Customize Models

In our design, a complete model is defined as a top-level module which contains several model components based on their functionalities.

  • model: a top-level module defines the type of the task, such as ImageClassifier for image classification, MAE for self-supervised leanrning, ImageToImageRetriever for image retrieval.

  • backbone: usually a feature extraction network that records the major differences between models, e.g., ResNet, MobileNet.

  • neck: the component between backbone and head, e.g., GlobalAveragePooling.

  • head: the component for specific tasks, e.g., ClsHead, ContrastiveHead.

  • loss: the component in the head for calculating losses, e.g., CrossEntropyLoss, LabelSmoothLoss.

  • target_generator: the component for self-supervised leanrning task specifically, e.g., VQKD, HOGGenerator.

Add a new model

Generally, for image classification and retrieval tasks, the pipelines are consistent. However, the pipelines are different from each self-supervised leanrning algorithms, like MAE and BEiT. Thus, in this section, we will explain how to add your self-supervised learning algorithm.

Add a new self-supervised learning algorithm

  1. Create a new file mmpretrain/models/selfsup/ and implement NewAlgorithm in it.

    from mmpretrain.registry import MODELS
    from .base import BaseSelfSupvisor
    class NewAlgorithm(BaseSelfSupvisor):
        def __init__(self, backbone, neck=None, head=None, init_cfg=None):
        # ``extract_feat`` function is defined in BaseSelfSupvisor, you could
        # overwrite it if needed
        def extract_feat(self, inputs, **kwargs):
        # the core function to compute the loss
        def loss(self, inputs, data_samples, **kwargs):
  2. Import the new algorithm module in mmpretrain/models/selfsup/

    from .new_algorithm import NewAlgorithm
    __all__ = [
  3. Use it in your config file.

    model = dict(

Add a new backbone

Here we present how to develop a new backbone component by an example of ResNet_CIFAR. As the input size of CIFAR is 32x32, which is much smaller than the default size of 224x224 in ImageNet, this backbone replaces the kernel_size=7, stride=2 to kernel_size=3, stride=1 and removes the MaxPooling after the stem layer to avoid forwarding small feature maps to residual blocks.

The easiest way is to inherit from ResNet and only modify the stem layer.

  1. Create a new file mmpretrain/models/backbones/

    import torch.nn as nn
    from mmpretrain.registry import MODELS
    from .resnet import ResNet
    class ResNet_CIFAR(ResNet):
        """ResNet backbone for CIFAR.
        short description of the backbone
            depth(int): Network depth, from {18, 34, 50, 101, 152}.
        def __init__(self, depth, deep_stem, **kwargs):
            # call ResNet init
            super(ResNet_CIFAR, self).__init__(depth, deep_stem=deep_stem, **kwargs)
            # other specific initializations
            assert not self.deep_stem, 'ResNet_CIFAR do not support deep_stem'
        def _make_stem_layer(self, in_channels, base_channels):
            # override the ResNet method to modify the network structure
            self.conv1 = build_conv_layer(
            self.norm1_name, norm1 = build_norm_layer(
                self.norm_cfg, base_channels, postfix=1)
            self.add_module(self.norm1_name, norm1)
            self.relu = nn.ReLU(inplace=True)
        def forward(self, x):
            # Customize the forward method if needed.
            x = self.conv1(x)
            x = self.norm1(x)
            x = self.relu(x)
            outs = []
            for i, layer_name in enumerate(self.res_layers):
                res_layer = getattr(self, layer_name)
                x = res_layer(x)
                if i in self.out_indices:
            # The return value needs to be a tuple with multi-scale outputs from different depths.
            # If you don't need multi-scale features, just wrap the output as a one-item tuple.
            return tuple(outs)
        def init_weights(self):
            # Customize the weight initialization method if needed.
            # Disable the weight initialization if loading a pretrained model.
            if self.init_cfg is not None and self.init_cfg['type'] == 'Pretrained':
            # Usually, we recommend using `init_cfg` to specify weight initialization methods
            # of convolution, linear, or normalization layers. If you have some special needs,
            # do these extra weight initialization here.


Replace original registry names from BACKBONES, NECKS, HEADS and LOSSES to MODELS in OpenMMLab 2.0 design.

