1.dict() creates a dictionary={}
2. Whether the "tiny" in model_arch string is in the string
3.img_size=416
The data set processing of yolo3 is also a highlight. Since yolo3 has requirements for the input of the data set, the specified photo input size must be 416, so there are a series of operations for the unsatisfactory photo size. If the direct resize operation will directly lose the photo Information, the network has to adapt to the problem of photo size in the process of learning classification, resulting in poor training effect,
In yolo3, the height and width are adjusted to the same size first, and then the resize of the upsampling is performed. At the same time, the coordinate position of the label must be modified, randomly flipped horizontally, and the size is randomly changed again, and then changed to 416 as the input.
#resize: The principle of the resize function is to calculate the existing pixel points through the interpolation algorithm, so as to obtain the new pixel point position and pixel value.
# Upsampling:
Upsampling refers to restoring the resolution of the feature map to the resolution of the original image
1.unpooling (unpooling): fill the gap with 0
The opposite process of maximum pooling, the place corresponding to the maximum value is filled with the maximum value, and other positions are filled with 0
2.unsampling fills the gap with the same value
3. deconvolution/transpose convolution (transpose convolution):
Add 0 to the feature maps, and then do the convolution operation
There are parameters to learn