Problems in replacing RESNET convolution of mmdet with ghost convolution group

1.SyntaxError: (unicode error) 'unicodeescape' codec can't decode bytes in pos

What this question means is python Chinese vs “\” Defined , The modified code contains this symbol in the file path It is likely to be an escape character , I made a mistake because of the comments copied from the Internet .

2.TypeError: FasterRCNN: ResNet: __init__() got an unexpected keyword argument‘dw_size’

This error is a prompt for unexpected characters , Generally, the error is reported because of protoc Version and python Version mismatch , So you can directly input the following commands in the terminal window to reinstall protoc：

pip install -U protobuf

I made this mistake because there is no definition dw_size This parameter , After modification .

3.ValueError: FasterRCNN: ResNet: out_channels must be divisible by groups

This error occurs because you cannot divide the number of input and output channels when grouping , Must be divisible . For a detailed explanation, you can take a look at this big guy's article ：

python Basic course ： Yes pytorch In the function of group Introduction to the function of parameters _ Programmer oyster's blog -CSDN Blog https://blog.csdn.net/chengxun03/article/details/1054442234.RuntimeError: Given groups=1, weight of size 32 3 3 3, expected input[1, 224, 224, 3] 

Or link up with the above problems , The probability of this problem is that the number of groups cannot be divided

5.RuntimeError: Sizes of tensors must match except in dimension 2. Got 198 and

This problem occurs because I use expand_as This method , The two tensor dimensions are different , in addition expand_as Only one-dimensional or n That's ok 1 Column .

6. RuntimeError: The expanded size of the tensor (272) must match the existing

And questions 5 Agreement , I have this problem because padding There is a wrong number , You cannot give integers directly , It can be used 3//2, This will not result in 272 and 270

7.pytorch How to look at all kinds of tensor data

You can take a look at this big guy's ：

【pyTorch Basics 】(1) The data type of the tensor _ state Sir The blog of -CSDN Blog _ The data type of the tensor https://blog.csdn.net/dgvv4/article/details/1241948838.RuntimeError: running_mean should contain 32 elements not 64

When doing convolution , Change the step size to 2 after , The size does not correspond well . Through a variety of print Understand the detailed process .

9. ResNet: 'list' object attribute 'append' is read-only

result.append('Senior'), You don't need an equal sign , Use append You can't take it with you at the back “=”.

10.'Bottleneck' object has no attribute 'convs'

There are no attributes in the residual block convs, Because I accidentally annotated .

11.PyTorch Medium contiguous Reading

I learned the principle of this method , See the article of the boss for details ：

PyTorch Medium contiguous Reading _DLANDML The blog of -CSDN Blog _pytorch in contiguous This article explains pytorch in contiguous The meaning of 、 Definition 、 Realization , as well as contiguous The reason for its existence , Not contiguous The solution to this problem . And compared numpy Medium contiguous.contiguous Itself is an adjective , It means continuous , About contiguous,PyTorch Provides is_contiguous、contiguous( The adjective uses ) Two methods , They are used to determine Tensor Whether it is contiguous Of , And guarantee Tensor yes contiguous Of .PyTorch Medium is_contiguous yes https://blog.csdn.net/l641208111/article/details/122683939

12.Sizes of tensors must match except in dimension 2. Got 50 and 100

Just like the problem above , Because it doesn't correspond well conv3 The number of input channels , Just modify it .

# width * scales,2022
        self.conv3 = build_conv_layer(
            self.conv_cfg,
            out1111_channels,
            self.planes * self.expansion,
            kernel_size=1,
            bias=False)

After so many problems , I made a query and solved it myself , Findings will be resnet and ghostnet Basically, I know about it , The process of possible problem solving is the process of ascension .

Some students also want to resnet The ordinary convolution in the network is replaced by Ghost Grouping convolution , You can communicate in the comments section .