def conv(x, filters, kernel_size, strides=(1, 1), padding='same', activation="Mish", use_bias=True):
x = Conv2D(filters, kernel_size, strides, padding, use_bias=use_bias)(x)
x = BatchNormalization()(x)
if activation == 'LeakRelu':
x = LeakyReLU()(x)
if activation == 'Mish':
x = Activation('Mish')(x)
if activation == 'Linear':
x = linear(x)
return x
def compute_similarity(self, repeated_context_vectors,
repeated_query_vectors):
element_wise_multiply = repeated_context_vectors * repeated_query_vectors
concatenated_tensor = K.concatenate([
repeated_context_vectors, repeated_query_vectors,
element_wise_multiply
],
axis=-1)
dot_product = K.squeeze(K.dot(concatenated_tensor, self.kernel),
axis=-1)
return linear(dot_product + self.bias)
def Conv2D_BN_Act(x,
filters,
size,
strides=1,
batch_norm=True,
activation="mish"):
padding = ('same' if strides == 1 else 'valid')
x = Conv2D(filters=filters,
kernel_size=size,
strides=strides,
padding=padding,
use_bias=not batch_norm,
kernel_regularizer=l2(5e-4))(x)
if batch_norm:
x = BatchNormalization()(x)
if activation in ["Mish", "mish"]:
x = Mish()(x)
elif activation in ["LeakyReLU", "leakyrelu"]:
x = LeakyReLU(alpha=0.1)(x)
elif activation in ["Linear", "linear"]:
x = linear(x)
return x