Skip to content

Fine-tune pretrained Convolutional Neural Networks with PyTorch

License

Notifications You must be signed in to change notification settings

huiyuan-Wang/pytorch-cnn-finetune

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

41 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Fine-tune pretrained Convolutional Neural Networks with PyTorch.

PyPI CircleCI codecov.io

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 the torchvision package:

  • ResNet (resnet18, resnet34, resnet50, resnet101, resnet152)
  • ResNeXt (resnext50_32x4d, resnext101_32x8d)
  • 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)
  • MobileNet V2 (mobilenet_v2)
  • ShuffleNet v2 (shufflenet_v2_x0_5, shufflenet_v2_x1_0)
  • AlexNet (alexnet)
  • GoogLeNet (googlenet)
  • 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)
  • PNASNet-5-Large (pnasnet5large)
  • PolyNet (polynet)

Requirements

  • Python 3.5+
  • PyTorch 1.1+

Installation

pip install cnn_finetune

Major changes:

Version 0.4

  • Default value for pretrained argument in make_model is changed from False to True. Now call make_model('resnet18', num_classes=10) is equal to make_model('resnet18', num_classes=10, pretrained=True)

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]

CIFAR10 Example

See examples/cifar10.py file (requires PyTorch 1.1+).

About

Fine-tune pretrained Convolutional Neural Networks with PyTorch

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 96.3%
  • Shell 3.7%