def train(self):
self.train_dataset, \
self.test_dataset = self.data_process.generate_data(batch_size=self.batch_size,
epoch=self.epoch)
fm = FM(user_num=self.user_num,
movie_num=self.movie_num,
year_num=self.year_num,
genre_num=self.genre_num,
embedding_size=self.embedding_size,
is_sigmoid=True)
fm_model = fm.get_fm_model()
fm_model.summary()
fm_model.compile(optimizer="adam",
loss=keras.losses.binary_crossentropy,
metrics=["accuracy"])
fm_model.fit(self.train_dataset,
epochs=self.epoch,
steps_per_epoch=1270010 // self.batch_size + 1,
callbacks=[
get_callback(self.test_dataset, self.model_path,
self.version)
])
def __init__(self,
user_num=4691,
movie_num=2514,
year_num=76,
genre_num=9,
embedding_size=16,
att_embedding_size=16,
head_num=2,
att_layer_num=3,
use_res=True):
self.user_num = user_num
self.movie_num = movie_num
self.year_num = year_num
self.genre_num = genre_num
self.embedding_size = embedding_size
self.att_layer_num = att_layer_num
self.att_embedding_size = att_embedding_size
self.head_num = head_num
self.use_res = use_res
self.fm = FM(user_num=self.user_num,
movie_num=self.movie_num,
year_num=self.year_num,
genre_num=self.genre_num,
embedding_size=self.embedding_size)
def __init__(self,
user_num=4691,
movie_num=2514,
year_num=76,
genre_num=9,
embedding_size=16,
dnn_hidden_units=(
128,
128,
),
cross_layer_num=2):
self.user_num = user_num
self.movie_num = movie_num
self.year_num = year_num
self.genre_num = genre_num
self.embedding_size = embedding_size
self.dnn_hidden_units = dnn_hidden_units
self.cross_layer_num = cross_layer_num
self.fm = FM(user_num=self.user_num,
movie_num=self.movie_num,
year_num=self.year_num,
genre_num=self.genre_num,
embedding_size=self.embedding_size)
def __init__(self,
user_num=4691,
movie_num=2514,
year_num=76,
genre_num=9,
embedding_size=16,
attention_fator=16):
"""
:param user_num:
:param movie_num:
:param year_num:
:param genre_num:
:param embedding_size:
:param attention_fator: attention dense num
"""
self.user_num = user_num
self.movie_num = movie_num
self.year_num = year_num
self.genre_num = genre_num
self.embedding_size = embedding_size
self.attention_fator = attention_fator
self.fm = FM(user_num=self.user_num,
movie_num=self.movie_num,
year_num=self.year_num,
genre_num=self.genre_num,
embedding_size=self.embedding_size)
class WDL(object):
def __init__(self,
user_num=4691,
movie_num=2514,
year_num=76,
genre_num=9,
embedding_size=16):
self.user_num = user_num
self.movie_num = movie_num
self.year_num = year_num
self.genre_num = genre_num
self.embedding_size = embedding_size
self.fm = FM(user_num=self.user_num,
movie_num=self.movie_num,
year_num=self.year_num,
genre_num=self.genre_num,
embedding_size=self.embedding_size)
"""deep 端"""
def get_deep_logit(self, sparse_embedding):
x = Concatenate(axis=2)(sparse_embedding)
x = Flatten()(x)
# x = Lambda(lambda x: tf.squeeze(x,axis=1))(x)
x = Dense(128, activation="relu")(x)
x = Dense(128, activation="relu")(x)
logit = Dense(1, use_bias=False)(x)
return logit
def get_wdl_model(self):
"""wide 端"""
linear_logit = self.fm.get_linear_logit()
"""deep 端"""
sparse_embedding = self.fm.get_sparse_embedding()
deep_logit = self.get_deep_logit(sparse_embedding)
x = Add()([linear_logit, deep_logit])
outputs = Dense(1, activation="sigmoid")(x)
model = Model(inputs=self.fm.inputs, outputs=outputs)
return model
class AFM(object):
def __init__(self,
user_num=4691,
movie_num=2514,
year_num=76,
genre_num=9,
embedding_size=16,
attention_fator=16):
"""
:param user_num:
:param movie_num:
:param year_num:
:param genre_num:
:param embedding_size:
:param attention_fator: attention dense num
"""
self.user_num = user_num
self.movie_num = movie_num
self.year_num = year_num
self.genre_num = genre_num
self.embedding_size = embedding_size
self.attention_fator = attention_fator
