def build(self, input_shape):
dnn_hidden = [int(h) for h in self.flags.deep_layers.split(',')]
cin_hidden = [int(h) for h in self.flags.cin_layers.split(',')]
self.fm_block = self.build_fm()
self.linear_block = Linear()
self.dnn_block = self.build_deep(dnn_hidden)
self.cin_block = self.build_cin(cin_hidden)
def build(self, input_shape):
n, p = input_shape
self.linear_block = Linear()
# interaction factors, randomly initialized
self.kernel = self.add_weight(shape=(self.k, p),
initializer='random_normal',
trainable=True)
def build(self, input_shape):
dnn_hidden = [int(h) for h in self.flags.deep_layers.split(',')]
cin_hidden = [int(h) for h in self.flags.cin_layers.split(',')]
self.embeddingsize = self.flags.embed_dim
self.numbericfeaturescnt = len(self.NUMERICAL_FEATURES)
self.fm_block = self.build_fm()
self.linear_block = Linear()
self.dnn_block = self.build_deep(dnn_hidden)
self.cin_block = self.build_cin(cin_hidden)
def build(self, input_shape):
self.linear_part = Linear()
self.field_nums = len(input_shape)
index = 0
self.field_dict = {}
for idx, val in enumerate(input_shape):
if val in self.NUMERICAL_FEATURES:
self.field_dict[index] = idx
index += 1
if val in self.CATEGORY_FEATURES:
for i in range(self.voc_emb_size[val][1]):
self.field_dict[index] = idx
index += 1
self.total_dims = len(self.field_dict)
self.kernel = self.add_weight('v', shape=[self.total_dims, self.field_nums, self.k],
initializer=tf.keras.initializers.TruncatedNormal(mean=0, stddev=0.01))
def build_wide(self, hidden=1):
return Linear(hidden)
def build_linear(self, hidden=1):
return Linear(hidden)