def __init__(self, n_filters, n_out):
super(ConvNet, self).__init__()
self.n_filters = n_filters
self.n_out = n_out
with self.init_scope():
self.l1 = BL.ConvBNBST(n_filters, 3)
self.l2 = BL.BinaryLinearBNSoftmax(n_out)
def __init__(self, n_in, n_out):
super(AlexNet, self).__init__()
self.n_out = n_out
with self.init_scope():
self.conv1 = BL.ConvPoolBNBST(n_in,
10,
5,
stride=1,
pad=0,
pksize=3,
pstride=2)
self.conv2 = BL.BinaryConvPoolBNBST(10,
10,
5,
stride=1,
pad=2,
pksize=3,
pstride=2)
self.conv3 = BL.BinaryConvBNBST(10, 10, 3, stride=1, pad=1)
self.conv4 = BL.BinaryConvBNBST(10, 10, 3, stride=1, pad=1)
self.conv5 = BL.BinaryConvPoolBNBST(10,
10,
3,
stride=1,
pad=1,
pksize=3,
pstride=2)
self.fc8 = BL.BinaryLinearBNSoftmax(40, n_out)
def __init__(self, n_filters, n_out):
super(ConvNet, self).__init__()
self.n_filters = n_filters
self.n_out = n_out
with self.init_scope():
self.l1 = BL.ConvPoolBNBST(1,
n_filters,
3,
stride=2,
pksize=3,
pstride=2)
self.l2 = BL.BinaryLinearBNSoftmax(360, n_out)
def __init__(self, n_in, n_out, n_filters):
super(BinaryConvNet, self).__init__()
self.n_in = n_in
self.n_out = n_out
self.n_filters = n_filters
with self.init_scope():
self.l1 = BL.BinaryConvPoolBNBST(n_in,
n_filters,
3,
stride=2,
pksize=3,
pstride=2)
self.l2 = BL.BinaryLinearBNSoftmax(None, n_out)
def __init__(self, n_filters, n_out):
super(ConvNet, self).__init__()
self.n_filters = n_filters
self.n_out = n_out
with self.init_scope():
print '(net.py)-init ConvNet'
# if the number of intermediate output(3) increase, the accuracy will increase as well
self.l1 = BL.ConvPoolBNBST(n_filters, 3)
print ' self.l1 is ', self.l1
# we can also try
# self.l1 = BL.BinaryConvBNBST(n_filters, 3)
# Softmax is more computational expensive
self.l2 = BL.BinaryLinearBNSoftmax(n_out)
print ' self.l2 is ', self.l2