def __init__(self, batchnorm_pool = False, **kwargs):
super(NegConvBNRelu, self).__init__(**kwargs)
self.conv1 = nn.Conv2D(channels=64, kernel_size=(3, 3), strides=(1,1), use_bias=False)
self.bn = nn.BatchNorm()
self.act = nn.Activation('relu')
self.pool = nn.AvgPool2D(pool_size=(4,4))
self.tailneg = TailNegBlock()
self.batchnorm_pool = batchnorm_pool
def __init__(self, **kwargs):
super(NegConvAdd, self).__init__(**kwargs)
self.conv1 = nn.Conv2D(channels=64, kernel_size=(3, 3), strides=(1,1), use_bias=False)
self.act = nn.Activation('relu')
self.pool = nn.AvgPool2D(pool_size=(4,4))
self.tailneg = TailNegBlock()
self.add_value = mx.gluon.Parameter('add_value', init=mx.init.Xavier(magnitude=2.24),
dtype='float32', allow_deferred_init=True)
def __init__(self, **kwargs):
super(NegConvBN, self).__init__(**kwargs)
self.conv1 = nn.Conv2D(channels=64,
kernel_size=(3, 3),
strides=(1, 1),
use_bias=False)
self.bn1 = nn.BatchNorm()
self.pool = nn.AvgPool2D(pool_size=(4, 4))
self.tailneg = TailNegBlock()
def __init__(self, use_bias, flatten, **kwargs):
super(NegFCReLU, self).__init__(**kwargs)
self.fc = nn.Dense(units=64, use_bias=use_bias, flatten=flatten)
self.act1 = nn.Activation('relu')
self.act2 = nn.Activation('sigmoid')
self.tail_neg = TailNegBlock()