def wider_pre_conv(layer, n_add_filters, weighted=True):
'''wider previous conv layer.
'''
n_dim = get_n_dim(layer)
if not weighted:
return get_conv_class(n_dim)(
layer.input_channel,
layer.filters + n_add_filters,
kernel_size=layer.kernel_size,
)
n_pre_filters = layer.filters
rand = np.random.randint(n_pre_filters, size=n_add_filters)
teacher_w, teacher_b = layer.get_weights()
student_w = teacher_w.copy()
student_b = teacher_b.copy()
# target layer update (i)
for i, _ in enumerate(rand):
teacher_index = rand[i]
new_weight = teacher_w[teacher_index, ...]
new_weight = new_weight[np.newaxis, ...]
student_w = np.concatenate((student_w, new_weight), axis=0)
student_b = np.append(student_b, teacher_b[teacher_index])
new_pre_layer = get_conv_class(n_dim)(layer.input_channel,
n_pre_filters + n_add_filters,
layer.kernel_size)
new_pre_layer.set_weights(
(add_noise(student_w, teacher_w), add_noise(student_b, teacher_b)))
return new_pre_layer
def wider_next_conv(layer, start_dim, total_dim, n_add, weighted=True):
'''wider next conv layer.
'''
n_dim = get_n_dim(layer)
if not weighted:
return get_conv_class(n_dim)(layer.input_channel + n_add,
layer.filters,
kernel_size=layer.kernel_size,
stride=layer.stride)
n_filters = layer.filters
teacher_w, teacher_b = layer.get_weights()
new_weight_shape = list(teacher_w.shape)
new_weight_shape[1] = n_add
new_weight = np.zeros(tuple(new_weight_shape))
student_w = np.concatenate(
(teacher_w[:, :start_dim, ...].copy(), add_noise(
new_weight, teacher_w), teacher_w[:, start_dim:total_dim,
...].copy()),
axis=1)
new_layer = get_conv_class(n_dim)(layer.input_channel + n_add,
n_filters,
kernel_size=layer.kernel_size,
stride=layer.stride)
new_layer.set_weights((student_w, teacher_b))
return new_layer
def create_new_layer(layer, n_dim):
''' create new layer for the graph
'''
input_shape = layer.output.shape
dense_deeper_classes = [StubDense, get_dropout_class(n_dim), StubReLU]
conv_deeper_classes = [
get_conv_class(n_dim),
get_batch_norm_class(n_dim), StubReLU
]
if is_layer(layer, "ReLU"):
conv_deeper_classes = [
get_conv_class(n_dim),
get_batch_norm_class(n_dim)
]
dense_deeper_classes = [StubDense, get_dropout_class(n_dim)]
elif is_layer(layer, "Dropout"):
dense_deeper_classes = [StubDense, StubReLU]
elif is_layer(layer, "BatchNormalization"):
conv_deeper_classes = [get_conv_class(n_dim), StubReLU]
layer_class = None
if len(input_shape) == 1:
# It is in the dense layer part.
layer_class = sample(dense_deeper_classes, 1)[0]
else:
# It is in the conv layer part.
layer_class = sample(conv_deeper_classes, 1)[0]
if layer_class == StubDense:
new_layer = StubDense(input_shape[0], input_shape[0])
elif layer_class == get_dropout_class(n_dim):
new_layer = layer_class(Constant.DENSE_DROPOUT_RATE)
elif layer_class == get_conv_class(n_dim):
new_layer = layer_class(input_shape[-1],
input_shape[-1],
sample((1, 3, 5), 1)[0],
stride=1)
elif layer_class == get_batch_norm_class(n_dim):
new_layer = layer_class(input_shape[-1])
elif layer_class == get_pooling_class(n_dim):
new_layer = layer_class(sample((1, 3, 5), 1)[0])
else:
new_layer = layer_class()
return new_layer
def _insert_conv_layer_chain2(self, start_node_id, end_node_id,filters_start, filters_end):
skip_output_id = start_node_id
num = len(self._get_pooling_layers(start_node_id, end_node_id))
# f = open("/root/trials/" + "/chain.log", "a+")
# f.write("num=" + str(num) + "\n")
if num >= 1:
new_layer = get_conv_class(self.n_dim)(filters_start, filters_end, 1, 2**num)
skip_output_id = self.add_layer(new_layer, skip_output_id)
skip_output_id = self.add_layer(self.batch_norm(self.node_list[skip_output_id].shape[-1]), skip_output_id)
elif filters_start != filters_end:
new_layer = get_conv_class(self.n_dim)(filters_start, filters_end, 1, 1)
skip_output_id = self.add_layer(new_layer, skip_output_id)
skip_output_id = self.add_layer(self.batch_norm(self.node_list[skip_output_id].shape[-1]), skip_output_id)
return skip_output_id
def deeper_conv_block(conv_layer, kernel_size, weighted=True):
'''deeper conv layer.
