def _construct_CNN(self, input_, p_size, params):
# Input block
k_size = [
params["k_size_input"], params["k_size_b1"], params["k_size_b2"],
params["k_size_b3"]
]
model = self._construct_CNN_input(input_, p_size, k_size)
# Block 3 res
min_3 = (self.CNN_sizes[p_size[3]] -
1) if (self.CNN_sizes[p_size[3]] - 1) > -1 else 0
filters_res = [min_3, min_3, self.CNN_sizes[p_size[3]]]
model = cmp.Conv1D_res_block(model, filters_res, k_size[3])
# Dropout
model = self._add_dropout(model, DNN=False)
# Extra blocks, gotta get deep
model = cmp.Conv1D_res_block(model, [128, 128, 128], kernel_shape=3)
# Dropout
model = self._add_dropout(model, DNN=False)
model = cmp.Conv1D_res_block(model,
[128, 128, self.CNN_sizes[p_size[3]]],
kernel_shape=3)
# Dropout
model = self._add_dropout(model, DNN=False)
if params["type_b4"] == 0:
# Block 4 ID
filters_id = [self.CNN_sizes[p_size[4]], self.CNN_sizes[p_size[3]]]
model = cmp.Conv1D_identity_block(model,
filters_id,
kernel_shape=3)
# Dropout
model = self._add_dropout(model, DNN=False)
# Flatten for dense
model = kl.Flatten()(model)
else:
# Flatten for dense
model = kl.Flatten()(model)
# Dense ID Block 4
model = cmp.dense_res_block(model, self.DNN_sizes[p_size[4]])
# Dropout
model = self._add_dropout(model, DNN=True)
return model
def _construct_CNN_input(self, input_model, p_size, kernels):
# CNN input block
model = cmp.Conv1D_input_layer(input_model,
self.CNN_sizes[p_size[0]],
kernel_shape=kernels[0])
# CNN Res Block
min_1 = (self.CNN_sizes[p_size[1]] -
1) if (self.CNN_sizes[p_size[1]] - 1) > -1 else 0
filters_res = [min_1, min_1, self.CNN_sizes[p_size[1]]]
model = cmp.Conv1D_res_block(model,
n_filters=filters_res,
kernel_shape=kernels[1])
# Dropout
model = self._add_dropout(model, DNN=False)
# CNN Identity Block
filters_id = [self.CNN_sizes[p_size[2]], self.CNN_sizes[p_size[1]]]
model = cmp.Conv1D_identity_block(model,
filters_id,
kernel_shape=kernels[2])
# Dropout
model = self._add_dropout(model, DNN=False)
return model