torchvision

class trojanvision.models.AlexNet(name='alexnet', model=_AlexNet, **kwargs)

AlexNet proposed by Alex Krizhevsky from Google in 2014.

Available model names:

{'alexnet'}

See also

• torchvision: torchvision.models.alexnet

• paper: One weird trick for parallelizing convolutional neural networks

self.features = nn.Sequential(
    nn.Conv2d(3, 64, kernel_size=11, stride=4, padding=2),
    nn.ReLU(inplace=True),
    nn.MaxPool2d(kernel_size=3, stride=2),
    nn.Conv2d(64, 192, kernel_size=5, padding=2),
    nn.ReLU(inplace=True),
    nn.MaxPool2d(kernel_size=3, stride=2),
    nn.Conv2d(192, 384, kernel_size=3, padding=1),
    nn.ReLU(inplace=True),
    nn.Conv2d(384, 256, kernel_size=3, padding=1),
    nn.ReLU(inplace=True),
    nn.Conv2d(256, 256, kernel_size=3, padding=1),
    nn.ReLU(inplace=True),
    nn.MaxPool2d(kernel_size=3, stride=2),
)

self.pool = nn.AdaptiveAvgPool2d((6, 6))
self.flatten = nn.Flatten(1)

self.classifier = nn.Sequential(
    nn.Dropout(),
    nn.Linear(256 * 6 * 6, 4096),
    nn.ReLU(inplace=True),
    nn.Dropout(),
    nn.Linear(4096, 4096),
    nn.ReLU(inplace=True),
    nn.Linear(4096, num_classes),
)

class trojanvision.models.DenseNet(name='densenet', layer=121, model=_DenseNet, **kwargs)

DenseNet proposed by Gao Huang from Cornell University in CVPR 2017.

Available model names:

{'densenet', 'densenet_comp',
 'densenet121', 'densenet169', 'densenet201', 'densenet161',
 'densenet121_comp', 'densenet169_comp', 'densenet201_comp', 'densenet161_comp'}

See also

• torchvision: torchvision.models.densenet121

• paper: Densely Connected Convolutional Networks

Note

_comp reduces the first convolutional layer from kernel_size=7, stride=2, padding=3

to kernel_size=3, stride=1, padding=1, and removes following norm0, relu0, pool0 (pool0 is torch.nn.MaxPool2d) before block layers.

class trojanvision.models.EfficientNet(name='efficientnet', layer='_b0', model=_EfficientNet, **kwargs)

EfficientNet proposed by Mingxing Tan from Google in ICML 2019.

Available model names:

{'efficientnet', 'efficientnet_comp',
 'efficientnet_b0', 'efficientnet_b1', 'efficientnet_b2',
 'efficientnet_b3', 'efficientnet_b4', 'efficientnet_b5',
 'efficientnet_b6', 'efficientnet_b7',
 'efficientnet_b0_comp', 'efficientnet_b1_comp', 'efficientnet_b2_comp',
 'efficientnet_b3_comp', 'efficientnet_b4_comp', 'efficientnet_b5_comp',
 'efficientnet_b6_comp', 'efficientnet_b7_comp'}

See also

• torchvision: torchvision.models.efficientnet_b0

• paper: EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

Note

_comp reduces the first convolutional layer from kernel_size=7, stride=2, padding=3

to kernel_size=3, stride=1, padding=1.

class trojanvision.models.MNASNet(name='mnasnet', mnas_alpha=1.0, model=_MNASNet, **kwargs)

MNASNet proposed by Mingxing Tan from Google in CVPR 2019.

Available model names:

{'mnasnet', 'mnasnet0_5', 'mnasnet0_75', 'mnasnet1_0', 'mnasnet1_3'}

See also

• torchvision: torchvision.models.mnasnet0_5

• paper: MnasNet: Platform-Aware Neural Architecture Search for Mobile

class trojanvision.models.MobileNet(name='mobilenet_v2', model=_MobileNet, **kwargs)

MobileNets proposed by Andrew Howard and Liang-Chieh Chen from Google in CVPR 2018.

Available model names:

['mobilenet_v2', 'mobilenet_v3_large', 'mobilenet_v3_small',
 'mobilenet_v2_comp', 'mobilenet_v3_large_comp', 'mobilenet_v3_small_comp']

See also

• MobileNet v2:

  • torchvision: torchvision.models.mobilenet_v2

  • paper: MobileNetV2: Inverted Residuals and Linear Bottlenecks

• MobileNet v3:

  • torchvision: torchvision.models.mobilenet_v3_small

  • paper: Searching for MobileNetV3

Note

_comp uses a small inverted_residual_setting and set first conv layer stride=1.

class trojanvision.models.ResNet(name='resnet', layer=18, model=_ResNet, **kwargs)

ResNet model series including ResNet, ResNext and WideResNet.

