mmcls.models

The models package contains several sub-packages for addressing the different components of a model.

• classifiers: The top-level module which defines the whole process of a classification model.

• backbones: Usually a feature extraction network, e.g., ResNet, MobileNet.

• necks: The component between backbones and heads, e.g., GlobalAveragePooling.

• heads: The component for specific tasks. In MMClassification, we provides heads for classification.

• losses: Loss functions.

• utils: Some helper functions and common components used in various networks.

  • data_preprocessor: The component before model to preprocess the inputs, e.g., ClsDataPreprocessor.

  • Common Components: Common components used in various networks.

  • Helper Functions: Helper functions.

Build Functions

build_classifier	Build classifier.
build_backbone	Build backbone.
build_neck	Build neck.
build_head	Build head.
build_loss	Build loss.

Classifiers

BaseClassifier	Base class for classifiers.
ImageClassifier	Image classifiers for supervised classification task.
TimmClassifier	Image classifiers for pytorch-image-models (timm) model.
HuggingFaceClassifier	Image classifiers for HuggingFace model.

Backbones

AlexNet	AlexNet backbone.
CSPDarkNet	CSP-Darknet backbone used in YOLOv4.
CSPNet	The abstract CSP Network class.
CSPResNeXt	CSP-ResNeXt backbone.
CSPResNet	CSP-ResNet backbone.
Conformer	Conformer backbone.
ConvMixer	ConvMixer.
ConvNeXt	ConvNeXt.
DaViT	DaViT.
DeiT3	DeiT3 backbone.
DenseNet	DenseNet.
DistilledVisionTransformer	Distilled Vision Transformer.
EdgeNeXt	EdgeNeXt.
EfficientFormer	EfficientFormer.
EfficientNet	EfficientNet backbone.
HorNet	A PyTorch impl of : HorNet: Efficient High-Order Spatial Interactions with Recursive Gated Convolutions Inspiration from https://github.com/raoyongming/HorNet :param arch: HorNet architecture. If use string, choose from 'tiny', 'small', 'base' and 'large'. If use dict, it should have below keys: - base_dim (int): The base dimensions of embedding. - depths (List[int]): The number of blocks in each stage. - orders (List[int]): The number of order of gnConv in each stage. - dw_cfg (List[dict]): The Config for dw conv. Defaults to 'tiny'. :type arch: str | dict :param in_channels: Number of input image channels. Defaults to 3. :type in_channels: int :param drop_path_rate: Stochastic depth rate. Defaults to 0. :type drop_path_rate: float :param scale: Scaling parameter of gflayer outputs. Defaults to 1/3. :type scale: float :param use_layer_scale: Whether to use use_layer_scale in HorNet block. Defaults to True. :type use_layer_scale: bool :param out_indices: Output from which stages. Default: (3, ). :type out_indices: Sequence[int] :param frozen_stages: Stages to be frozen (stop grad and set eval mode). -1 means not freezing any parameters. Defaults to -1. :type frozen_stages: int :param with_cp: Use checkpoint or not. Using checkpoint will save some memory while slowing down the training speed. Defaults to False. :type with_cp: bool :param gap_before_final_norm: Whether to globally average the feature map before the final norm layer. In the official repo, it's only used in classification task. Defaults to True. :type gap_before_final_norm: bool :param init_cfg: The Config for initialization. Defaults to None. :type init_cfg: dict, optional.
HRNet	HRNet backbone.
InceptionV3	Inception V3 backbone.
LeNet5	LeNet5 backbone.
MViT	Multi-scale ViT v2.
MlpMixer	Mlp-Mixer backbone.
MobileNetV2	MobileNetV2 backbone.
MobileNetV3	MobileNetV3 backbone.
MobileOne	MobileOne backbone.
MobileViT	MobileViT backbone.
PCPVT	The backbone of Twins-PCPVT.
PoolFormer	PoolFormer.
RegNet	RegNet backbone.
RepMLPNet	RepMLPNet backbone.
RepVGG	RepVGG backbone.
Res2Net	Res2Net backbone.
ResNeSt	ResNeSt backbone.
ResNeXt	ResNeXt backbone.
ResNet	ResNet backbone.
ResNetV1c	ResNetV1c backbone.
ResNetV1d	ResNetV1d backbone.
ResNet_CIFAR	ResNet backbone for CIFAR.
SEResNeXt	SEResNeXt backbone.
SEResNet	SEResNet backbone.
SVT	The backbone of Twins-SVT.
ShuffleNetV1	ShuffleNetV1 backbone.
ShuffleNetV2	ShuffleNetV2 backbone.
SwinTransformer	Swin Transformer.
SwinTransformerV2	Swin Transformer V2.
T2T_ViT	Tokens-to-Token Vision Transformer (T2T-ViT)
TIMMBackbone	Wrapper to use backbones from timm library.
TNT	Transformer in Transformer.
VAN	Visual Attention Network.
VGG	VGG backbone.
VisionTransformer	Vision Transformer.

Necks

GlobalAveragePooling	Global Average Pooling neck.
GeneralizedMeanPooling	Generalized Mean Pooling neck.
HRFuseScales	Fuse feature map of multiple scales in HRNet.

Heads

ClsHead	Classification head.
LinearClsHead	Linear classifier head.
StackedLinearClsHead	Classifier head with several hidden fc layer and a output fc layer.
VisionTransformerClsHead	Vision Transformer classifier head.
EfficientFormerClsHead	EfficientFormer classifier head.
DeiTClsHead	Distilled Vision Transformer classifier head.
ConformerHead	Linear classifier head.
MultiLabelClsHead	Classification head for multilabel task.
MultiLabelLinearClsHead	Linear classification head for multilabel task.

Losses

CrossEntropyLoss	Cross entropy loss.
LabelSmoothLoss	Initializer for the label smoothed cross entropy loss.
FocalLoss	Focal loss.
AsymmetricLoss	asymmetric loss.
SeesawLoss	Implementation of seesaw loss.

models.utils

This package includes some helper functions and common components used in various networks.

Common Components

InvertedResidual	Inverted Residual Block.
SELayer	Squeeze-and-Excitation Module.
WindowMSA	Window based multi-head self-attention (W-MSA) module with relative position bias.
WindowMSAV2	Window based multi-head self-attention (W-MSA) module with relative position bias.
ShiftWindowMSA	Shift Window Multihead Self-Attention Module.
MultiheadAttention	Multi-head Attention Module.
ConditionalPositionEncoding	The Conditional Position Encoding (CPE) module.
PatchEmbed	Image to Patch Embedding.
PatchMerging	Merge patch feature map.
HybridEmbed	CNN Feature Map Embedding.
LayerScale	LayerScale layer.

Helper Functions

channel_shuffle	Channel Shuffle operation.
make_divisible	Make divisible function.
resize_pos_embed	Resize pos_embed weights.
resize_relative_position_bias_table	Resize relative position bias table.
to_ntuple	A to_tuple function generator.
is_tracing	Determine whether the model is called during the tracing of code with torch.jit.trace.