def __init__(self,
encoder_dir=None,
decoder_dir=None,
temp=None,
pretrainRes=False,
uselayer=4):
'''
For switchable concenration loss
Using Resnet18
'''
super(Model_switchGTfixdot_swCC_Res, self).__init__()
self.gray_encoder = encoder_res18(pretrained=pretrainRes,
uselayer=uselayer)
self.rgb_encoder = encoder3(reduce=True)
self.nlm = NLM_woSoft()
self.decoder = decoder3(reduce=True)
self.temp = temp
self.softmax = nn.Softmax(dim=1)
self.cos_window = torch.Tensor(np.outer(np.hanning(40),
np.hanning(40))).cuda()
self.normalize = normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
if (not encoder_dir is None):
print("Using pretrained encoders: %s." % encoder_dir)
self.rgb_encoder.load_state_dict(torch.load(encoder_dir))
if (not decoder_dir is None):
print("Using pretrained decoders: %s." % decoder_dir)
self.decoder.load_state_dict(torch.load(decoder_dir))
for param in self.decoder.parameters():
param.requires_grad = False
for param in self.rgb_encoder.parameters():
param.requires_grad = False
def __init__(self,
pretrained,
encoder_dir=None,
decoder_dir=None,
temp=1,
Resnet="r18",
color_switch=True,
coord_switch=True):
super(track_match_comb, self).__init__()
if Resnet in "r18":
self.gray_encoder = encoder_res18(pretrained=pretrained,
uselayer=4)
elif Resnet in "r50":
self.gray_encoder = encoder_res50(pretrained=pretrained,
uselayer=4)
self.rgb_encoder = encoder3(reduce=True)
self.decoder = decoder3(reduce=True)
self.rgb_encoder.load_state_dict(torch.load(encoder_dir))
self.decoder.load_state_dict(torch.load(decoder_dir))
for param in self.decoder.parameters():
param.requires_grad = False
for param in self.rgb_encoder.parameters():
param.requires_grad = False
self.nlm = NLM_woSoft()
self.normalize = normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
self.softmax = nn.Softmax(dim=1)
self.temp = temp
self.grid_flat = None
self.grid_flat_crop = None
self.color_switch = color_switch
self.coord_switch = coord_switch
def __init__(self,
pretrained,
encoder_dir=None,
decoder_dir=None,
temp=1,
Resnet="r50",
color_switch=True,
coord_switch=True):
super(track_match_comb, self).__init__()
if Resnet in "r18":
self.gray_encoder = encoder_res18(pretrained=pretrained,
uselayer=4)
elif Resnet in "r50":
self.gray_encoder = encoder_res50(pretrained=pretrained,
uselayer=4)
self.rgb_encoder = encoder3(reduce=True)
self.decoder = decoder3(reduce=True)
self.CAM = DeepLab_org(base='densenet169', c_output=1)
#self.CAM = nn.DataParallel(self.model_org).cuda()
self.CAM.load_state_dict(
convert_state_dict(torch.load('./libs/sal.pth')))
for param in self.CAM.parameters():
param.requires_grad = False
self.CAM.eval()
self.rgb_encoder.load_state_dict(torch.load(encoder_dir))
self.decoder.load_state_dict(torch.load(decoder_dir))
for param in self.decoder.parameters():
param.requires_grad = False
for param in self.rgb_encoder.parameters():
param.requires_grad = False
self.nlm = NLM_woSoft()
self.normalize = normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
self.softmax = nn.Softmax(dim=1)
self.temp = temp
self.grid_flat = None
self.grid_flat_crop = None
self.color_switch = color_switch
self.coord_switch = coord_switch
self.value_num = 128
self.key_num = 32
self.channel = 512
self.pool = nn.MaxPool2d(kernel_size=2)
self.fusion3_1 = Residual_block(self.value_num * 2, self.value_num * 2)
self.fusion3_2 = Residual_block(self.value_num * 2, self.value_num)
self.fusion2_1 = Residual_block(self.value_num, self.value_num)
self.fusion2_2 = Residual_block(self.value_num, self.value_num)
self.weak_conv1 = GC(self.value_num, self.value_num // 2)
self.weak_bn = nn.BatchNorm2d(self.value_num // 2)
#self.weak_conv2 = nn.Conv2d(self.value_num,1,1, bias=True)
self.weak_conv2_1 = nn.Conv2d(self.value_num // 2,
1,
kernel_size=3,
stride=1,
dilation=6,
padding=6)
self.weak_conv2_2 = nn.Conv2d(self.value_num // 2,
1,
kernel_size=3,
stride=1,
dilation=12,
padding=12)
self.weak_conv2_3 = nn.Conv2d(self.value_num // 2,
1,
kernel_size=3,
stride=1,
dilation=18,
padding=18)
self.memory_encoder = self._make_encoder(self.channel,
key_num=self.key_num,
value_num=self.value_num)
self.linear_e = nn.Linear(self.key_num, self.key_num, bias=False)
self.prelu = nn.ReLU(inplace=True)
self.sigmoid = nn.Sigmoid()
def __init__(self,
encoder_dir=None,
decoder_dir=None,
temp=None,
pretrainRes=False,
uselayer=4):
'''
For switchable concenration loss
Using Resnet18
'''
super(Model_switchGTfixdot_swCC_Res_ZVOS, self).__init__()
self.gray_encoder = encoder_res18(pretrained=pretrainRes,
uselayer=uselayer)
self.rgb_encoder = encoder3(reduce=True)
self.nlm = NLM_woSoft()
self.decoder = decoder3(reduce=True)
self.temp = temp
self.softmax = nn.Softmax(dim=1)
self.cos_window = torch.Tensor(np.outer(np.hanning(40),
np.hanning(40))).cuda()
self.normalize = normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
self.value_num = 256
self.key_num = 64
self.channel = 512
self.CAM = DeepLab_org(base='densenet169', c_output=1)
#self.CAM = nn.DataParallel(self.model_org).cuda()
self.CAM.load_state_dict(
convert_state_dict(torch.load('./libs/sal.pth')))
for param in self.CAM.parameters():
param.requires_grad = False
self.pool = nn.MaxPool2d(kernel_size=2)
self.weak_conv1 = GC(512 * 2, self.value_num)
self.weak_bn = nn.BatchNorm2d(self.value_num)
self.weak_conv2_1 = nn.Conv2d(self.value_num,
1,
kernel_size=3,
stride=1,
dilation=6,
padding=6)
self.weak_conv2_2 = nn.Conv2d(self.value_num,
1,
kernel_size=3,
stride=1,
dilation=12,
padding=12)
self.weak_conv2_3 = nn.Conv2d(self.value_num,
1,
kernel_size=3,
stride=1,
dilation=18,
padding=18)
self.prelu = nn.ReLU(inplace=True)
self.sigmoid = nn.Sigmoid()
if (not encoder_dir is None):
print("Using pretrained encoders: %s." % encoder_dir)
self.rgb_encoder.load_state_dict(torch.load(encoder_dir))
if (not decoder_dir is None):
print("Using pretrained decoders: %s." % decoder_dir)
self.decoder.load_state_dict(torch.load(decoder_dir))
for param in self.decoder.parameters():
param.requires_grad = False
for param in self.rgb_encoder.parameters():
param.requires_grad = False
