yolov5 improvement (1) Add attention focus mechanism

(1) Self-attention mechanism

What I want to learn is the attention mechanism, but it seems to be running out of bounds at first, and I learned the self-attention mechanism.Not to mention, it's pretty good.
NTU Li Hongyi Self-Attention Mechanism
input: vector set

muti-head: may have different connections

Application in image:

Think of rgb on a pixel as a vector

Applications on the model include: ①self-attention GAN
②DETR

Comparison of CNN and Self-attention:
CNN only considers the receptive field, and Self-attention considers the overall situation.So think of cnn as a small (simplified) Self-attention

②Small data volume is superior to CNN, while large volume Self-attention will surpass CNN
Li Hongyi's statement for the reason is: Self-Attention is more elastic, CNN is less elastic

RNN&SA
①SA is parallelized, RNN cannot parallel words
②Data memory

(2) Attention mechanism

The next step is the attention mechanism (Attention)
First upload the information first

a>
pytorch application:
First go to the information
In fact, it is on csdnThere are online courses but poor children really have no money to spend recently, but we can still learn according to his framework

1. Understand the attention mechanism

Attention is divided into four basic types according to the different dimensions of attention: channel attention, spatial attention, temporal attention and branch attention,And two combined attentions: channel-spatial attention and spatial-temporal attention.

spatial: space
temporal: time
> Draw a 3D coordinate axis like this:

2. Enter the attention module

If you encounter problems, please look at B-led lesson

The functions that this Xiaobai does not know, the example is better to understand
1) cat: splice

2) view: change the arrangement of cols and rows

3) torch.mean channel average &torch.maxChannel max
torch.nn.AdaptiveAvgPool2d(output_size): Provides a 2-dimensional adaptive average pooling operation. For any input size input, the output size can be specified as H*Wp>

Compared with global average pooling, it can be understood that the slicing method is different!!!

The attention mechanism is a plug-and-play module that can theoretically be placed behind any feature layer.
Since placing on the backbone will make the pretrained weights of the network unavailable, apply the attention mechanism to enhancing the feature extraction network

How come someone even wrote the actual combat?Still so well written?yolov5 adds an attention-focusing mechanismdownloaded.
If there is any problem in actual use, I will add it!!I feel that the b guide has already said it very well
1. If you add an independent attention mechanism layer, it may affect the number of subsequent layers (the number of layers of the feature map layer received from the backbone will change)
2. Generally not added to the backbone extraction network to avoid affecting the pre-training weights