class LinearAC(LinearOnly):
ac = ActivationParam()
def __init__(
self,
start_shape,
stop_shape,
min_features=10,
max_features=128,
depth=2,
bias_prob=0.0,
):
LinearOnly.__init__(
self, start_shape, stop_shape, min_features, max_features, depth, bias_prob
)
self.use_ac = BinaryParam(name="")
self.use_ac.randomize(true_prob=0.7)
def generate(self):
limits = self.limits
unit_list = []
for j in range(self.depth):
if j == 0:
in_shape = self.start_shape
else:
in_shape = unit_list[-1].out_shape
if j == self.depth - 1:
out_shape = self.stop_shape
else:
out_shape = None
unit_list.append(self.layers[0](in_shape, out_shape))
if self.use_ac.value and out_shape is None:
unit_list.append(
self.ac.val(unit_list[-1].out_shape, unit_list[-1].out_shape)
)
self.layers[0].out_features.limits = (limits[0], unit_list[-1].out_shape[0])
self.layers[0].out_features.limits = limits
return unit_list
def __call__(self, num_nets, startnum=1):
self.name = "Linear{}".format(self.ac.val.__class__.__name__)
return _Net.__call__(self, num_nets, startnum)
class ConvAC(ConvOnly):
ac = ActivationParam()
def __init__(self, start_shape, stop_shape, depth, convclass, bias_prob):
ConvOnly.__init__(self, start_shape, stop_shape, depth, convclass,
bias_prob)
self.use_ac = BinaryParam(name="")
self.use_ac.randomize(true_prob=0.7)
def generate(self, resnet_cover=256):
t = len(self.start_shape) - 1
self.layers[0].kernel_size.randomize(
limits=((1, ) * t,
(min(list(self.start_shape[1:]) + [resnet_cover]), ) * t))
limits = self.layers[0].kernel_size.limits
unit_list = []
for j in range(self.depth):
if j == 0:
in_shape = self.start_shape
else:
in_shape = unit_list[-1].out_shape
if j == self.depth - 1:
out_shape = self.stop_shape
else:
out_shape = None
unit_list.append(self.layers[0](in_shape, out_shape))
if self.use_ac.value and out_shape is None:
unit_list.append(
self.ac.val(unit_list[-1].out_shape,
unit_list[-1].out_shape))
out_shape = unit_list[-1].out_shape
t = len(out_shape) - 1
self.layers[0].kernel_size.randomize(
limits=((1, ) * t, (min(out_shape[1:]), ) * t))
self.layers[0].kernel_size.randomize(limits=limits)
return unit_list
def __call__(self, num_nets, startnum=1):
self.name = "{}".format(
self.__class__.__name__.replace("AC",
self.ac.val.__class__.__name__))
return _Net.__call__(self, num_nets, startnum)