  1. Import the new backbone module in mmpretrain/models/backbones/

    from .resnet_cifar import ResNet_CIFAR
    __all__ = [
        ..., 'ResNet_CIFAR'
  2. Modify the correlated settings in your config file.

    model = dict(

Add a new backbone for self-supervised learning

For some self-supervised learning algorithms, the backbones are kind of different, such as MAE, BEiT, etc. Their backbones need to deal with mask in order to extract features from visible tokens.

Take MAEViT as an example, we need to overwrite forward function to compute with mask. We also defines init_weights to initialize parameters and random_masking to generate mask for MAE pre-training.

class MAEViT(VisionTransformer):
    """Vision Transformer for MAE pre-training"""

    def __init__(mask_ratio, **kwargs) -> None:
        # position embedding is not learnable during pretraining
        self.pos_embed.requires_grad = False
        self.mask_ratio = mask_ratio
        self.num_patches = self.patch_resolution[0] * self.patch_resolution[1]

    def init_weights(self) -> None:
        """Initialize position embedding, patch embedding and cls token."""
        # define what if needed

    def random_masking(
        x: torch.Tensor,
        mask_ratio: float = 0.75
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """Generate the mask for MAE Pre-training."""

    def forward(
        x: torch.Tensor,
        mask: Optional[bool] = True
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """Generate features for masked images.

        The function supports two kind of forward behaviors. If the ``mask`` is
        ``True``, the function will generate mask to masking some patches
        randomly and get the hidden features for visible patches, which means
        the function will be executed as masked imagemodeling pre-training;
        if the ``mask`` is ``None`` or ``False``, the forward function will
        call ``super().forward()``, which extract features from images without
        if mask is None or False:
            return super().forward(x)

            B = x.shape[0]
            x = self.patch_embed(x)[0]
            # add pos embed w/o cls token
            x = x + self.pos_embed[:, 1:, :]

            # masking: length -> length * mask_ratio
            x, mask, ids_restore = self.random_masking(x, self.mask_ratio)

            # append cls token
            cls_token = self.cls_token + self.pos_embed[:, :1, :]
            cls_tokens = cls_token.expand(B, -1, -1)
            x =, x), dim=1)

            for _, layer in enumerate(self.layers):
                x = layer(x)
            # Use final norm
            x = self.norm1(x)

            return (x, mask, ids_restore)

Add a new neck

Here we take GlobalAveragePooling as an example. It is a very simple neck without any arguments. To add a new neck, we mainly implement the forward function, which applies some operations on the output from the backbone and forwards the results to the head.

  1. Create a new file in mmpretrain/models/necks/

    import torch.nn as nn
    from mmpretrain.registry import MODELS
    class GlobalAveragePooling(nn.Module):
        def __init__(self):
   = nn.AdaptiveAvgPool2d((1, 1))
        def forward(self, inputs):
            # we regard inputs as tensor for simplicity
            outs =
            outs = outs.view(inputs.size(0), -1)
            return outs
  2. Import the new neck module in mmpretrain/models/necks/

    from .gap import GlobalAveragePooling
    __all__ = [
        ..., 'GlobalAveragePooling'
  3. Modify the correlated settings in your config file.

    model = dict(

Add a new head

Based on ClsHead

Here we present how to develop a new head by the example of simplified VisionTransformerClsHead as the following. To implement a new head, we need to implement a pre_logits method for processes before the final classification head and a forward method.

Why do we need the pre_logits method?

In classification tasks, we usually use a linear layer to do the final classification. And sometimes, we need to obtain the feature before the final classification, which is the output of the pre_logits method.

  1. Create a new file in mmpretrain/models/heads/

    import torch.nn as nn
    from mmpretrain.registry import MODELS
    from .cls_head import ClsHead
    class VisionTransformerClsHead(ClsHead):
        def __init__(self, num_classes, in_channels, hidden_dim, **kwargs):
            self.in_channels = in_channels
            self.num_classes = num_classes
            self.hidden_dim = hidden_dim
            self.fc1 = nn.Linear(in_channels, hidden_dim)
            self.act = nn.Tanh()
            self.fc2 = nn.Linear(hidden_dim, num_classes)
        def pre_logits(self, feats):
            # The output of the backbone is usually a tuple from multiple depths,
            # and for classification, we only need the final output.
            feat = feats[-1]
            # The final output of VisionTransformer is a tuple of patch tokens and
            # classification tokens. We need classification tokens here.
            _, cls_token = feat
            # Do all works except the final classification linear layer.
            return self.act(self.fc1(cls_token))
        def forward(self, feats):
            pre_logits = self.pre_logits(feats)
            # The final classification linear layer.
            cls_score = self.fc2(pre_logits)
            return cls_score
  2. Import the module in mmpretrain/models/heads/

    from .vit_head import VisionTransformerClsHead
    __all__ = [
        ..., 'VisionTransformerClsHead'
  3. Modify the correlated settings in your config file.

