Image Classification Networks from torchvision — mlr

Well-known image classification networks from the torchvision package.

Each of these learners wraps one of the model_*() generators of torchvision. The section Available Learners below lists for every learner the original architecture that it implements, the paper that introduced this architecture, and the file in torchvision that contains the implementation, so the architecture can be inspected directly.

Available Learners

The table below lists for each learner the original architecture that it implements, the paper that introduced this architecture, and the file in torchvision that contains the implementation.

Learner	Architecture	Reference	Implementation
`classif.alexnet`	AlexNet	Krizhevsky (2017) and Krizhevsky (2014)	models-alexnet.R
`classif.convnext_tiny_1k`	ConvNeXt	Liu (2022)	models-convnext.R
`classif.convnext_tiny_22k`	ConvNeXt	Liu (2022)	models-convnext.R
`classif.convnext_small_22k`	ConvNeXt	Liu (2022)	models-convnext.R
`classif.convnext_base_1k`	ConvNeXt	Liu (2022)	models-convnext.R
`classif.convnext_base_22k`	ConvNeXt	Liu (2022)	models-convnext.R
`classif.convnext_large_1k`	ConvNeXt	Liu (2022)	models-convnext.R
`classif.convnext_large_22k`	ConvNeXt	Liu (2022)	models-convnext.R
`classif.efficientnet_b0`	EfficientNet	Tan (2019)	models-efficientnet.R
`classif.efficientnet_b1`	EfficientNet	Tan (2019)	models-efficientnet.R
`classif.efficientnet_b2`	EfficientNet	Tan (2019)	models-efficientnet.R
`classif.efficientnet_b3`	EfficientNet	Tan (2019)	models-efficientnet.R
`classif.efficientnet_b4`	EfficientNet	Tan (2019)	models-efficientnet.R
`classif.efficientnet_b5`	EfficientNet	Tan (2019)	models-efficientnet.R
`classif.efficientnet_b6`	EfficientNet	Tan (2019)	models-efficientnet.R
`classif.efficientnet_b7`	EfficientNet	Tan (2019)	models-efficientnet.R
`classif.efficientnet_v2_s`	EfficientNetV2	Tan (2021)	models-efficientnetv2.R
`classif.efficientnet_v2_m`	EfficientNetV2	Tan (2021)	models-efficientnetv2.R
`classif.efficientnet_v2_l`	EfficientNetV2	Tan (2021)	models-efficientnetv2.R
`classif.inception_v3`	Inception v3	Szegedy (2016)	models-inception.R
`classif.maxvit`	MaxViT	Tu (2022)	models-maxvit.R
`classif.mobilenet_v2`	MobileNetV2	Sandler (2018)	models-mobilenetv2.R
`classif.mobilenet_v3_large`	MobileNetV3	Howard (2019)	models-mobilenetv3.R
`classif.mobilenet_v3_small`	MobileNetV3	Howard (2019)	models-mobilenetv3.R
`classif.resnet18`	ResNet	He (2016)	models-resnet.R
`classif.resnet34`	ResNet	He (2016)	models-resnet.R
`classif.resnet50`	ResNet	He (2016)	models-resnet.R
`classif.resnet101`	ResNet	He (2016)	models-resnet.R
`classif.resnet152`	ResNet	He (2016)	models-resnet.R
`classif.resnext50_32x4d`	ResNeXt	He (2016) and Xie (2017)	models-resnet.R
`classif.resnext101_32x8d`	ResNeXt	He (2016) and Xie (2017)	models-resnet.R
`classif.wide_resnet50_2`	Wide ResNet	He (2016) and Zagoruyko (2016)	models-resnet.R
`classif.wide_resnet101_2`	Wide ResNet	He (2016) and Zagoruyko (2016)	models-resnet.R
`classif.vgg11`	VGG	Simonyan (2014)	models-vgg.R
`classif.vgg11_bn`	VGG	Simonyan (2014)	models-vgg.R
`classif.vgg13`	VGG	Simonyan (2014)	models-vgg.R
`classif.vgg13_bn`	VGG	Simonyan (2014)	models-vgg.R
`classif.vgg16`	VGG	Simonyan (2014)	models-vgg.R
`classif.vgg16_bn`	VGG	Simonyan (2014)	models-vgg.R
`classif.vgg19`	VGG	Simonyan (2014)	models-vgg.R
`classif.vgg19_bn`	VGG	Simonyan (2014)	models-vgg.R
`classif.vit_b_16`	Vision Transformer (ViT)	Dosovitskiy (2021)	models-vit.R
`classif.vit_b_32`	Vision Transformer (ViT)	Dosovitskiy (2021)	models-vit.R
`classif.vit_l_16`	Vision Transformer (ViT)	Dosovitskiy (2021)	models-vit.R
`classif.vit_l_32`	Vision Transformer (ViT)	Dosovitskiy (2021)	models-vit.R
`classif.vit_h_14`	Vision Transformer (ViT)	Dosovitskiy (2021)	models-vit.R

Parameters

Parameters from LearnerTorchImage and

pretrained :: logical(1)
Whether to use the pretrained model. The final linear layer will be replaced with a new nn_linear with the number of classes inferred from the Task.

classif.inception_v3 additionally has

aux_logits :: logical(1)
Whether to enable the auxiliary classifier, which acts as a regularizer during training and is not applied when predicting. Defaults to FALSE. The configured loss is applied to the predictions of both classifiers and does not have to be changed for this. Note that the auxiliary classifier raises the minimum input size from 75x75 to 299x299, because it pools and convolves the feature map further than the main classifier does.
aux_weight :: numeric(1)
The weight with which the loss of the auxiliary classifier is added to the loss of the main classifier. Defaults to 0.4, the value used by the Inception v3 paper. Only has an effect if aux_logits is TRUE.

