def __init__(self,
n_classes,
bc_learning=True,
nobias=False,
dr_ratio=0.5):
super(ConvNet, self).__init__()
self.dr_ratio = dr_ratio
self.bc_learning = bc_learning
if self.bc_learning:
self.loss = kl_divergence
else:
self.loss = F.softmax_cross_entropy
# architecture
kwargs = {'ksize': 3, 'stride': 1, 'pad': 1, 'nobias': nobias}
with self.init_scope():
self.conv1_1 = Conv2DBNActiv(3, 64, **kwargs)
self.conv1_2 = Conv2DBNActiv(64, 64, **kwargs)
self.conv2_1 = Conv2DBNActiv(64, 128, **kwargs)
self.conv2_2 = Conv2DBNActiv(128, 128, **kwargs)
self.conv3_1 = Conv2DBNActiv(128, 256, **kwargs)
self.conv3_2 = Conv2DBNActiv(256, 256, **kwargs)
self.conv3_3 = Conv2DBNActiv(256, 256, **kwargs)
self.conv3_4 = Conv2DBNActiv(256, 256, **kwargs)
self.fc4 = L.Linear(1024,
initialW=Uniform(1. / math.sqrt(256 * 4 * 4)))
self.fc5 = L.Linear(1024, initialW=Uniform(1. / math.sqrt(1024)))
self.fc6 = L.Linear(n_classes,
initialW=Uniform(1. / math.sqrt(1024)))
def __init__(self, n_in, n_hidden):
super(DependentSelectionLayer, self).__init__()
# input dim is incoming feature size (n_in) + previous selection z (1)
with self.init_scope():
self.rnn = RCNN(n_in + 1, n_hidden, 2)
self.l = L.Linear(
n_in + n_hidden, 1, initialW=Uniform(0.05),
initial_bias=Uniform(0.05))
def __init__(self, n_classes):
super(ConvNet, self).__init__(
conv11=ConvBNReLU(3, 64, 3, pad=1),
conv12=ConvBNReLU(64, 64, 3, pad=1),
conv21=ConvBNReLU(64, 128, 3, pad=1),
conv22=ConvBNReLU(128, 128, 3, pad=1),
conv31=ConvBNReLU(128, 256, 3, pad=1),
conv32=ConvBNReLU(256, 256, 3, pad=1),
conv33=ConvBNReLU(256, 256, 3, pad=1),
conv34=ConvBNReLU(256, 256, 3, pad=1),
fc4=L.Linear(256 * 4 * 4, 1024, initialW=Uniform(1. / math.sqrt(256 * 4 * 4))),
fc5=L.Linear(1024, 1024, initialW=Uniform(1. / math.sqrt(1024))),
fc6=L.Linear(1024, n_classes, initialW=Uniform(1. / math.sqrt(1024)))
)