MobileNetV2 Class Reference

Implementation of MobileNetV2 as done in py-torchvision See: // https://github.com/pytorch/vision/blob/main/torchvision/models/mobilenetv2.py#L66. More...

#include <mobilenet_v2.h>

Inheritance diagram for MobileNetV2:

Collaboration diagram for MobileNetV2:

Public Member Functions
	MobileNetV2 (int num_classes=1000, float width_mult=1.0f, int round_nearest=8, float dropout=0.2)
	Construct a new MobileNetV2 object.
torch::Tensor	forward (torch::Tensor x)
	Performs the forward pass.
void	initialize_weights ()
	Initialize conv/bn/linear similar to torchvision defaults.
void	load_torchvision_weights (std::string pt)
	Loads a .pt weight file containing a dict with key/parameter pairs.
int	getNinputChannelsOfClassifier () const
	Gets the number of input channels of the classifier.
void	replaceClassifier (torch::nn::Sequential &newClassifier)
	Replaces classifier with a new one.
void	setFeaturesLearning (bool doLearn)
	Enables/disables learning in the feature layers.
torch::nn::Sequential	getClassifier () const
	Gets the Classifier object.

Static Public Member Functions
static torch::Tensor	preprocess (cv::Mat img, bool resizeOnly=false)
	Preprocessing of an openCV image for inference or learning.

Static Public Attributes
static constexpr char	featuresModuleName [] = "features"
	Name of the features submodule.
static constexpr char	classifierModuleName [] = "classifier"
	Name of the classifier submodule.

Detailed Description

Implementation of MobileNetV2 as done in py-torchvision See: // https://github.com/pytorch/vision/blob/main/torchvision/models/mobilenetv2.py#L66.

Constructor & Destructor Documentation

MobileNetV2()

MobileNetV2::MobileNetV2 ( int  num_classes = 1000, float  width_mult = 1.0f, int  round_nearest = 8, float  dropout = 0.2  )

inline

Construct a new MobileNetV2 object.

If you want to load the weight files from torchvision into this class use the default values for the parameters.

Parameters

num_classes	Number of classes.
width_mult	Width multiplier - adjusts number of channels in each layer by this amount.
round_nearest	Round the number of channels in each layer to be a multiple of this number.
dropout	Dropout probability for the dropout layer in the classifier.

Member Function Documentation

forward()

torch::Tensor MobileNetV2::forward ( torch::Tensor  x )

inline

Performs the forward pass.

Parameters

x	The batch of input images.

Returns

torch::Tensor The category scores for the different labels.

getClassifier()

torch::nn::Sequential MobileNetV2::getClassifier ( ) const

inline

Gets the Classifier object.

Gets a shared pointer to the classifier, for example, to attach an optimiser for transfer learning.

Returns

torch::nn::Sequential

getNinputChannelsOfClassifier()

int MobileNetV2::getNinputChannelsOfClassifier ( ) const

inline

Gets the number of input channels of the classifier.

This will make it easy to replace the classifier with anything the user wants by creating their own torch::nn::Sequential() for the classifier.

Returns

int The number of intput channels of classifier class "classfier".

load_torchvision_weights()

void MobileNetV2::load_torchvision_weights ( std::string  pt )

inline

Loads a .pt weight file containing a dict with key/parameter pairs.

See https://github.com/pytorch/pytorch/issues/36577 The difference between pytorch and libtorch is that pytorch just has named parameters but libtorch has both named parameters and named buffers. This method makes sure that the key/parameters pairs are loaded into both named buffers and named parameters.

Parameters

pt	filename of the .pt weight file.

preprocess()

static torch::Tensor MobileNetV2::preprocess ( cv::Mat  img, bool  resizeOnly = false  )

inlinestatic

Preprocessing of an openCV image for inference or learning.

The images are resized to 256x256, followed by a central crop of 224x224. Finally the values are first rescaled to [0.0, 1.0] and then normalized using mean=[0.485, 0.456, 0.406] and std=[0.229, 0.224, 0.225].

Parameters

img	8bit BGR openCV image with an aspect ratio of 1:1.
resizeOnly	If true the image is only resized to 224x224 but not cropped. Default: false.

Returns

torch::Tensor The image as a tensor ready to be used for inference and learning.

replaceClassifier()

void MobileNetV2::replaceClassifier ( torch::nn::Sequential &  newClassifier )

inline

Replaces classifier with a new one.

For transfer learning the default classifier is replaced with a new one.

Parameters

newClassifier

The new classifier.

setFeaturesLearning()

void MobileNetV2::setFeaturesLearning ( bool  doLearn )

inline

Enables/disables learning in the feature layers.

For transfer learning one needs to disable learning in the features submodule.

Member Data Documentation

classifierModuleName

constexpr char MobileNetV2::classifierModuleName[] = "classifier"

staticconstexpr

Name of the classifier submodule.

This appears as part of the key in named_parametes and named_buffers.

featuresModuleName

constexpr char MobileNetV2::featuresModuleName[] = "features"

staticconstexpr

Name of the features submodule.

This appears as part of the key in named_parametes and named_buffers.

---

The documentation for this class was generated from the following file:

• mobilenet_v2.h