def get_text_model(embedding_size=EMBEDDING_SIZE,
input_length=MAX_DOCUMENT_LENGTH):
inputs = keras.Input(shape=(input_length, embedding_size))
x = layers.Dropout(0.1)(inputs)
x = layers.Convolution1D(
16,
kernel_size=4,
activation='relu',
strides=1,
padding='same',
kernel_constraint=constraints.MaxNorm(max_value=3))(x)
x = layers.Dropout(0.5)(x)
x = layers.Convolution1D(
12,
kernel_size=8,
activation='relu',
strides=2,
padding='same',
kernel_constraint=constraints.MaxNorm(max_value=3))(x)
x = layers.Dropout(0.5)(x)
x = layers.Convolution1D(
8,
kernel_size=16,
activation='relu',
strides=2,
padding='same',
kernel_constraint=constraints.MaxNorm(max_value=3))(x)
x = layers.Dropout(0.5)(x)
outputs = layers.Flatten()(x)
# outputs = layers.Dense(2, activation='relu', kernel_constraint=constraints.MaxNorm(max_value=3))(x)
return keras.Model(inputs, outputs)
def get_model_head(embedding_size=EMBEDDING_SIZE,
input_length=MAX_DOCUMENT_LENGTH,
link_embedding_size=LINK_EMBEDDING_SIZE,
link_input_length=LINK_INPUT_LENGTH):
he_inputs = keras.Input(shape=(input_length, embedding_size))
en_inputs = keras.Input(shape=(input_length, embedding_size))
link_inputs = keras.Input(shape=(link_input_length,
link_embedding_size))
he_outputs = CnnClfEnsemble.get_text_model()(he_inputs)
en_outputs = CnnClfEnsemble.get_text_model()(en_inputs)
link_outputs = CnnClfEnsemble.get_link_model()(link_inputs)
x = layers.Concatenate()([he_outputs, en_outputs, link_outputs])
x = layers.Dense(128,
activation='relu',
kernel_constraint=constraints.MaxNorm(max_value=3))(x)
x = layers.Dropout(0.5)(x)
x = layers.Dense(64,
activation='relu',
kernel_constraint=constraints.MaxNorm(max_value=3))(x)
outputs = layers.Dense(
2,
activation='softmax',
kernel_constraint=constraints.MaxNorm(max_value=3))(x)
return keras.Model(inputs=[he_inputs, en_inputs, link_inputs],
outputs=outputs)
def get_mlp_model(self, num_classes, name):
inputs = keras.Input((self.input_length, self.embedding_size))
x = layers.Dropout(0.5)(inputs)
x = layers.Dense(128, activation='relu', kernel_constraint=constraints.MaxNorm(max_value=3))(x)
x = layers.Dropout(0.5)(x)
x = layers.Dense(64, activation='relu', kernel_constraint=constraints.MaxNorm(max_value=3))(x)
x = layers.Dropout(0.5)(x)
x = layers.Dense(32, activation='relu', kernel_constraint=constraints.MaxNorm(max_value=3))(x)
x = layers.Dropout(0.5)(x)
outputs = layers.Dense(num_classes, activation='softmax', kernel_constraint=constraints.MaxNorm(max_value=3))(x)
return keras.Model(inputs, outputs, name=name)
def test_cosinedense_reg_constraint():
layer_test(core.CosineDense,
kwargs={
'units': 3,
'kernel_regularizer': regularizers.l2(0.01),
'bias_regularizer': regularizers.l1(0.01),
'activity_regularizer': regularizers.l2(0.01),
'kernel_constraint': constraints.MaxNorm(1),
'bias_constraint': constraints.MaxNorm(1)
},
input_shape=(3, 2))
def _set_constraints(self):
if self.scope in self.specs['maxnorm_scope']:
constr = k_con.MaxNorm(2.)
print('Setting constraint for {}, to MaxNorm'.format(self.scope))
else:
constr = None
return constr
def __init__(self,
model_name,
klass_name,
embedding_matrix,
embedding_size=EMBEDDING_SIZE,
input_length=MAX_DOCUMENT_LENGTH):
self.klass_name = klass_name
self.model = models.Sequential(name=f'{model_name}-model')
self.model.add(
layers.Embedding(
embedding_matrix.shape[0],
embedding_size,
input_length=input_length,
embeddings_initializer=initializers.Constant(embedding_matrix),
trainable=False))
# model.add(layers.Embedding(len(tokenizer.word_index)+1, embedding_size, input_length=MAX_DOCUMENT_LENGTH)) # for trainable embedding layer
self.model.add(layers.Dropout(0.1))
self.model.add(
layers.Convolution1D(
16,
kernel_size=4,
activation='relu',
strides=1,
padding='same',
kernel_constraint=constraints.MaxNorm(max_value=3)))
self.model.add(layers.Dropout(0.5))
self.model.add(
layers.Convolution1D(
12,
kernel_size=8,
activation='relu',
strides=2,
padding='same',
kernel_constraint=constraints.MaxNorm(max_value=3)))
self.model.add(layers.Dropout(0.5))
self.model.add(
layers.Convolution1D(
8,
kernel_size=16,
activation='relu',
strides=2,
padding='same',
kernel_constraint=constraints.MaxNorm(max_value=3)))
self.model.add(layers.Dropout(0.5))
self.model.add(layers.Flatten())
self.model.add(
layers.Dense(128,
activation='relu',
kernel_constraint=constraints.MaxNorm(max_value=3)))
self.model.add(layers.Dropout(0.5))
self.model.add(
layers.Dense(64,
activation='relu',
kernel_constraint=constraints.MaxNorm(max_value=3)))
self.model.add(layers.Dropout(0.5))
self.model.add(
layers.Dense(2,
activation='softmax',
kernel_constraint=constraints.MaxNorm(max_value=3)))
self.model.compile(
optimizer=optimizers.Adam(), #learning_rate=0.001),
loss=losses.CategoricalCrossentropy(from_logits=False),
metrics=[
metrics.CategoricalAccuracy(),
metrics.Recall(class_id=0),
metrics.Precision(class_id=0)
])