'''
n_dim = get_n_dim(conv_layer)
filter_shape = (kernel_size, ) * 2
n_filters = conv_layer.filters
weight = np.zeros((n_filters, n_filters) + filter_shape)
center = tuple(map(lambda x: int((x - 1) / 2), filter_shape))
for i in range(n_filters):
filter_weight = np.zeros((n_filters, ) + filter_shape)
index = (i, ) + center
filter_weight[index] = 1
weight[i, ...] = filter_weight
bias = np.zeros(n_filters)
new_conv_layer = get_conv_class(n_dim)(conv_layer.filters,
n_filters,
kernel_size=kernel_size)
bn = get_batch_norm_class(n_dim)(n_filters)
if weighted:
new_conv_layer.set_weights(
(add_noise(weight,
np.array([0, 1])), add_noise(bias, np.array([0, 1]))))
new_weights = [
add_noise(np.ones(n_filters, dtype=np.float32), np.array([0, 1])),
add_noise(np.zeros(n_filters, dtype=np.float32), np.array([0, 1])),
add_noise(np.zeros(n_filters, dtype=np.float32), np.array([0, 1])),
add_noise(np.ones(n_filters, dtype=np.float32), np.array([0, 1])),
]
bn.set_weights(new_weights)
return [StubReLU(), new_conv_layer, bn]
def to_concat_skip_model(self, start_id, end_id):
"""Add a weighted add concatenate connection from after start node to end node.
Args:
start_id: The convolutional layer ID, after which to start the skip-connection.
end_id: The convolutional layer ID, after which to end the skip-connection.
"""
self.operation_history.append(
("to_concat_skip_model", start_id, end_id))
filters_end = self.layer_list[end_id].output.shape[-1]
filters_start = self.layer_list[start_id].output.shape[-1]
start_node_id = self.layer_id_to_output_node_ids[start_id][0]
pre_end_node_id = self.layer_id_to_input_node_ids[end_id][0]
end_node_id = self.layer_id_to_output_node_ids[end_id][0]
skip_output_id = self._insert_conv_layer_chain3(
start_node_id, end_node_id, filters_start)
# skip_output_id = self._insert_pooling_layer_chain(
# start_node_id, end_node_id)
concat_input_node_id = self._add_node(
deepcopy(self.node_list[end_node_id]))
self._redirect_edge(pre_end_node_id, end_node_id, concat_input_node_id)
concat_layer = StubConcatenate()
concat_layer.input = [
self.node_list[concat_input_node_id],
self.node_list[skip_output_id],
]
concat_output_node_id = self._add_node(Node(concat_layer.output_shape))
self._add_edge(concat_layer, concat_input_node_id,
concat_output_node_id)
self._add_edge(concat_layer, skip_output_id, concat_output_node_id)
concat_layer.output = self.node_list[concat_output_node_id]
self.node_list[concat_output_node_id].shape = concat_layer.output_shape
# Add the concatenate layer.
new_conv_layer = get_conv_class(self.n_dim)(
filters_start + filters_end, filters_end, 1)
self._add_edge(new_conv_layer, concat_output_node_id, end_node_id)
new_conv_layer.input = self.node_list[concat_output_node_id]
new_conv_layer.output = self.node_list[end_node_id]
self.node_list[end_node_id].shape = new_conv_layer.output_shape
if self.weighted:
filter_shape = (1, ) * self.n_dim
weights = np.zeros((filters_end, filters_end) + filter_shape)
for i in range(filters_end):
filter_weight = np.zeros((filters_end, ) + filter_shape)
center_index = (i, ) + (0, ) * self.n_dim
filter_weight[center_index] = 1
weights[i, ...] = filter_weight
weights = np.concatenate(
(weights,
np.zeros((filters_end, filters_start) + filter_shape)),
axis=1)
bias = np.zeros(filters_end)
new_conv_layer.set_weights((add_noise(weights, np.array([0, 1])),
add_noise(bias, np.array([0, 1]))))
def to_deeper_graph2(graph):
''' deeper graph
'''
weighted_layer_ids = graph.deep_layer_ids2()
if len(weighted_layer_ids) >= Constant.MAX_LAYERS:
return None
deeper_layer_ids = sample(weighted_layer_ids, 1)#选一层
for layer_id in deeper_layer_ids:
layer = graph.layer_list[layer_id]
input_shape = layer.output.shape
layer_class = get_conv_class(graph.n_dim)
new_layer = layer_class(input_shape[-1], input_shape[-1], 3, stride=1)
output_id = graph.to_deeper_model(layer_id, new_layer)
layer_id2 = graph.get_layers_id(output_id)
layer2 = graph.layer_list[layer_id2]
input_shape2 = layer2.output.shape
layer_class = get_batch_norm_class(graph.n_dim)
new_layer2 = layer_class(input_shape2[-1])
output_id2=graph.to_deeper_model(layer_id2, new_layer2)
layer_id3 = graph.get_layers_id(output_id2)
graph.to_deeper_model(layer_id3, StubReLU())
return graph
def _insert_conv_layer_chain3(self, start_node_id, end_node_id,filters_start):
skip_output_id = start_node_id
num = len(self._get_pooling_layers(start_node_id, end_node_id))
if num >= 1:
new_layer = get_conv_class(self.n_dim)(filters_start, filters_start, 1, 2**num)
skip_output_id = self.add_layer(new_layer, skip_output_id)
skip_output_id = self.add_layer(self.batch_norm(self.node_list[skip_output_id].shape[-1]), skip_output_id)
return skip_output_id
def __init__(self, n_output_node, input_shape):
super(CnnGenerator, self).__init__(n_output_node, input_shape)
self.n_dim = len(self.input_shape) - 1
if len(self.input_shape) > 4:
raise ValueError("The input dimension is too high.")
