def from_config(cls, config):
# Use layer build function to initialise new NeuralGraphOutput
inner_layer_config = config.pop('inner_layer_config')
create_inner_layer_fn = lambda: layer_from_config(
deepcopy(inner_layer_config))
layer = cls(create_inner_layer_fn, **config)
return layer
def __init__(self, inner_layer_arg, **kwargs):
# Initialise based on one of the three initialisation methods
# Case 1: Check if inner_layer_arg is conv_width
if isinstance(inner_layer_arg, (int)):
self.conv_width = inner_layer_arg
# Keras2: we assume all the kwargs should be passed to the Dense layer
# dense_layer_kwargs, kwargs = filter_func_args(layers.Dense.__init__,
# kwargs, overrule_args=['name'])
self.create_inner_layer_fn = lambda: layers.Dense(
self.conv_width, **kwargs) #dense_layer_kwargs)
# Case 2: Check if an initialised keras layer is given
elif isinstance(inner_layer_arg, layers.Layer):
assert inner_layer_arg.built == False, 'When initialising with a keras layer, it cannot be built.'
_, self.conv_width = inner_layer_arg.compute_output_shape(
(None, 1))
# layer_from_config will mutate the config dict, therefore create a get fn
self.create_inner_layer_fn = lambda: layer_from_config(
dict(class_name=inner_layer_arg.__class__.__name__,
config=inner_layer_arg.get_config()))
# Case 3: Check if a function is provided that returns a initialised keras layer
elif callable(inner_layer_arg):
example_instance = inner_layer_arg()
assert isinstance(
example_instance, layers.Layer
), 'When initialising with a function, the function has to return a keras layer'
assert example_instance.built == False, 'When initialising with a keras layer, it cannot be built.'
_, self.conv_width = example_instance.compute_output_shape(
(None, 1))
self.create_inner_layer_fn = inner_layer_arg
else:
raise ValueError(
'NeuralGraphHidden has to be initialised with 1). int conv_widht, 2). a keras layer instance, or 3). a function returning a keras layer instance.'
)
super(NeuralGraphHidden, self).__init__(
) # Keras2: all the kwargs will be passed to the Dense layer only
def create_mobilenet_v2(input_shape=(32, 32, 3),
num_classes=10,
is_pretrained=False,
mobilenet_shape=(224, 224, 3),
num_change_strides=0,
seperate_softmax=True):
assert (input_shape[0] == input_shape[1])
assert (224 % input_shape[0] == 0)
mobilenet = None
if is_pretrained:
mobilenet = MobileNetV2(input_shape=mobilenet_shape,
include_top=True,
alpha=1.0,
weights='imagenet',
seperate_softmax=seperate_softmax,
backend=keras.backend,
layers=keras.layers,
models=keras.models,
utils=keras.utils)
return mobilenet
else:
mobilenet = MobileNetV2(input_shape=mobilenet_shape,
include_top=True,
weights=None,
classes=num_classes,
seperate_softmax=seperate_softmax,
output_stride=None,
backend=keras.backend,
layers=keras.layers,
models=keras.models,
utils=keras.utils)
scale_factor = mobilenet_shape[0] // input_shape[0]
output_residual = queue.Queue(2)
input_net = Input(input_shape)
if scale_factor > 1:
x = UpSampling2D((scale_factor, scale_factor))(input_net)
else:
x = input_net
for layer in mobilenet.layers[1:-2]:
config = layer.get_config()
# check if stride has to be changed
if num_change_strides > 0:
if layer.__class__ == keras.layers.ZeroPadding2D:
continue
if 'strides' in config.keys():
if config['strides'] == (2, 2):
config['strides'] = (1, 1)
config['padding'] = 'same'
num_change_strides -= 1
next_layer = layer_from_config({
'class_name': layer.__class__.__name__,
'config': config
})
if isinstance(layer, Add):
x = next_layer([x, output_residual.get()])
else:
x = next_layer(x)
if "_project_BN" in layer.name:
if output_residual.full():
output_residual.get()
output_residual.put(x)
if "_add" in layer.name:
if output_residual.full():
output_residual.get()
output_residual.put(x)
x = Dropout(rate=0.2)(x)
x = Dense(num_classes, activation=None)(x)
x = Activation('softmax')(x)
return Model(inputs=input_net, outputs=x, name='mobilenetv2')
def create_inner_layer_fn():
return layer_from_config(deepcopy(inner_layer_config))