Use of torchversion.transforms

1. First look at the source code:

The content entered by clicking transform is:

Just select the circled .transforms in ctrl and enter the transform.py file. In fact, the implementation of the torchversion.transforms package is in this transform.py file

2. Introduce the Compose class:

This class is to combine multiple transforms, just like the example he gave, combine the two transforms transforms.CenterCrop(10) and transforms.Totensor() together, the output of transforms.CenterCrop(10)It is the input of transforms.Totensor().

3.ToTensor class:

Convert a PIL.Image or numpy.array data to a tensor type

4. Commonly used classes for transforms :

5. Here's a little more about:

Normalize class in transforms: (at 11:29): Common Transforms (1)_bilibili _bilibili

Risize class in transforms: (0:0) Common Transforms (2)_bilibili _bilibili

RandomCrop class: (at 10:36) Common Transforms (2) _ beep mile _bilibili

6. The usage of transform:

(1) Background knowledge: what is the tensor data type:

Why do we need the tensor data type: As shown in the following figure: The first picture shows the description of the picture in PIL.Image format, and the second picture shows the description of the picture in tensor format.We can obviously find that the tensor data type contains some content required for neural network training, such as gradient grad, etc.

PIL.Image:

tensor format:

(2) Background knowledge What is the __call__() function?

Take the following code as an example to explain:

In short, if there is a __call__() function implemented, then you can directly use the instance to pass parameters to call

(3) How to use transform:

First instantiate an instance by calling the class in transform as the tool to be used at the moment

(Of course, the parameters or precautions required for each type of instantiation are different, you need to check the official documentation yourself)

Then use this instantiated tool to process the content to be processed

from torchvision import transformsfrom PIL import Imageimg_path = "2222.jpg" #The relative path of the pictureimg = Image.open(img_path)print((img))tensor_trans = transforms.ToTensor() #Call the ToTensor class in transform to instantiate an instance tensor_trans as the tool to be used to convert PIL.Image to tensor type datatensor_img = tensor_trans(img)print(tensor_img)