def wider_conv2d(self, model, layer_name, new_width, config):
# modify graph
new_graph = model.graph.copy()
new_node = new_graph.get_nodes(layer_name)[0]
assert new_node.type == 'Conv2D' or new_node.type == 'Conv2D_Pooling', 'must wide a conv'
new_node.config['filters'] = new_width
# logger.debug(new_graph.to_json())
# construct model
new_model = MyModel(config=config, graph=new_graph)
self.copy_weight(model, new_model)
# inherit weight
w_conv1, b_conv1 = model.get_layers(layer_name)[0].get_weights()
w_conv2, b_conv2 = model.get_layers(layer_name,
next_layer=True,
type=['Conv2D', 'Conv2D_Pooling'
])[0].get_weights()
new_w_conv1, new_b_conv1, new_w_conv2 = Net2Net._wider_conv2d_weight(
w_conv1, b_conv1, w_conv2, new_width, "net2wider")
new_model.get_layers(layer_name)[0].set_weights(
[new_w_conv1, new_b_conv1])
new_model.get_layers(layer_name,
next_layer=True,
type=['Conv2D', 'Conv2D_Pooling'
])[0].set_weights([new_w_conv2, b_conv2])
return new_model
def deeper_conv2d(self,
model,
layer_name,
config,
with_pooling,
kernel_size=3,
filters='same'):
# construct graph
new_graph = model.graph.copy()
if with_pooling == False:
type = 'Conv2D'
else:
type = 'Conv2D_Pooling'
new_node = Node(type=type,
name='new',
config={
'kernel_size': kernel_size,
'filters': filters
})
new_graph.deeper(layer_name, new_node)
# logger.debug(new_graph.to_json())
# construct model
new_model = MyModel(config=config, graph=new_graph)
# inherit weight
# in fact there is no need to get w_conv1 and b_conv1
# what we actually need is only w_conv1's shape
# more specifically, only kh, kw and filters are needed, num_channel is not necessary
node_type = model.graph.get_nodes(layer_name)[0].type
if node_type == 'Conv2D' or node_type == 'Conv2D_Pooling':
w_conv1, b_conv1 = new_model.get_layers(
layer_name)[0].get_weights()
# tensorflow kernel format: [filter_height, filter_width, in_channels, out_channels] channels_last
# theano kernel format: [output_channels, input_channels, filter_rows, filter_columns] channels_first
if K.image_data_format() == "channels_first": # theano format
# convert_kernel function Converts a Numpy kernel matrix from Theano format to TensorFlow format, vise versa
w_conv1 = convert_kernel(w_conv1)
kh, kw, num_channel, filters = w_conv1.shape
elif node_type == 'Group':
kh, kw, filters = 3, 3, model.graph.get_nodes(
layer_name)[0].config['filters']
w_conv2, b_conv2 = new_model.get_layers(
layer_name, next_layer=True)[0].get_weights()
new_w_conv2, new_b_conv2 = Net2Net._deeper_conv2d_weight(
kh=kh, kw=kw, filters=filters)
new_model.get_layers(layer_name, next_layer=True)[0].set_weights(
[new_w_conv2, new_b_conv2])
self.copy_weight(model, new_model)
return new_model
def wider_group_conv2d(self, model, layer_name, new_width, config):
# modify graph
new_graph = model.graph.copy()
node = new_graph.get_nodes(layer_name)[0]
assert node.type == 'Group', 'must wide a group layer'
node.config['filters'] = new_width
new_model = MyModel(config=config, graph=new_graph)
self.copy_weight(model, new_model)
filter_num = int(new_width / node.config['group_num'])
for i in range(node.config['group_num']):
if i == node.config['group_num'] - 1: # last group
filter_num = new_width - i * int(
(new_width / node.config['group_num']))
layer_name_conv_1 = layer_name + '_conv2d_' + str(i) + '_1'
layer_name_conv_2 = layer_name + '_conv2d_' + str(i) + '_2'
# use keras get_layer function
w_conv1, b_conv1 = model.model.get_layer(
layer_name_conv_1).get_weights()
w_conv2, b_conv2 = model.model.get_layer(
layer_name_conv_2).get_weights()
new_w_conv1, new_b_conv1, new_w_conv2 = Net2Net._wider_conv2d_weight(
w_conv1, b_conv1, w_conv2, filter_num, "net2wider")
new_model.model.get_layer(layer_name_conv_1).set_weights(
[new_w_conv1, new_b_conv1])
# new_model.model.get_layer(layer_name_conv_2).set_weights([new_w_conv2, b_conv2])
if i == 0:
w_conv_group = new_w_conv2
b_conv_group = b_conv2
else:
w_conv_group = np.concatenate((w_conv_group, new_w_conv2),
axis=3)
b_conv_group = np.concatenate((b_conv_group, b_conv2), axis=0)
# find group layer's next conv layer
w_conv2, b_conv2 = model.get_layers(
layer_name,
next_layer=True,
type=['Conv2D', 'Conv2D_Pooling', 'Group'])[0].get_weights()
new_w_conv_group, new_b_conv_group, new_w_conv2 = Net2Net._wider_conv2d_weight(
w_conv_group, b_conv_group, w_conv2, new_width, "net2wider")
for i in range(node.config['group_num']):
layer_name_conv_2 = layer_name + '_conv2d_' + str(i) + '_2'
if i == node.config['group_num'] - 1:
new_model.model.get_layer(layer_name_conv_2).set_weights([
new_w_conv_group[:, :, :, i * filter_num:],
new_b_conv_group[i * filter_num:]
])
else:
new_model.model.get_layer(layer_name_conv_2).set_weights([
new_w_conv_group[:, :, :,
i * filter_num:(i + 1) * filter_num],
new_b_conv_group[i * filter_num:(i + 1) * filter_num]
])
new_model.get_layers(layer_name,
next_layer=True,
type=['Conv2D', 'Conv2D_Pooling', 'Group'
])[0].set_weights([new_w_conv2, b_conv2])
return new_model
