def shareable_dilated_conv_layers(n_channels, dilation_rate):
extent = 1 + (2 * dilation_rate)
shared_cnv = layers.AtrousConv1D(n_channels,
extent,
atrous_rate=dilation_rate,
border_mode="same")
shared_bnm = layers.BatchNormalization(mode=2)
return shared_cnv, shared_bnm
def _res_unit(self, inputs, nb_filter, size=3, stride=1, atrous=1,
stage=1, block=1):
name = '%02d-%02d/' % (stage, block)
id_name = '%sid_' % (name)
res_name = '%sres_' % (name)
# Residual branch
# 1x1 down-sample conv
x = kl.BatchNormalization(name=res_name + 'bn1')(inputs)
x = kl.Activation('relu', name=res_name + 'act1')(x)
w_reg = kr.WeightRegularizer(l1=self.l1_decay, l2=self.l2_decay)
x = kl.Conv1D(nb_filter[0], 1,
name=res_name + 'conv1',
subsample_length=stride,
init=self.init,
W_regularizer=w_reg)(x)
# LxL conv
x = kl.BatchNormalization(name=res_name + 'bn2')(x)
x = kl.Activation('relu', name=res_name + 'act2')(x)
w_reg = kr.WeightRegularizer(l1=self.l1_decay, l2=self.l2_decay)
x = kl.AtrousConv1D(nb_filter[1], size,
atrous_rate=atrous,
name=res_name + 'conv2',
border_mode='same',
init=self.init,
W_regularizer=w_reg)(x)
# 1x1 up-sample conv
x = kl.BatchNormalization(name=res_name + 'bn3')(x)
x = kl.Activation('relu', name=res_name + 'act3')(x)
w_reg = kr.WeightRegularizer(l1=self.l1_decay, l2=self.l2_decay)
x = kl.Conv1D(nb_filter[2], 1,
name=res_name + 'conv3',
init=self.init,
W_regularizer=w_reg)(x)
# Identity branch
if nb_filter[-1] != inputs._keras_shape[-1] or stride > 1:
w_reg = kr.WeightRegularizer(l1=self.l1_decay, l2=self.l2_decay)
identity = kl.Conv1D(nb_filter[2], 1,
name=id_name + 'conv1',
subsample_length=stride,
init=self.init,
W_regularizer=w_reg)(inputs)
else:
identity = inputs
x = kl.merge([identity, x], name=name + 'merge', mode='sum')
return x
def dilated_conv_block(x, dilation_rate, skip_inputs=None):
extent = 1 + (2 * dilation_rate)
n_channels = int(x.get_shape()[-1])
cnv = layers.AtrousConv1D(n_channels,
extent,
atrous_rate=dilation_rate,
border_mode="same")(x)
bnm = layers.BatchNormalization()(cnv)
act = layers.LeakyReLU(0.2)(bnm)
if skip_inputs is None:
return act
else:
return act, layers.merge(skip_inputs + [act], mode="sum")