What is Transfrom?
It can be seen as a toolbox for image processing, and by looking at the Transform class you can find different ways to process images
More accurately, Transform has various class definitions that we can inherit from or construct, and then call the methods inside them to realize the corresponding functions
You can conveniently view the classes and methods in the transform through the structure, and then just implement the corresponding objects.Where the __call__ function acts as a function that is automatically called when this object is calledFor example, this is the definition of the Totensor class, as you can see it's all about converting images in other formats to the tensor format
The specific code is as follows, the output is the tensor type image
from torchvision import transforms
from PIL import Image
image_path="dataset/hymenoptera_data/train/ants_image/7759525_1363d24e88.jpg"
image=(image_path)
transform_class = () # tectonic (geology)transformhit the nail on the headToTensorresemble
image_tensor =transform_class(image) # 调用resemble将图像转化为tensorspecification（__call__function (math.)）
print(type(image_tensor))
What's the Tensor format for
Compared to PIL's image format and numpy's array format, the tensor image format is useful because it contains some attributes used for neural networks, such as backpropagation, whether the CPU or GPU is used for training, etc., so the tensor image format is more suitable for neural network training
from import SummaryWriter
from torchvision import transforms
from PIL import Image
image_path="dataset/hymenoptera_data/train/ants_image/7759525_1363d24e88.jpg"
image=(image_path)
transform_class = () # tectonic (geology)transformhit the nail on the headToTensorresemble
image_tensor =transform_class(image) # 调expense or outlayresemble将图像转化为tensorspecification（__call__function (math.)）
writer = SummaryWriter("logs")
writer.add_image("tensor_image",image_tensor,1) # expense or outlaytensorspecification显示图像
()
In addition, we can see that the image format processed using PIL is in Jpeg format, the image processed by opencv is in numpy's array format, and the image processed by transform is in tensor format, and the three are not the same
The various formats and the corresponding calling methods are listed below: