def _make_convs(self):
head_width = self._hyper_params['head_width']
# feature adjustment
self.r_z_k = conv_bn_relu(head_width,
head_width,
1,
3,
0,
has_relu=False)
self.c_z_k = conv_bn_relu(head_width,
head_width,
1,
3,
0,
has_relu=False)
self.r_x = conv_bn_relu(head_width,
head_width,
1,
3,
0,
has_relu=False)
self.c_x = conv_bn_relu(head_width,
head_width,
1,
3,
0,
has_relu=False)
def __init__(self, block=creat_residual_block):
super(ResNet18_M, self).__init__()
self.block = block
self.stage1 = nn.Sequential(
conv_bn_relu(3,
32,
stride=2,
kszie=3,
pad=3,
has_bn=True,
has_relu=True,
bias=False),
conv_bn_relu(32,
32,
stride=1,
kszie=3,
pad=1,
has_bn=True,
has_relu=True,
bias=False),
conv_bn_relu(32,
32,
stride=1,
kszie=3,
pad=1,
has_bn=True,
has_relu=True,
bias=False), nn.MaxPool2d(3, 2, 1, ceil_mode=False))
self.stage2 = self.__make_stage(self.block, 32, 64, 2, 1)
self.stage3 = self.__make_stage(self.block, 64, 128, 2, 2)
self.stage4 = self.__make_stage(self.block, 128, 256, 2, 2)
self.stage5 = self.__make_stage(self.block, 256, 256, 2, 2)
def __init__(self):
super(TinyConv, self).__init__()
self.conv1 = conv_bn_relu(3, 32, stride=2, kszie=3, pad=0)
self.pool1 = nn.MaxPool2d(3, stride=2, padding=0, ceil_mode=True)
self.conv2a = conv_bn_relu(32, 64, stride=1, kszie=1, pad=0)
self.conv2b = conv_bn_relu(64, 64, stride=2, kszie=7, pad=0, groups=64)
self.conv3a = conv_bn_relu(64, 64, stride=1, kszie=3, pad=0)
self.conv3b = conv_bn_relu(64,
64,
stride=1,
kszie=1,
pad=0,
has_relu=False)
# initialization
for m in self.modules():
if isinstance(m, nn.Conv2d) or isinstance(m, nn.Linear):
import scipy.stats as stats
stddev = m.stddev if hasattr(m, 'stddev') else 0.1
X = stats.truncnorm(-2, 2, scale=stddev)
values = torch.as_tensor(X.rvs(m.weight.numel()),
dtype=m.weight.dtype)
values = values.view(m.weight.size())
with torch.no_grad():
m.weight.copy_(values)
elif isinstance(m, nn.BatchNorm2d):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def __init__(self):
super(DenseboxHead, self).__init__()
self.cls_score_p5 = conv_bn_relu(256,
1,
stride=1,
kszie=1,
pad=0,
has_relu=False)
self.ctr_score_p5 = conv_bn_relu(256,
1,
stride=1,
kszie=1,
pad=0,
has_relu=False)
self.bbox_offsets_p5 = conv_bn_relu(256,
4,
stride=1,
kszie=1,
pad=0,
has_relu=False)
self.bi = torch.nn.Parameter(torch.tensor(0.).type(torch.Tensor))
self.si = torch.nn.Parameter(torch.tensor(1.).type(torch.Tensor))
self.cls_convs = []
self.bbox_convs = []
def __init__(self, inplanes, outplanes, stride, has_proj=False):
super(creat_residual_block, self).__init__()
self.has_proj = has_proj
if self.has_proj:
self.proj_conv = conv_bn_relu(inplanes,
outplanes,
stride=stride,
kszie=1,
pad=0,
has_bn=True,
has_relu=False,
bias=False)
self.conv1 = conv_bn_relu(inplanes,
outplanes,
stride=stride,
kszie=3,
pad=1,
has_bn=True,
has_relu=True,
bias=False)
self.conv2 = conv_bn_relu(outplanes,
outplanes,
stride=1,
kszie=3,
pad=1,
has_bn=True,
has_relu=False,
bias=False)
self.relu = nn.ReLU()
def __init__(self):
super(AlexNet, self).__init__()
self.conv1 = conv_bn_relu(3, 96, stride=2, kszie=11, pad=0)
