def __init__(self, net_path=None, **kargs):
super(TrackerSiamFC, self).__init__(name='SiamFC',
is_deterministic=True)
self.cfg = parse_args(**kargs)
self.para = param()
# setup GPU device if available
self.cuda = torch.cuda.is_available()
self.device = torch.device('cuda:0' if self.cuda else 'cpu')
# setup model
self.net = SiamFC()
# if net_path is not None:
# self.net.load_state_dict(torch.load(
# net_path, map_location=lambda storage, loc: storage))
self.net = self.net.to(self.device)
# setup optimizer
self.optimizer = optim.SGD(
self.net.parameters(),
lr=self.cfg.initial_lr, # self.cfg.initial_lr = 0.01
weight_decay=self.cfg.
weight_decay, # self.cfg.weight_decay = 0.0005
momentum=self.cfg.momentum) # self.cfg.momentum = 0.9
# setup lr scheduler
# self.cfg.lr_decay = 0.8685113737513527
self.lr_scheduler = ExponentialLR(self.optimizer,
gamma=self.cfg.lr_decay)
def __init__(self, **kargs):
super(SiamFC, self).__init__()
# 这里讲一下conv2中的group参数
# 假设当前的输入channel数为8， 输出为12，卷积核大小为3，
# 则当group=1时， 总共有12个神经元，每个神经元的参数大小为3*3*8，每个神经元生成一个channel
# 假设当group=4的时候，每一组的神经元输入channel=2，输出=3，总共仍然有12个神经元，只不过每个神经元的大小为3*3*2，而不是3*3*8
# Conv2d(96, 256, kernel_size=(5, 5), stride=(1, 1), groups=2)
# self.feature[4].weight.data.shape = torch.Size([256, 48, 5, 5])
# setup GPU device if available
self.cuda = torch.cuda.is_available()
self.cfg = parse_args(**kargs)
self.device = torch.device('cuda:0' if self.cuda else 'cpu')
self.para = param()
self.feature = nn.Sequential(
# conv1
nn.Conv2d(3, 96, 11, 2), # 这里的权值是[96, 3, 11, 11]
nn.BatchNorm2d(96, eps=1e-6, momentum=0.05),
nn.ReLU(inplace=True),
nn.MaxPool2d(3, 2),
# conv2
nn.Conv2d(96, 256, 5, 1, groups=2),
nn.BatchNorm2d(256, eps=1e-6, momentum=0.05),
nn.ReLU(inplace=True),
nn.MaxPool2d(3, 2),
# conv3
nn.Conv2d(256, 384, 3, 1),
nn.BatchNorm2d(384, eps=1e-6, momentum=0.05),
nn.ReLU(inplace=True),
# conv4
nn.Conv2d(384, 384, 3, 1, groups=2),
nn.BatchNorm2d(384, eps=1e-6, momentum=0.05),
nn.ReLU(inplace=True),
# conv5
nn.Conv2d(384, 256, 3, 1, groups=2))
self._initialize_weights()
self.deconv = nn.Conv2d(256 * self.para.prior_frames_num, 256, 1,
1) # 用于将堆叠后的特征对齐
def __init__(self, seq_dataset, **kargs):
super(Pairwise, self).__init__()
self.cfg = self.parse_args(**kargs)
self.para = param()
self.seq_dataset = seq_dataset # 这是一个got10k的dataset类
self.indices = np.random.permutation(
len(seq_dataset)) # 随机打乱seq_dataset的数据，打乱后的索引放到indices
# augmentation for exemplar and instance images
self.transform_z = Compose([
RandomStretch(max_stretch=0.05), # 这里是对图像进行随机尺度[0.95 1.05)的缩放
CenterCrop(self.cfg.instance_sz -
8), # cfg.instance_sz = 255 从图像中间裁剪，尺寸为255-8大小
RandomCrop(self.cfg.instance_sz -
2 * 8), # cfg.instance_sz = 255 对图像进行随机裁剪，尺寸为255 - 2*8
CenterCrop(self.cfg.exemplar_sz
), # cfg.exemplar_sz = 127 然后在中心裁剪成127 * 127
ToTensor()
])
self.transform_x = Compose([
RandomStretch(max_stretch=0.05),
CenterCrop(self.cfg.instance_sz - 8),
RandomCrop(self.cfg.instance_sz - 2 * 8),
ToTensor()
])
# 将file_path中的预训练模型读入net
def load_pre_train_mode(net, file_path):
# 读取预训练的网络
pretrained_dict = torch.load(file_path)
my_model_dict = net.state_dict()
pretrained_dict = {
k: v
for k, v in pretrained_dict.items() if k in my_model_dict
}
my_model_dict.update(pretrained_dict)
net.load_state_dict(my_model_dict)
if __name__ == '__main__':
# setup dataset
para = param()
name = 'GOT-10k'
assert name in ['VID', 'GOT-10k']
if name == 'GOT-10k':
root_dir = '/home/fanfu/data/GOT-10k'
seq_dataset = GOT10k(root_dir, subset='train')
elif name == 'VID':
root_dir = '/home/fanfu/data/ILSVRC2015/Data/VID/'
seq_dataset = ImageNetVID(root_dir, subset=('train', 'val'))
pair_dataset = Pairwise(seq_dataset)
# setup data loader
cuda = torch.cuda.is_available()
loader = DataLoader(pair_dataset,
batch_size=8,
shuffle=True,