【孪生网络siamfc代码学习】

 1.Fully-Convolutional Siamese Networks for Object Tracking【ECCV2016 workshop】

使用了这位老兄的代码学习一下：

huanglianghua/siamfc-pytorch: A clean PyTorch implementation of SiamFC tracking/training, evaluated on 7 datasets. (github.com)https://github.com/huanglianghua/siamfc-pytorch

 2.参考了siamfc-pytorch代码讲解（一）：backbone&head - 简书 (jianshu.com)

3.在以上基础上：思考了如下：

新建一个main.py

并如下打上断点，可以看到F.conv操作过程中的细节

from backbones import AlexNetV1
from heads import SiamFC
if __name__ == '__main__':
    alexnetv1 = AlexNetV1()
    import torch
    z = torch.randn(8, 3, 127, 127)
    output1 = alexnetv1(z)
    print(output1.shape)  # torch.Size([1, 256, 6, 6])
    x = torch.randn(8, 3, 256, 256)
    output2 = alexnetv1(x)
    print(output2.shape)  # torch.Size([1, 256, 22, 22])
    siamfc = SiamFC()
    out = siamfc(output1, output2)

    print(out.shape)
    # 换成AlexNetV2依次是：
    # torch.Size([1, 32, 17, 17])、torch.Size([1, 32, 49, 49])
    # 换成AlexNetV3依次是：
    # torch.Size([1, 512, 6, 6])、torch.Size([1, 512, 22, 22])

 

 

 

 假如模板z:

torch.Size([8, 256, 6, 6])

x:

torch.Size([8, 256, 22, 22]

minibatch=1,in_channels=256*8,out_channels=8,groups=8;

x.shape
Out[10]: torch.Size([1, 2048, 22, 22])
z.shape
Out[11]: torch.Size([8, 256, 6, 6])
#最后相关图shape,8是batch_size,
torch.Size([8, 1, 17, 17])

TORCH.NN.FUNCTIONAL.CONV2D