def __init__(self, in_planes, planes, stride=1, test=False):
super(BasicBlock, self).__init__()
self.test = test
self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(planes)
self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(planes)
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != self.expansion*planes:
self.shortcut = nn.Sequential(
nn.Conv2d(in_planes, self.expansion*planes, kernel_size=1, stride=stride, bias=False),
nn.BatchNorm2d(self.expansion*planes)
)
# Gate layers
self.fc1 = nn.Conv2d(in_planes, 16, kernel_size=1)
self.fc1bn = nn.BatchNorm1d(16)
self.fc2 = nn.Conv2d(16, 2, kernel_size=1)
# initialize the bias of the last fc for
# initial opening rate of the gate of about 85%
self.fc2.bias.data[0] = 0.1
self.fc2.bias.data[1] = 2
self.gs = GumbleSoftmax()
self.gs.cuda()
def __init__(self, num_gates_fixed_open, num_gates, num_filters_per_gate):
super(SpecialGumble, self).__init__()
self.num_gates_fixed_open = num_gates_fixed_open
self.num_gates = num_gates
self.num_filters_per_gate = num_filters_per_gate
self.gs = GumbleSoftmax()
def __init__(self, in_planes, planes, stride=1, test=False):
super(Bottleneck, self).__init__()
self.test = test
self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=1, bias=False)
self.bn1 = nn.BatchNorm2d(planes)
self.conv2 = nn.Conv2d(planes,
planes,
kernel_size=3,
stride=stride,
padding=1,
bias=False)
self.bn2 = nn.BatchNorm2d(planes)
self.conv3 = nn.Conv2d(planes,
self.expansion * planes,
kernel_size=1,
bias=False)
self.bn3 = nn.BatchNorm2d(self.expansion * planes)
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != self.expansion * planes:
self.shortcut = nn.Sequential(
nn.Conv2d(in_planes,
self.expansion * planes,
kernel_size=1,
stride=stride,
bias=False), nn.BatchNorm2d(self.expansion * planes))
# Gate layers
self.w = nn.Parameter(torch.cuda.FloatTensor([.1, 4]).view((2, 1, 1)))
self.gs = GumbleSoftmax()
self.gs.cuda()
def __init__(self, shape, unit_test_init=False):
super(GumbleRelu, self).__init__()
self.gs = GumbleSoftmax()
self.fc1_weights = nn.Parameter(torch.zeros(
(1, shape[1], shape[2], shape[3], 1)),
requires_grad=True)
self.fc1_bias_initial = nn.Parameter(torch.zeros(
(1, shape[1], shape[2], shape[3], 1)),
requires_grad=True)
self.fc1_bias = nn.Parameter(torch.zeros(
(1, shape[1], shape[2], shape[3], 2)),
requires_grad=True)
if unit_test_init:
self.fc1_weights.data.fill_(0.1)
self.fc1_bias.data.fill_(0.1)
else:
torch.nn.init.xavier_uniform(self.fc1_weights)
self.fc1_bias.data[:, :, :, :, 1] = 4