Properties

Supported task types: "classif"
Predict Types: "response" and "prob"
Feature Types: "lazy_tensor"
Required packages: "mlr3torch", "torch", "torchvision"

References

Krizhevsky A, Sutskever I, Hinton G (2017). “Imagenet classification with deep convolutional neural networks.” Communications of the ACM, 60(6), 84–90. Krizhevsky A (2014). “One weird trick for parallelizing convolutional neural networks.” arXiv preprint arXiv:1404.5997. Liu Z, Mao H, Wu C, Feichtenhofer C, Darrell T, Xie S (2022). “A ConvNet for the 2020s.” In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 11976–11986. Tan M, Le Q (2019). “EfficientNet: Rethinking model scaling for convolutional neural networks.” In Proceedings of the 36th International Conference on Machine Learning (ICML), 6105–6114. PMLR. Tan M, Le Q (2021). “EfficientNetV2: Smaller models and faster training.” In Proceedings of the 38th International Conference on Machine Learning (ICML), 10096–10106. PMLR. Szegedy C, Vanhoucke V, Ioffe S, Shlens J, Wojna Z (2016). “Rethinking the inception architecture for computer vision.” In Proceedings of the IEEE conference on computer vision and pattern recognition, 2818–2826. Tu Z, Talebi H, Zhang H, Yang F, Milanfar P, Bovik A, Li Y (2022). “MaxViT: Multi-axis vision transformer.” In Proceedings of the European Conference on Computer Vision (ECCV), 459–479. Sandler M, Howard A, Zhu M, Zhmoginov A, Chen L (2018). “Mobilenetv2: Inverted residuals and linear bottlenecks.” In Proceedings of the IEEE conference on computer vision and pattern recognition, 4510–4520. Howard A, Sandler M, Chu G, Chen L, Chen B, Tan M, Wang W, Zhu Y, Pang R, Vasudevan V, Le Q, Adam H (2019). “Searching for MobileNetV3.” In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 1314–1324. He K, Zhang X, Ren S, Sun J (2016). “Deep residual learning for image recognition.” In Proceedings of the IEEE conference on computer vision and pattern recognition, 770–778. Xie S, Girshick R, Dollár P, Tu Z, He K (2017). “Aggregated residual transformations for deep neural networks.” In Proceedings of the IEEE conference on computer vision and pattern recognition, 1492–1500. Zagoruyko S, Komodakis N (2016). “Wide residual networks.” In Proceedings of the British Machine Vision Conference (BMVC). 1605.07146, https://arxiv.org/abs/1605.07146. Simonyan K, Zisserman A (2014). “Very deep convolutional networks for large-scale image recognition.” arXiv preprint arXiv:1409.1556. Dosovitskiy A, Beyer L, Kolesnikov A, Weissenborn D, Zhai X, Unterthiner T, Dehghani M, Minderer M, Heigold G, Gelly S, Uszkoreit J, Houlsby N (2021). “An image is worth 16x16 words: Transformers for image recognition at scale.” In The Ninth International Conference on Learning Representations (ICLR). 2010.11929, https://arxiv.org/abs/2010.11929.

Super classes

mlr3::Learner -> LearnerTorch -> LearnerTorchImage -> LearnerTorchVision

Methods

Inherited methods

`LearnerTorchVision$new()`

Creates a new instance of this R6 class.

Usage

LearnerTorchVision$new(
  name,
  module_generator,
  label,
  optimizer = NULL,
  loss = NULL,
  callbacks = list(),
  jittable = FALSE,
  extra_param_set = NULL,
  network_args = character(0)
)

Arguments

name: (character(1))
The name of the network.
module_generator: (function(pretrained, num_classes))
Function that generates the network.
label: (character(1))
The label of the network.
optimizer: (TorchOptimizer)
The optimizer to use for training. Per default, adam is used.
loss: (TorchLoss)
The loss used to train the network. Per default, mse is used for regression and cross_entropy for classification.
callbacks: (list() of TorchCallbacks)
The callbacks. Must have unique ids.
jittable: (logical(1))
Whether to use jitting.
extra_param_set: (ParamSet or NULL)
Parameters that this network has in addition to pretrained, or NULL if it has none. They are added to the learner's ParamSet.
network_args: (character())
The ids of those parameters of extra_param_set that are passed on to module_generator. The remaining ones are interpreted by the learner itself, such as aux_weight, which is used to wrap the loss and never reaches the network.

`LearnerTorchVision$clone()`

The objects of this class are cloneable with this method.

Usage

LearnerTorchVision$clone(deep = FALSE)

Arguments

deep: Whether to make a deep clone.