def _init_param(self):
nu.init_module(self.Q_noise)
nu.init_module(self.R_noise)
nu.init_module(self.pred2loc)
nu.init_lstm_module(self.pred_lstm)
nu.init_lstm_module(self.Q_lstm)
nu.init_lstm_module(self.R_lstm)
def _init_param(self):
nu.init_module(self.loc2feat)
nu.init_module(self.vel2feat)
nu.init_module(self.pred2loc)
nu.init_lstm_module(self.pred_lstm)
nu.init_lstm_module(self.refine_lstm)
def __init__(self, arch_name, roi_name, down_ratio, roi_kernel):
super(Model, self).__init__()
self.base = dla_up.__dict__[arch_name](
pretrained_base='imagenet', down_ratio=down_ratio)
num_channel = self.base.channels[int(np.log2(down_ratio))]
# We use roialign with kernel size = 7 in our experiments
assert ('align' in roi_name or 'pool' in roi_name)
assert (roi_kernel == 7)
if 'align' in roi_name:
print('Using RoIAlign')
self.roi_pool = ROIAlign(
(roi_kernel, roi_kernel),
1.0 / down_ratio,
0)
elif 'pool' in roi_name:
print('Using RoIPool')
self.roi_pool = ROIPool(
(roi_kernel, roi_kernel),
1.0 / down_ratio)
self.dim = nn.Sequential(
nn.Conv2d(num_channel, num_channel,
kernel_size=3, stride=1, padding=0, bias=False),
nn.BatchNorm2d(num_channel),
nn.ReLU(inplace=True),
nn.Conv2d(num_channel, num_channel,
kernel_size=3, stride=1, padding=0, bias=False),
nn.BatchNorm2d(num_channel),
nn.ReLU(inplace=True),
nn.Conv2d(num_channel, num_channel,
kernel_size=3, stride=1, padding=0, bias=False),
nn.BatchNorm2d(num_channel),
nn.ReLU(inplace=True),
nn.Conv2d(num_channel, 3, kernel_size=1,
stride=1, padding=0, bias=True)) # 3 dim
self.rot = nn.Sequential(
nn.Conv2d(num_channel, num_channel,
kernel_size=3, stride=1, padding=0, bias=False),
nn.BatchNorm2d(num_channel),
nn.ReLU(inplace=True),
nn.Conv2d(num_channel, num_channel,
kernel_size=3, stride=1, padding=0, bias=False),
nn.BatchNorm2d(num_channel),
nn.ReLU(inplace=True),
nn.Conv2d(num_channel, num_channel,
kernel_size=3, stride=1, padding=0, bias=False),
nn.BatchNorm2d(num_channel),
nn.ReLU(inplace=True),
nn.Conv2d(num_channel, 8, kernel_size=1,
stride=1, padding=0, bias=True)) # 1 + 1 + 2
self.dep = nn.Sequential(
nn.Conv2d(num_channel, num_channel,
kernel_size=3, stride=1, padding=0, bias=False),
nn.BatchNorm2d(num_channel),
nn.ReLU(inplace=True),
nn.Conv2d(num_channel, num_channel,
kernel_size=3, stride=1, padding=0, bias=False),
nn.BatchNorm2d(num_channel),
nn.ReLU(inplace=True),
nn.Conv2d(num_channel, num_channel,
kernel_size=3, stride=1, padding=0, bias=False),
nn.BatchNorm2d(num_channel),
nn.ReLU(inplace=True),
nn.Conv2d(num_channel, 1, kernel_size=1,
stride=1, padding=0, bias=True),
nn.Sigmoid())
self.cen = nn.Sequential(
nn.Conv2d(num_channel, num_channel,
kernel_size=3, stride=1, padding=0, bias=False),
nn.BatchNorm2d(num_channel),
nn.ReLU(inplace=True),
nn.Conv2d(num_channel, num_channel,
kernel_size=3, stride=1, padding=0, bias=False),
nn.BatchNorm2d(num_channel),
nn.ReLU(inplace=True),
nn.Conv2d(num_channel, num_channel,
kernel_size=3, stride=1, padding=0, bias=False),
nn.BatchNorm2d(num_channel),
nn.ReLU(inplace=True),
nn.Conv2d(num_channel, 2, kernel_size=1,
stride=1, padding=0, bias=True))
nu.init_module(self.base)
nu.init_module(self.dim)
nu.init_module(self.rot)
nu.init_module(self.dep)
nu.init_module(self.cen)