def __init__(self, class_num=8, dim_in=2048, trans_param_num=3, detach_network=False, pretrained=True, temporal_out=4, T=None, hist=1):
super(CycleTime, self).__init__()
dim = 512
#print("Pretrained imagenet:", pretrained)
resnet = resnet_res4s1.resnet18(pretrained=pretrained)
#print("RESNET 18:", resnet)
#resnet_50 = resnet_res4s1.resnet50(pretrained=pretrained)
#print("RESNET 50:", resnet_50)
self.encoderVideo = inflated_resnet.InflatedResNet(copy.deepcopy(resnet))
#print("INFLATED RESNET 18:", self.encoderVideo)
self.detach_network = detach_network
self.hist = hist
self.div_num = 512
self.T = self.div_num**-.5 if T is None else T
print('self.T:', self.T)
self.afterconv1 = nn.Conv3d(256, 18, kernel_size=1, bias=False)
self.spatial_out1 = 30
self.spatial_out2 = 10
self.temporal_out = temporal_out
self.afterconv3_trans = nn.Conv2d(self.spatial_out1 * self.spatial_out1, 128, kernel_size=4, padding=0, bias=False)
self.afterconv4_trans = nn.Conv2d(128, 64, kernel_size=4, padding=0, bias=False)
corrdim = 64 * 4 * 4
corrdim_trans = 64 * 4 * 4
self.linear2 = nn.Linear(corrdim_trans, trans_param_num)
self.leakyrelu = nn.LeakyReLU(0.1, inplace=True)
self.relu = nn.ReLU(inplace=True)
self.avgpool = nn.AvgPool2d(7, stride=1)
self.avgpool3d = nn.AvgPool3d((4, 2, 2), stride=(1, 2, 2))
self.maxpool2d = nn.MaxPool2d(2, stride=2)
# initialization
nn.init.kaiming_normal(self.afterconv1.weight, mode='fan_out')
nn.init.kaiming_normal(self.afterconv3_trans.weight, mode='fan_out')
nn.init.kaiming_normal(self.afterconv4_trans.weight, mode='fan_out')
# assuming no fc pre-training
for m in self.modules():
if isinstance(m, nn.Linear):
m.weight.data.normal_(0, 0.01)
m.bias.data.zero_()
# transformation
self.geometricTnf = GeometricTnfAffine(geometric_model='affine',
tps_grid_size=3,
tps_reg_factor=0.2,
out_h=self.spatial_out2, out_w=self.spatial_out2,
offset_factor=227/210)
xs = np.linspace(-1,1,80)
xs = np.meshgrid(xs, xs)
xs = np.stack(xs, 2)
self.xs = xs
self.criterion_inlier = WeakInlierCountPool(geometric_model='affine', tps_grid_size=3, tps_reg_factor=0.2, h_matches=30, w_matches=30, use_conv_filter=False, dilation_filter=0, normalize_inlier_count=True)
self.criterion_synth = TransformedGridLoss(use_cuda=True, geometric_model='affine')
def __init__(self, class_num=8, dim_in=2048, trans_param_num=6, detach_network=False, pretrained=True, temporal_out=4, T=None):
super(CycleTime, self).__init__()
dim = 512
print(pretrained)
resnet = resnet_res4s1.resnet50(pretrained=pretrained)
self.encoderVideo = inflated_resnet.InflatedResNet(copy.deepcopy(resnet))
self.detach_network = detach_network
self.div_num = 512
self.T = self.div_num**-.5 if T is None else T
print('self.T:', self.T)
# self.encoderVideo = resnet3d.resnet50(pretrained=False)
self.afterconv1 = nn.Conv3d(1024, 512, kernel_size=1, bias=False)
self.spatial_out1 = 30
self.spatial_out2 = 10
self.temporal_out = temporal_out
self.afterconv3_trans = nn.Conv2d(self.spatial_out1 * self.spatial_out1, 128, kernel_size=4, padding=0, bias=False)
self.afterconv4_trans = nn.Conv2d(128, 64, kernel_size=4, padding=0, bias=False)
corrdim = 64 * 4 * 4
corrdim_trans = 64 * 4 * 4
