Load pre training model ( There are major updates )

Believe recently (2022 year 7 month ) Installed or updated PyTorch and torchvision My comrades may have encountered one of the following errors when running the code ：

• UserWarning: The parameter ‘pretrained’ is deprecated since 0.13 and will be removed in 0.15, please use ‘weights’ instead.

• UserWarning: Arguments other than a weight enum or None for ‘weights’ are deprecated since 0.13 and will be removed in 0.15. The current behavior is equivalent to passing weights=ResNet50_Weights.IMAGENET1K_V1. You can also use weights=ResNet50_Weights.DEFAULT to get the most up-to-date weights.

• Expected type ‘Optional[ResNet50_Weights]’, got ‘str’ instead

This is because torchvision 0.13 A major update has been made to the loading method of the pre training model . Today, you can put it on at one time 3 strip Bug Destroy all .

from torchvision 0.13 Start ,torchvision Offer a brand new Multi weight support API (Multi-weight support API) , It supports loading different versions of weight parameter files into the model .

1. Comparison between new and old versions

from torchvision 0.13 Start , Load the parameters of the pre training model function from pretrained = True Change it to weights= Pre training model parameter version . And the old version will be written in the future torchvision 0.15 Version is Deprecated .

for instance ：

from torchvision import models

#  The old version is written , In the future torchvision 0.15 Version is Deprecated
model_old = models.resnet50(pretrained=True) # deprecated
model_old = models.resnet50(True) # deprecated

# torchvision 0.13 And later new versions 
model_new = models.resnet50(weights=models.ResNet50_Weights.IMAGENET1K_V1)
#  No pre training model loaded 
model = models.resnet50(weights=None)
model = models.resnet50()

among , The first 8 Line code IMAGENET1K_V1 It means ResNet-50 stay ImageNet The first version of the weight parameter file for pre training on the dataset . Is a version identifier .

2. The benefits of the new writing

In the old version pretrained = True in , We don't have much choice for pre training weight parameters , The default pre training weight file version should be used as soon as it is implemented . But the problem is , The development of deep learning is changing with each passing day , Soon, a more powerful model was born .

And use the new version weights= Pre training model parameter version , It is equivalent to that we have the option of pre training weight parameter file . We can enjoy using more accurate, faster and more updated pre training weight parameter files , Help our research to a higher level .

for instance ：

from torchvision import models

#  The loading accuracy is 76.130% Old weight parameter file V1
model_v1 = models.resnet50(weights=models.ResNet50_Weights.IMAGENET1K_V1)
#  Equivalent writing 
model_v1 = models.resnet50(weights="IMAGENET1K_V1")

#  The loading accuracy is 80.858% New weight parameter file V2
model_v2 = models.resnet50(weights=models.ResNet50_Weights.IMAGENET1K_V2)
#  Equivalent writing 
model_v1 = models.resnet50(weights="IMAGENET1K_V2")

If you don't know which version of the weight file is the latest , No problem , Choose the default DEFAULT that will do . The official will follow torchvision Upgrade and give way DEFAULT The weight file version is kept up to date . As shown in the following code ：

from torchvision import models

#  If you don't know which version is the latest ,  Choose the default DEFAULT that will do 
model_new = models.resnet50(weights=models.ResNet50_Weights.DEFAULT)