def layers(self):
bn = True
return [
Conv((7, 7, 96), init=Kaiming(), activation=Explin(), batch_norm=bn, padding=3, strides=1)\
if self.bn_first_layer else\
Conv((7, 7, 96), init=Kaiming(), bias=Constant(0), activation=Explin(), padding=3, strides=1),
Pooling(3, strides=2, padding=1),
Conv((7, 7, 128), init=Kaiming(), activation=Explin(), batch_norm=bn, padding=3, strides=1),
Pooling(3, strides=2, padding=1),
Conv((5, 5, 256), init=Kaiming(), activation=Explin(), batch_norm=bn, padding=2, strides=1),
Pooling(3, strides=2, padding=1),
Conv((3, 3, 384), init=Kaiming(), activation=Explin(), batch_norm=bn, padding=1, strides=1),
Conv((3, 3, 384), init=Kaiming(), activation=Explin(), batch_norm=bn, padding=1, strides=1),
Conv((3, 3, 384), init=Kaiming(), activation=Explin(), batch_norm=bn, padding=1, strides=1),
Pooling(3, strides=2, padding=1, op='avg'),
Affine(nout=self.noutputs, init=Kaiming(), activation=Explin(), batch_norm=bn),
Affine(nout=self.noutputs, init=Kaiming(), activation=Explin(), batch_norm=bn),
Affine(nout=self.noutputs, init=Kaiming(), bias=Constant(0),
activation=Softmax() if self.use_softmax else Logistic(shortcut=True))
]
def layers(self):
bn = True
return [
# input 128
Conv((7, 7, 96),
init=Kaiming(),
bias=Constant(0),
activation=Explin(),
padding=3,
strides=1),
Pooling(3, strides=2, padding=1),
# 64
Conv((7, 7, 128),
init=Kaiming(),
activation=Explin(),
batch_norm=bn,
padding=3,
strides=1),
Pooling(3, strides=2, padding=1),
# 32
Conv((5, 5, 256),
init=Kaiming(),
activation=Explin(),
batch_norm=bn,
padding=2,
strides=1),
Pooling(3, strides=2, padding=1),
# 16
Conv((3, 3, 384),
init=Kaiming(),
activation=Explin(),
batch_norm=bn,
padding=1,
strides=1),
Conv((3, 3, 384),
init=Kaiming(),
activation=Explin(),
batch_norm=bn,
padding=1,
strides=1),
Conv((3, 3, 384),
init=Kaiming(),
activation=Explin(),
batch_norm=bn,
padding=1,
strides=1),
Pooling(3, strides=2, padding=1),
# 8
Conv((3, 3, 8192),
init=Kaiming(),
activation=Explin(),
batch_norm=bn,
padding=1,
strides=1),
Pooling('all', op='avg'),
Affine(nout=self.noutputs,
init=Kaiming(),
bias=Constant(0),
activation=Softmax() if self.use_softmax else Logistic(
shortcut=True))
]
def layers(self):
bn = True
return [
# input 128
Conv((7, 7, 64),
init=Kaiming(),
bias=Constant(0),
activation=Explin(),
padding=3,
strides=1),
Pooling(3, strides=2, padding=1),
# 64
Conv((3, 3, 96),
init=Kaiming(),
activation=Explin(),
batch_norm=bn,
padding=1,
strides=1),
Conv((3, 3, 96),
init=Kaiming(),
activation=Explin(),
batch_norm=bn,
padding=1,
strides=1),
Pooling(3, strides=2, padding=1),
# 32
Conv((3, 3, 192),
init=Kaiming(),
activation=Explin(),
batch_norm=bn,
padding=1,
strides=1),
Conv((3, 3, 192),
init=Kaiming(),
activation=Explin(),
batch_norm=bn,
padding=1,
strides=1),
Pooling(3, strides=2, padding=1),
# 16
Conv((3, 3, 384),
init=Kaiming(),
activation=Explin(),
batch_norm=bn,
padding=1,
strides=1),
Conv((3, 3, 384),
init=Kaiming(),
activation=Explin(),
batch_norm=bn,
padding=1,
strides=1),
Conv((3, 3, 384),
init=Kaiming(),
activation=Explin(),
batch_norm=bn,
padding=1,
strides=1),
# this 4th deep layer may have been in for vgg3pool64all run? can not fit for 6fold so commented
#Conv((3, 3, 384), init=Kaiming(), activation=Explin(), batch_norm=bn, padding=1, strides=1),
Pooling(3, strides=2, padding=1),
# 8
Conv((3, 3, 6144),
init=Kaiming(),
activation=Explin(),
batch_norm=bn,
padding=1,
strides=1),
Pooling('all', op='avg'),
Affine(nout=self.noutputs,
init=Kaiming(),
bias=Constant(0),
activation=Softmax() if self.use_softmax else Logistic(
shortcut=True))
]