self.linear2 = nn.Linear(corrdim_trans, trans_param_num)
self.leakyrelu = nn.LeakyReLU(0.1, inplace=True)
self.relu = nn.ReLU(inplace=True)
self.avgpool = nn.AvgPool2d(7, stride=1)
self.avgpool3d = nn.AvgPool3d((4, 2, 2), stride=(1, 2, 2))
self.maxpool2d = nn.MaxPool2d(2, stride=2)
# initialization
nn.init.kaiming_normal_(self.afterconv1.weight, mode='fan_out', nonlinearity='relu')
nn.init.kaiming_normal_(self.afterconv3_trans.weight, mode='fan_out', nonlinearity='relu')
nn.init.kaiming_normal_(self.afterconv4_trans.weight, mode='fan_out', nonlinearity='relu')
# assuming no fc pre-training
for m in self.modules():
if isinstance(m, nn.Linear):
m.weight.data.normal_(0, 0.01)
m.bias.data.zero_()
# transformation
self.geometricTnf = GeometricTnfAffine(geometric_model='affine',
tps_grid_size=3,
tps_reg_factor=0.2,
out_h=self.spatial_out2, out_w=self.spatial_out2,
offset_factor=227/210)
self.geometricTnf_img = GeometricTnfAffine(geometric_model='affine',
tps_grid_size=3,
tps_reg_factor=0.2,
out_h=80, out_w=80,
offset_factor=227/210)
xs = np.linspace(-1,1,80)
xs = np.meshgrid(xs, xs)
xs = np.stack(xs, 2)
self.xs = xs
def __init__(self, class_num=8, dim_in=2048, trans_param_num=3, detach_network=False, pretrained=True, temporal_out=4, T=None, hist=1):
'''
Args:
class_num: 类别数
dim_in: 输入特征维度
trans_param_num: 描述变换的参数数量，此处只需要三个参数即可
temporal_out: 视频时序长度
'''
super(CycleTime, self).__init__()
dim = 512
print(pretrained)
resnet = resnet_res4s1.resnet50(pretrained=pretrained)
self.encoderVideo = inflated_resnet.InflatedResNet(copy.deepcopy(resnet))
self.detach_network = detach_network
self.hist = hist
self.div_num = 512
self.T = self.div_num**-.5 if T is None else T
print('self.T:', self.T)
self.afterconv1 = nn.Conv3d(1024, 512, kernel_size=1, bias=False)
self.spatial_out1 = 30
self.spatial_out2 = 10
self.temporal_out = temporal_out
self.afterconv3_trans = nn.Conv2d(self.spatial_out1 * self.spatial_out1, 128, kernel_size=4, padding=0, bias=False)
self.afterconv4_trans = nn.Conv2d(128, 64, kernel_size=4, padding=0, bias=False)
corrdim = 64 * 4 * 4
corrdim_trans = 64 * 4 * 4
self.linear2 = nn.Linear(corrdim_trans, trans_param_num)
self.leakyrelu = nn.LeakyReLU(0.1, inplace=True)
self.relu = nn.ReLU(inplace=True)
# initialization
nn.init.kaiming_normal_(self.afterconv1.weight, mode='fan_out', nonlinearity='relu')
nn.init.kaiming_normal_(self.afterconv3_trans.weight, mode='fan_out', nonlinearity='relu')
nn.init.kaiming_normal_(self.afterconv4_trans.weight, mode='fan_out', nonlinearity='relu')
# assuming no fc pre-training
for m in self.modules():
if isinstance(m, nn.Linear):
m.weight.data.normal_(0, 0.01)
m.bias.data.zero_()
# transformation
self.geometricTnf = GeometricTnfAffine(geometric_model='affine',
tps_grid_size=3,
tps_reg_factor=0.2,
out_h=self.spatial_out2, out_w=self.spatial_out2,
offset_factor=227/210)
self.criterion_inlier = WeakInlierCountPool(geometric_model='affine', tps_grid_size=3, tps_reg_factor=0.2, h_matches=30, w_matches=30, use_conv_filter=False, dilation_filter=0, normalize_inlier_count=True)
self.criterion_synth = TransformedGridLoss(use_cuda=True, geometric_model='affine')