self.pool1 = nn.MaxPool2d(3, 2, 0, ceil_mode=True)
self.conv2 = conv_bn_relu(96, 256, 1, 5, 0)
self.pool2 = nn.MaxPool2d(3, 2, 0, ceil_mode=True)
self.conv3 = conv_bn_relu(256, 384, 1, 3, 0)
self.conv4 = conv_bn_relu(384, 384, 1, 3, 0)
self.conv5 = conv_bn_relu(384, 256, 1, 3, 0, has_relu=False)
def __init__(self, backbone, head, loss):
super(SiamTrack, self).__init__()
self.basemodel = backbone
# feature adjustment
self.r_z_k = conv_bn_relu(256, 256, 1, 3, 0, has_relu=False)
self.c_z_k = conv_bn_relu(256, 256, 1, 3, 0, has_relu=False)
self.r_x = conv_bn_relu(256, 256, 1, 3, 0, has_relu=False)
self.c_x = conv_bn_relu(256, 256, 1, 3, 0, has_relu=False)
# head
self.head = head
# loss
self.loss = loss
# initialze head
conv_list = [
self.r_z_k.conv, self.c_z_k.conv, self.r_x.conv, self.c_x.conv
]
for ith in range(len(conv_list)):
conv = conv_list[ith]
torch.nn.init.normal_(conv.weight, std=0.01)
def _make_conv_output(self):
head_width = self._hyper_params['head_width']
self.cls_score_p5 = conv_bn_relu(head_width,
1,
stride=1,
kszie=1,
pad=0,
has_relu=False)
self.ctr_score_p5 = conv_bn_relu(head_width,
1,
stride=1,
kszie=1,
pad=0,
has_relu=False)
self.bbox_offsets_p5 = conv_bn_relu(head_width,
4,
stride=1,
kszie=1,
pad=0,
has_relu=False)
def _make_conv_output(self):
head_width = self._hyper_params['head_width']
self.cls_score_p5 = conv_bn_relu(head_width,
1,
stride=1,
kszie=1,
pad=0,
has_relu=False)
self.cls_score2_p5 = conv_bn_relu_withweight(256,
1,
stride=1,
kszie=1,
pad=0,
has_relu=False)
self.ctr_score_p5 = conv_bn_relu(head_width,
1,
stride=1,
kszie=1,
pad=0,
has_relu=False)
'''
self.ctr_score2_p5 = conv_bn_relu_withweight(256,
1,
stride=1,
kszie=1,
pad=0,
has_relu=False)
'''
self.bbox_offsets_p5 = conv_bn_relu(head_width,
4,
stride=1,
kszie=1,
pad=0,
has_relu=False)
self.bbox_offsets2_p5 = conv_bn_relu_withweight(256,
4,
stride=1,
kszie=3,
pad=1,
has_bn=True,
has_relu=False)
def _make_convs(self):
head_width = self._hyper_params['head_width']
neck_conv_bias = self._hyper_params['neck_conv_bias']
# feature adjustment
self.r_z_k = conv_bn_relu(
head_width,
head_width,
1,
3,
0,
has_relu=False,
bias=neck_conv_bias[0],
#has_bn = False
)
self.r_x = conv_bn_relu(head_width,
head_width,
1,
3,
0,
has_relu=False,
bias=neck_conv_bias[1])
self.c_z_k = conv_bn_relu(head_width,
head_width,
1,
3,
0,
has_relu=False,
bias=neck_conv_bias[2])
self.c_x = conv_bn_relu(head_width,
head_width,
1,
3,
0,
has_relu=False,
bias=neck_conv_bias[3])
def _make_conv3x3(self):
num_conv3x3 = self._hyper_params['num_conv3x3']
head_conv_bn = self._hyper_params['head_conv_bn']
self.cls_conv3x3_list = []
self.bbox_conv3x3_list = []
for i in range(num_conv3x3):
# is_last_conv = (i >= num_conv3x3)
cls_conv3x3 = conv_bn_relu(256,
256,
stride=1,
kszie=3,
pad=0,
has_bn=head_conv_bn[i])
bbox_conv3x3 = conv_bn_relu(256,
256,
stride=1,
kszie=3,
pad=0,
has_bn=head_conv_bn[i])
setattr(self, 'cls_p5_conv%d' % (i + 1), cls_conv3x3)
setattr(self, 'bbox_p5_conv%d' % (i + 1), bbox_conv3x3)
self.cls_conv3x3_list.append(cls_conv3x3)
self.bbox_conv3x3_list.append(bbox_conv3x3)