    model = dict(

Based on BaseModule

Here is an example of MAEPretrainHead, which is based on BaseModule and implemented for mask image modeling task. It is required to implement loss function to generate loss, but the other helper functions are optional.

# Copyright (c) OpenMMLab. All rights reserved.
import torch
from mmengine.model import BaseModule

from mmpretrain.registry import MODELS

class MAEPretrainHead(BaseModule):
    """Head for MAE Pre-training."""

    def __init__(self,
                 loss: dict,
                 norm_pix: bool = False,
                 patch_size: int = 16) -> None:
        self.norm_pix = norm_pix
        self.patch_size = patch_size
        self.loss_module =

    def patchify(self, imgs: torch.Tensor) -> torch.Tensor:
        """Split images into non-overlapped patches."""
        p = self.patch_size
        assert imgs.shape[2] == imgs.shape[3] and imgs.shape[2] % p == 0

        h = w = imgs.shape[2] // p
        x = imgs.reshape(shape=(imgs.shape[0], 3, h, p, w, p))
        x = torch.einsum('nchpwq->nhwpqc', x)
        x = x.reshape(shape=(imgs.shape[0], h * w, p**2 * 3))
        return x

    def construct_target(self, target: torch.Tensor) -> torch.Tensor:
        """Construct the reconstruction target."""
        target = self.patchify(target)
        if self.norm_pix:
            # normalize the target image
            mean = target.mean(dim=-1, keepdim=True)
            var = target.var(dim=-1, keepdim=True)
            target = (target - mean) / (var + 1.e-6)**.5

        return target

    def loss(self, pred: torch.Tensor, target: torch.Tensor,
             mask: torch.Tensor) -> torch.Tensor:
        """Generate loss."""
        target = self.construct_target(target)
        loss = self.loss_module(pred, target, mask)

        return loss

After implementation, the following step is the same as the step-2 and step-3 in Based on ClsHead

Add a new loss

To add a new loss function, we mainly implement the forward function in the loss module. We should register the loss module as MODELS as well. In addition, it is helpful to leverage the decorator weighted_loss to weight the loss for each element. Assuming that we want to mimic a probabilistic distribution generated from another classification model, we implement an L1Loss to fulfill the purpose as below.

  1. Create a new file in mmpretrain/models/losses/

    import torch
    import torch.nn as nn
    from mmpretrain.registry import MODELS
    from .utils import weighted_loss
    def l1_loss(pred, target):
        assert pred.size() == target.size() and target.numel() > 0
        loss = torch.abs(pred - target)
        return loss
    class L1Loss(nn.Module):
        def __init__(self, reduction='mean', loss_weight=1.0):
            super(L1Loss, self).__init__()
            self.reduction = reduction
            self.loss_weight = loss_weight
        def forward(self,
            assert reduction_override in (None, 'none', 'mean', 'sum')
            reduction = (
                reduction_override if reduction_override else self.reduction)
            loss = self.loss_weight * l1_loss(
                pred, target, weight, reduction=reduction, avg_factor=avg_factor)
            return loss
  2. Import the module in mmpretrain/models/losses/

    from .l1_loss import L1Loss
    __all__ = [
        ..., 'L1Loss'
  3. Modify loss field in the head configs.

    model = dict(
            loss=dict(type='L1Loss', loss_weight=1.0),

Finally, we can combine all the new model components in a config file to create a new model for best practices. Because ResNet_CIFAR is not a ViT-based backbone, we do not implement VisionTransformerClsHead here.

model = dict(
        out_indices=(3, ),
        loss=dict(type='L1Loss', loss_weight=1.0),
        topk=(1, 5),


For convenience, the same model components could inherit from existing config files, refers to Learn about configs for more details.

Read the Docs v: dev
On Read the Docs
Project Home

Free document hosting provided by Read the Docs.