self.fm = FM(user_num=self.user_num,
movie_num=self.movie_num,
year_num=self.year_num,
genre_num=self.genre_num,
embedding_size=self.embedding_size)
def get_afm_logit(self, sparse_embedding):
afm_out = AFMLayer(self.attention_fator)(sparse_embedding)
return afm_out
def get_afm_model(self):
linear_logit = self.fm.get_linear_logit()
sparse_embedding = self.fm.get_sparse_embedding()
afm_logit = self.get_afm_logit(sparse_embedding)
x = Add()([linear_logit, afm_logit])
outputs = Dense(1, activation="sigmoid")(x)
model = Model(inputs=self.fm.inputs, outputs=outputs)
return model
def __init__(self,
user_num=4691,
movie_num=2514,
year_num=76,
genre_num=9,
embedding_size=16):
self.user_num = user_num
self.movie_num = movie_num
self.year_num = year_num
self.genre_num = genre_num
self.embedding_size = embedding_size
self.fm = FM(user_num=self.user_num,
movie_num=self.movie_num,
year_num=self.year_num,
genre_num=self.genre_num,
embedding_size=self.embedding_size)
class AutoInt(object):
def __init__(self,
user_num=4691,
movie_num=2514,
year_num=76,
genre_num=9,
embedding_size=16,
att_embedding_size=16,
head_num=2,
att_layer_num=3,
use_res=True):
self.user_num = user_num
self.movie_num = movie_num
self.year_num = year_num
self.genre_num = genre_num
self.embedding_size = embedding_size
self.att_layer_num = att_layer_num
self.att_embedding_size = att_embedding_size
self.head_num = head_num
self.use_res = use_res
self.fm = FM(user_num=self.user_num,
movie_num=self.movie_num,
year_num=self.year_num,
genre_num=self.genre_num,
embedding_size=self.embedding_size)
def get_deep_output(self, inputs):
x = Flatten()(inputs)
x = Dense(128, activation="relu")(x)
out = Dense(128, activation="relu")(x)
return out
def get_transformer_logit(self, sparse_embedding):
inputs = Concatenate(axis=1)(sparse_embedding)
x = inputs
for i in range(self.att_layer_num):
x = MutiHeadSelfAttention(self.att_embedding_size, self.head_num,
self.use_res)(x)
att_output = Flatten()(x)
deep_output = self.get_deep_output(inputs)
x = Concatenate()([att_output, deep_output])
out = Dense(1, use_bias=False)(x)
return out
def get_autoint_model(self):
linear_logit = self.fm.get_linear_logit()
sparse_embedding = self.fm.get_sparse_embedding()
cross_logit = self.get_transformer_logit(sparse_embedding)
x = Add()([linear_logit, cross_logit])
output = Dense(1, activation="sigmoid")(x)
model = Model(inputs=self.fm.inputs, outputs=output)
return model
class DCN(object):
def __init__(self,
user_num=4691,
movie_num=2514,
year_num=76,
genre_num=9,
embedding_size=16,
dnn_hidden_units=(
128,
128,
),
cross_layer_num=2):
self.user_num = user_num
self.movie_num = movie_num
self.year_num = year_num
self.genre_num = genre_num
self.embedding_size = embedding_size
self.dnn_hidden_units = dnn_hidden_units
self.cross_layer_num = cross_layer_num
self.fm = FM(user_num=self.user_num,
movie_num=self.movie_num,
year_num=self.year_num,
genre_num=self.genre_num,
embedding_size=self.embedding_size)
"""deep 端 并没有进行 Dense(1) 处理"""
def get_deep_output(self, sparse_embedding):
x = Concatenate(axis=2)(sparse_embedding)
x = Flatten()(x)
for unit in list(self.dnn_hidden_units):
x = Dense(unit, activation="relu")(x)
return x
def get_cross_output(self, sparse_embedding):
x = Concatenate(axis=2)(sparse_embedding)
x = Flatten()(x)
cross_output = CrossLayer(cross_layer_num=self.cross_layer_num)(x)
return cross_output
def get_dcn_model(self):
linear_logit = self.fm.get_linear_logit()
sparse_embedding = self.fm.get_sparse_embedding()
deep_output = self.get_deep_output(sparse_embedding)
cross_output = self.get_cross_output(sparse_embedding)
x = Concatenate(axis=-1)([deep_output, cross_output])
cross_logit = Dense(1)(x)
logit = Add()([linear_logit, cross_logit])
output = Dense(1, activation="sigmoid")(logit)
model = Model(inputs=self.fm.inputs, outputs=output)
return model