if len(self.input_shape) < 2:
raise ValueError("The input dimension is too low.")
self.conv = get_conv_class(self.n_dim)
self.dropout = get_dropout_class(self.n_dim)
self.global_avg_pooling = get_global_avg_pooling_class(self.n_dim)
self.pooling = get_pooling_class(self.n_dim)
self.batch_norm = get_batch_norm_class(self.n_dim)
def _insert_pooling_layer_chain(self, start_node_id, end_node_id):
skip_output_id = start_node_id
for layer in self._get_pooling_layers(start_node_id, end_node_id):
new_layer = deepcopy(layer)
if is_layer(new_layer, "Conv"):
filters = self.node_list[start_node_id].shape[-1]
new_layer = get_conv_class(self.n_dim)(filters, filters, 1,
layer.stride)
if self.weighted:
init_conv_weight(new_layer)
else:
new_layer = deepcopy(layer)
skip_output_id = self.add_layer(new_layer, skip_output_id)
skip_output_id = self.add_layer(StubReLU(), skip_output_id)
return skip_output_id
def to_add_skip_model(self, start_id, end_id):
"""Add a weighted add skip-connection from after start node to end node.
Args:
start_id: The convolutional layer ID, after which to start the skip-connection.
end_id: The convolutional layer ID, after which to end the skip-connection.
"""
self.operation_history.append(("to_add_skip_model", start_id, end_id))
filters_end = self.layer_list[end_id].output.shape[-1]
filters_start = self.layer_list[start_id].output.shape[-1]
start_node_id = self.layer_id_to_output_node_ids[start_id][0]
pre_end_node_id = self.layer_id_to_input_node_ids[end_id][0]
end_node_id = self.layer_id_to_output_node_ids[end_id][0]
skip_output_id = self._insert_pooling_layer_chain(
start_node_id, end_node_id)
# Add the conv layer
new_conv_layer = get_conv_class(self.n_dim)(filters_start, filters_end,
1)
skip_output_id = self.add_layer(new_conv_layer, skip_output_id)
# Add the add layer.
add_input_node_id = self._add_node(
deepcopy(self.node_list[end_node_id]))
add_layer = StubAdd()
self._redirect_edge(pre_end_node_id, end_node_id, add_input_node_id)
self._add_edge(add_layer, add_input_node_id, end_node_id)
self._add_edge(add_layer, skip_output_id, end_node_id)
add_layer.input = [
self.node_list[add_input_node_id],
self.node_list[skip_output_id],
]
add_layer.output = self.node_list[end_node_id]
self.node_list[end_node_id].shape = add_layer.output_shape
# Set weights to the additional conv layer.
if self.weighted:
filter_shape = (1, ) * self.n_dim
weights = np.zeros((filters_end, filters_start) + filter_shape)
bias = np.zeros(filters_end)
new_conv_layer.set_weights((add_noise(weights, np.array([0, 1])),
add_noise(bias, np.array([0, 1]))))
def __init__(self, input_shape, weighted=True):
"""Initializer for Graph.
"""
self.input_shape = input_shape
self.weighted = weighted
self.node_list = []
self.layer_list = []
# node id start with 0
self.node_to_id = {}
self.layer_to_id = {}
self.layer_id_to_input_node_ids = {}
self.layer_id_to_output_node_ids = {}
self.adj_list = {}
self.reverse_adj_list = {}
self.operation_history = []
self.n_dim = len(input_shape) - 1
self.conv = get_conv_class(self.n_dim)
self.batch_norm = get_batch_norm_class(self.n_dim)
self.vis = None
self._add_node(Node(input_shape))
