第Pytorch教程內置模型源碼實現(xiàn)/docs/stable/torchvision/models.html

主要講解了torchvision.models的使用

torchvision.models

torchvision.models中包含了如下模型

AlexNet

ResNet

SqueezeNet

DenseNet

Inceptionv3

隨機初始化模型

importtorchvision.modelsasmodels

resnet18=models.resnet18()

alexnet=models.alexnet()

vgg16=models.vgg16()

squeezenet=models.squeezenet1_0()

desnet=models.densenet161()

inception=models.inception_v3()

使用預訓練好的參數(shù)

pytorch提供了預訓練的模型，使用torch.utils.model_zoo，通過讓參數(shù)pretrained=True來構建訓練好的模型

方法如下

resnet18=models.resnet18(pretrained=True)

alexnet=models.alexnet(pretrained=True)

squeezenet=models.squeezenet1_0(pretrained=True)

vgg16=models.vgg16(pretrained=True)

densenet=models.densenet161(pretrained=True)

inception=models.inception_v3(pretrained=True)

實例化一個預訓練好的模型會自動下載權重到緩存目錄，這個權重存儲路徑可以通過環(huán)境變量TORCH_MODEL_ZOO來指定，詳細的參考torch.utils.model_zoo.load_url()這個函數(shù)

有的模型試驗了不同的訓練和評估，例如batchnormalization。使用model.train()和model.eval()來轉換，查看train()oreval()來了解更多細節(jié)

所有的預訓練網絡希望使用相同的方式進行歸一化，例如圖片是mini-batch形式的3通道RGB圖片(3HW),H和W最少是244,。圖像必須加載到[0,1]范圍內，然后使用均值=[0.485,0.456,0.406]和std=[0.229,0.224,0.225]進行歸一化。

您可以使用以下轉換來normalzie:

normalize=trainform.Normalize9mean=[0.485,0.456,0.406],std=[0.229,0.224,0.225])

在這里我們可以找到一個在Imagenet上的這樣的例子

/pytorch/examples/blob/42e5b996718797e45c46a25c55b031e6768f8440/imagenet/main.py#L89-L101

目前這些模型的效果如下

下面是模型源碼的具體實現(xiàn)，具體實現(xiàn)大家可以閱讀源碼

###ALEXNET

torchvision.models.alexnet(pretrained=False,**kwargs)[SOURCE]

AlexNetmodelarchitecturefromthe“Oneweirdtrick…”paper.

Parameters:pretrained(bool)–IfTrue,returnsamodelpre-trainedonImageNet

###VGG

torchvision.models.vgg11(pretrained=False,**kwargs)[SOURCE]

VGG11-layermodel(configuration“A”)

Parameters:pretrained(bool)–IfTrue,returnsamodelpre-trainedonImageNet

torchvision.models.vgg11_bn(pretrained=False,**kwargs)[SOURCE]

VGG11-layermodel(configuration“A”)withbatchnormalization

Parameters:pretrained(bool)–IfTrue,returnsamodelpre-trainedonImageNet

torchvision.models.vgg13(pretrained=False,**kwargs)[SOURCE]

VGG13-layermodel(configuration“B”)

Parameters:pretrained(bool)–IfTrue,returnsamodelpre-trainedonImageNet

torchvision.models.vgg13_bn(pretrained=False,**kwargs)[SOURCE]

VGG13-layermodel(configuration“B”)withbatchnormalization

Parameters:pretrained(bool)–IfTrue,returnsamodelpre-trainedonImageNet

torchvision.models.vgg16(pretrained=False,**kwargs)[SOURCE]

VGG16-layermodel(configuration“D”)

Parameters:pretrained(bool)–IfTrue,returnsamodelpre-trainedonImageNet

torchvision.models.vgg16_bn(pretrained=False,**kwargs)[SOURCE]

VGG16-layermodel(configuration“D”)withbatchnormalization

Parameters:pretrained(bool)–IfTrue,returnsamodelpre-trainedonImageNet

torchvision.models.vgg19(pretrained=False,**kwargs)[SOURCE]

VGG19-layermodel(configuration“E”)

Parameters:pretrained(bool)–IfTrue,returnsamodelpre-trainedonImageNet

torchvision.models.vgg19_bn(pretrained=False,**kwargs)[SOURCE]

VGG19-layermodel(configuration‘E')withbatchnormalization

Parameters:pretrained(bool)–IfTrue,returnsamodelpre-trainedonImageNet

RESNET

torchvision.models.resnet18(pretrained=False,**kwargs)[SOURCE]

ConstructsaResNet-18model.

Parameters:pretrained(bool)–IfTrue,returnsamodelpre-trainedonImageNet

torchvision.models.resnet34(pretrained=False,**kwargs)[SOURCE]

ConstructsaResNet-34model.

Parameters:pretrained(bool)–IfTrue,returnsamodelpre-trainedonImageNet

torchvision.models.resnet50(pretrained=False,**kwargs)[SOURCE]

ConstructsaResNet-50model.

Parameters:pretrained(bool)–IfTrue,returnsamodelpre-trainedonImageNet

torchvision.models.resnet101(pretrained=False,**kwargs)[SOURCE]

ConstructsaResNet-101model.

Parameters:pretrained(bool)–IfTrue,returnsamodelpre-trainedonImageNet

torchvision.models.resnet152(pretrained=False,**kwargs)[SOURCE]

ConstructsaResNet-152model.

Parameters:pretrained(bool)–IfTrue,returnsamodelpre-trainedonImageNet

SQUEEZENET

torchvision.models.squeezenet1_0(pretrained=False,**kwargs)[SOURCE]

SqueezeNetmodelarchitecturefromthe“SqueezeNet:AlexNet-levelaccuracywith50xfewerparametersand0.5MBmodelsize”paper.

Parameters:pretrained(bool)–IfTrue,returnsamodelpre-trainedonImageNet

torchvision.models.squeezenet1_1(pretrained=False,**kwargs)[SOURCE]

SqueezeNet1.1modelfromtheofficialSqueezeNetrepo.SqueezeNet1.1has2.4xlesscomputationandslightlyfewerparametersthanSqueezeNet1.0,withoutsacrificingaccuracy.

Parameters:pretrained(bool)–IfTrue,returnsamodelpre-trainedonImageNet

DENSENET

torchvision.models.densenet121(pretrained=False,**kwargs)[SOURCE]

Densenet-121modelfrom“DenselyConnectedConvolutionalNetworks”

Parameters:pretrained(bool)–IfTrue,returnsamodelpre-trainedonImageNet

torchvision.models.densenet169(pretrained=False,**kwargs)[SOURCE]

Densenet-169modelfrom“DenselyConnectedConvolutionalNetworks”

Parameters:pretrained(bool)–IfTrue,returnsamodelpre-trainedonImageNet

torchvision.models.densenet161(pretrained=False,**kwargs)[SOURCE]

Densenet-161modelfrom“DenselyConnectedConvolutionalNetworks”

Parameters:pretrained(bool)–IfTrue,returnsamodelpre-trainedonImageNet

torchvision.models.densenet201(pretrained=False,**kwargs)[SOURCE]

Densenet-201modelfrom“DenselyConnectedConvolutionalNetworks”

Parameters:pretrained(bool)–IfTrue,returnsamodelpre-trainedonImageNet

INCEPTIONV3

torchvisi