def __init__(self, in_features, out_features):#num_units=4096): super(BNN_1blk_100, self).__init__() # self.infl_ratio=1 self.num_internal_blocks = 1 self.output_size = out_features self.input_size = in_features self.name = 'bnn_1blk_100' self.fc1 = BinaryLinear(self.input_size, 100*self.infl_ratio) self.bn1 = nn.BatchNorm1d(100*self.infl_ratio, eps=1e-4) self.htanh1 = BinaryStraightThrough() self.fc5 = BinaryLinear(100*self.infl_ratio, self.output_size) self.logsoftmax = nn.LogSoftmax()
def __init__(self, in_features, out_features, num_units=2048):
super(BinaryConnect, self).__init__()
self.net = nn.Sequential(
BinaryLinear(in_features, num_units),
nn.BatchNorm1d(num_units, eps=1e-4, momentum=0.15),
nn.ReLU(),
BinaryLinear(num_units, num_units),
nn.BatchNorm1d(num_units, eps=1e-4, momentum=0.15),
nn.ReLU(),
BinaryLinear(num_units, num_units),
nn.BatchNorm1d(num_units, eps=1e-4, momentum=0.15),
nn.ReLU(),
BinaryLinear(num_units, out_features),
nn.BatchNorm1d(out_features, eps=1e-4, momentum=0.15),
#nn.LogSoftmax()
)
def __init__(self, in_features, out_features, num_units=4096):
super(BinaryNet, self).__init__()
self.net = nn.Sequential(
nn.Dropout(p=0.2),
BinaryLinear(in_features, num_units),
nn.BatchNorm1d(num_units, eps=1e-4),
BinaryStraightThrough(),
nn.Dropout(),
BinaryLinear(num_units, num_units),
nn.BatchNorm1d(num_units, eps=1e-4),
BinaryStraightThrough(),
nn.Dropout(),
BinaryLinear(num_units, num_units),
nn.BatchNorm1d(num_units, eps=1e-4),
BinaryStraightThrough(),
nn.Dropout(),
BinaryLinear(num_units, 10),
nn.BatchNorm1d(10, eps=1e-4),
nn.LogSoftmax()
)
def __init__(self, topology, batch_norm):
super(BinaryConnect, self).__init__()
assert len(topology) > 1
for item in topology:
assert item > 0
layers = []
for dim_in, dim_out in zip(topology[:-1], topology[1:]):
layers.append(BinaryLinear(dim_in, dim_out))
if batch_norm:
layers.append(nn.BatchNorm1d(dim_out, eps=1e-4, momentum=0.15))
layers.append(nn.ReLU())
layers.pop()
layers.append(nn.LogSoftmax())
self.net = nn.Sequential(*layers)