用PyTorch微调预训练卷积神经网络

本是古典何须时尚 2021-12-15 00:17 398阅读 0赞

如果不懂微调的小伙伴请参考这篇：[CNN 之 fine-tune methods][CNN _ fine-tune methods]  
**resnet18, resnet34, resnet50, resnet101, resnet152, densenet121, densenet169, densenet201, densenet161, alexnet, vgg11, vgg11\_bn, vgg13, vgg13\_bn, vgg16, vgg16\_bn, vgg19, vgg19\_bn, squeezenet1\_0, squeezenet1\_1, inception\_v3, resnext101\_32x4d, resnext101\_64x4d, nasnetalarge, nasnetamobile, inceptionresnetv2, dpn68, dpn68b, dpn92, dpn98, dpn131, dpn107, inception\_v4, xception, senet154, se\_resnet50, se\_resnet101, se\_resnet152, se\_resnext50\_32x4d, se\_resnext101\_32x4d, pnasnet5large, polynet这些网络都可用。**

# Features #

*  Gives access to the most popular CNN architectures pretrained on ImageNet.
 *  Automatically replaces classifier on top of the network, which allows you to train a network with a dataset that has a different number of classes.
 *  Allows you to use images with any resolution (and not only the resolution that was used for training the original model on ImageNet).
 *  Allows adding a Dropout layer or a custom pooling layer.

# Supported architectures and models #

From [torchvision][] package:

*  ResNet (resnet18, resnet34, resnet50, resnet101, resnet152)
 *  DenseNet (densenet121, densenet169, densenet201, densenet161)
 *  Inception v3 (inception\_v3)
 *  VGG (vgg11, vgg11\_bn, vgg13, vgg13\_bn, vgg16, vgg16\_bn, vgg19, vgg19\_bn)
 *  SqueezeNet (squeezenet1\_0, squeezenet1\_1)
 *  AlexNet (alexnet)

From [Pretrained models for PyTorch][] package:

*  ResNeXt (resnext101\_32x4d, resnext101\_64x4d)
 *  NASNet-A Large (nasnetalarge)
 *  NASNet-A Mobile (nasnetamobile)
 *  Inception-ResNet v2 (inceptionresnetv2)
 *  Dual Path Networks (dpn68, dpn68b, dpn92, dpn98, dpn131, dpn107)
 *  Inception v4 (inception\_v4)
 *  Xception (xception)
 *  Squeeze-and-Excitation Networks (senet154, se\_resnet50, se\_resnet101, se\_resnet152, se\_resnext50\_32x4d, se\_resnext101\_32x4d)

# Requirements #

*  Python 3.5以上
 *  PyTorch 0.3以上

# Installation #

pip install cnn_finetune

# Example usage: #

##### Make a model with ImageNet weights for 10 classes #####

from cnn_finetune import make_model
    
    model = make_model('resnet18', num_classes=10, pretrained=True)

##### Make a model with Dropout #####

model = make_model('nasnetalarge', num_classes=10, pretrained=True, dropout_p=0.5)

#### Make a model with Global Max Pooling instead of Global Average Pooling ####

import torch.nn as nn
    
    model = make_model('inceptionresnetv2', num_classes=10, pretrained=True, pool=nn.AdaptiveMaxPool2d(1))

#### Make a VGG16 model that takes images of size 256x256 pixels ####

VGG and AlexNet models use fully-connected layers, so you have to additionally pass the input size of images when constructing a new model. This information is needed to determine the input size of fully-connected layers.

model = make_model('vgg16', num_classes=10, pretrained=True, input_size=(256, 256))

#### Make a VGG16 model that takes images of size 256x256 pixels and uses a custom classifier ####

import torch.nn as nn
    
    def make_classifier(in_features, num_classes):
        return nn.Sequential(
            nn.Linear(in_features, 4096),
            nn.ReLU(inplace=True),
            nn.Linear(4096, num_classes),
        )
    
    model = make_model('vgg16', num_classes=10, pretrained=True, input_size=(256, 256), classifier_factory=make_classifier)

#### Show preprocessing that was used to train the original model on ImageNet ####

>> model = make_model('resnext101_64x4d', num_classes=10, pretrained=True)
    >> print(model.original_model_info)
    ModelInfo(input_space='RGB', input_size=[3, 224, 224], input_range=[0, 1], mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
    >> print(model.original_model_info.mean)
    [0.485, 0.456, 0.406]

[CNN _ fine-tune methods]: https://blog.csdn.net/qq_38410428/article/details/94597145
[torchvision]: https://github.com/pytorch/vision/
[Pretrained models for PyTorch]: https://github.com/Cadene/pretrained-models.pytorch