,
user_feature_columns,
l2_reg_embedding,
init_std,
seed,
embedding_matrix_dict=embedding_matrix_dict)
user_dnn_input = combined_dnn_input(user_sparse_embedding_list,
user_dense_value_list)
item_features = build_input_features(item_feature_columns)
item_inputs_list = list(item_features.values())
item_sparse_embedding_list, item_dense_value_list = input_from_feature_columns(
item_features,
item_feature_columns,
l2_reg_embedding,
init_std,
seed,
embedding_matrix_dict=embedding_matrix_dict)
item_dnn_input = combined_dnn_input(item_sparse_embedding_list,
item_dense_value_list)
user_dnn_out = DNN(
user_dnn_hidden_units,
dnn_activation,
l2_reg_dnn,
dnn_dropout,
dnn_use_bn,
seed,
)(user_dnn_input)
item_dnn_out = DNN(item_dnn_hidden_units, dnn_activation, l2_reg_dnn,
dnn_dropout, dnn_use_bn, seed)(item_dnn_input)
score = Similarity(type=metric)([user_dnn_out, item_dnn_out])
output = PredictionLayer("binary", False)(score)
model = Model(inputs=user_inputs_list + item_inputs_list, outputs=output)
model.__setattr__("user_input", user_inputs_list)
model.__setattr__("item_input", item_inputs_list)
model.__setattr__("user_embedding", user_dnn_out)
model.__setattr__("item_embedding", item_dnn_out)
return model