Available model names:

{'resnet', 'resnet_comp', 'resnet_s',
 'resnet18', 'resnet34', 'resnet50', 'resnet101', 'resnet152',
 'resnet18_comp', 'resnet34_comp', 'resnet50_comp', 'resnet101_comp', 'resnet152_comp',
 'resnext50_32x4d', 'resnext101_32x8d',
 'resnext50_32x4d_comp', 'resnext101_32x8d_comp',
 'wide_resnet50_2', 'wide_resnet101_2',
 'wide_resnet50_2_comp', 'wide_resnet101_2_comp',

 'resnet18_s', 'resnet34_s', 'resnet50_s', 'resnet101_s', 'resnet152_s',
 'resnet18_ap_comp'}

See also

• ResNet:

  • torchvision: torchvision.models.resnet18

  • paper: Deep Residual Learning for Image Recognition

• ResNext:

  • torchvision: torchvision.models.resnext50_32x4d

  • paper: Aggregated Residual Transformations for Deep Neural Networks

• WideResNet:

  • torchvision: torchvision.models.wide_resnet50_2

  • paper: Wide Residual Networks

Note

• _comp reduces the first convolutional layer from kernel_size=7, stride=2, padding=3

  to kernel_size=3, stride=1, padding=1, and removes the maxpool layer before block layers.

• _s further reduces the conv channels and number of blocks based on _comp.

  ResNetS is used in NIPS 2017 paper by Facebook Research about continual learning.

  • paper: Gradient Episodic Memory for Continual Learning (GEM)

  • code: https://github.com/facebookresearch/GradientEpisodicMemory/blob/master/model/common.py

• _ap conducts average pooling in all blocks rather than a global pooling layer after feature extractor.

  ResNetAP is used in ICLR 2021 paper about dataset condensation.

  • paper: Dataset Condensation with Gradient Matching

  • code: https://github.com/VICO-UoE/DatasetCondensation/blob/master/networks.py

class trojanvision.models.ShuffleNetV2(name='shufflenet_v2', layer='_x0_5', model=_ShuffleNetV2, **kwargs)

ShuffleNet v2 proposed by Ningning Ma from Megvii in ECCV 2018.

Available model names:

{'shufflenet_v2', 'shufflenet_v2_comp',
 'shufflenet_v2_x0_5', 'shufflenet_v2_x1_0',
 'shufflenet_v2_x1_5', 'shufflenet_v2_x2_0',
 'shufflenet_v2_x0_5_comp', 'shufflenet_v2_x1_0_comp',
 'shufflenet_v2_x1_5_comp', 'shufflenet_v2_x2_0_comp'}

See also

• torchvision: torchvision.models.shufflenet_v2_x0_5

• paper: ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design

Note

_comp reduces the first convolutional layer from kernel_size=7, stride=2, padding=3

to kernel_size=3, stride=1, padding=1, and removes the maxpool layer before block layers.

class trojanvision.models.VGG(name='vgg', layer=13, model=_VGG, **kwargs)

VGG model proposed by Karen Simonyan from University of Oxford in ICLR 2015.

Available model names:

{'vgg', 'vgg_bn', 'vgg_comp', 'vgg_bn_comp', 'vgg_s', 'vgg_bn_s',
 'vgg11', 'vgg13', 'vgg16', 'vgg19',
 'vgg11_bn', 'vgg13_bn', 'vgg16_bn', 'vgg19_bn',
 'vgg11_comp', 'vgg13_comp', 'vgg16_comp', 'vgg19_comp',
 'vgg11_bn_comp', 'vgg13_bn_comp', 'vgg16_bn_comp', 'vgg19_bn_comp'
 'vgg11_s', 'vgg13_s', 'vgg16_s', 'vgg19_s',
 'vgg11_bn_s', 'vgg13_bn_s', 'vgg16_bn_s', 'vgg19_bn_s'}

See also

• torchvision: torchvision.models.vgg11

• paper: Very Deep Convolutional Networks for Large-Scale Image Recognition

Note

• _comp sets torch.nn.AdaptiveAvgPool2d from (7, 7) to (1, 1), update the intermediate feature dimension from 4096 to 512 in self.classifier.

• _s further makes self.classifier only one single linear layer based on _comp.