Community Computer Vision Course documentation

Swin Transformer

Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

Swin Transformer

Introduced in the 2021 paper, Swin Transformer: Hierarchical Vision Transformer using Shifted Windows, the Swin Transformer architecture optimizes for latency and performance using a shifted window (as opposed to sliding window) approach which reduces the number of operations required. Swin is considered a hierarchical backbone for computer vision. Swin can be used for tasks like image classification.

A backbone, in terms of deep learning, is a part of a neural network that does feature extraction. Additional layers can be added to the backbone to do a variety of vision tasks. Hierarchical backbones have tiered structures, sometimes with varying resolutions. This is in contrast to the non-hierarchical plain backbone in VitDet model.

Main Highlights

Shifted windows

In the original ViT, attention is done between each patch and all other patches, which gets computationally intensive. Swin optimizes this process by reducing the normally quadratic complexity ViT into linear complexity (with respect to image size). Swin achieves this using a technique similar to CNN, where patches only attend to other patches in the same window, as opposed to all other patches, and then are gradually merged with neighboring patches. This is what makes Swin a hierarchical model.

Architecture Diagram of Swin vs Vit, taken from Swin transformer paper Image taken from Swin Transformer paper

Advantages

Computational efficiency

Swin is more performant than completely patch-based approaches like ViT.

Large datasets

SwinV2 is one of the first 3B parameter models. As training size goes up, Swin outperforms CNN. The large number of parameters enables increased capacity for learning and more complex representations.

Swin Transformer V2 (paper)

Swin Transformer V2 is a large vision model that can support up to 3B parameters and capable of training with high resolution images. It improves upon the original Swin Transformer by stabilizing training, transfer models pre-trained with low-resolution images to high-resolution tasks, and using SimMIM, a self-supervised training approach that reduces the number of labeled images required for training.

Applications in Image Restoration

SwinIR (paper)

SwinIR is a model for turning low resolution images into high resolution images based on Swin Transformer.

Swin2SR (paper)

Swin2SR is another image restoration model. It is an improvement on SwinIR by incorporating Swin Transformer V2, applying the benefits of Swin V2 like training stability and higher image resolution capacity.

Overview of PyTorch Implementation of Swin

Key parts of the implementation of Swin from the original paper is outlined below:

Swin Transformer class

  1. Initialize Parameters. Among various other dropout and normalization parameters, these parameters include:

    • window_size: Size of the windows for local self-attention.
    • ape (bool): If True, add absolute position embedding to the patch embedding.
    • fused_window_process: Optional hardware optimization
  2. Apply Patch Embedding: Similar to ViT, Images are split into non-overlapping patches and linearly embedded using Conv2D.

  3. Apply Positional Embeddings: SwinTransformer optionally uses absolute position embeddings (ape), added to the patch embeddings. Absolute positional embeddings often help the model learn to use positional information about each patch to make more informed predictions.

  4. Apply Depth Decay: Depth decay helps with regularization and preventing overfitting. Depth decay usually done by skipping layers during training. In this Swin implementation, stochastic depth decay is used, which means the deeper the layer, the higher the chance it might be skipped.

  5. Layer Construction:

    • The model is composed of multiple layers (BasicLayer) of SwinTransformerBlocks, each downsampling the feature map for hierarchical processing using PatchMerging.
    • The dimensionality of features and resolution of feature maps change across layers.
  6. Classification Head: Similar to ViT, it uses an Multi-Layer Perceptron (MLP) head for classification tasks, as defined in self.head, as the last step

class SwinTransformer(nn.Module):
    def __init__(
        self,
        img_size=224,
        patch_size=4,
        in_chans=3,
        num_classes=1000,
        embed_dim=96,
        depths=[2, 2, 6, 2],
        num_heads=[3, 6, 12, 24],
        window_size=7,
        mlp_ratio=4.0,
        qkv_bias=True,
        qk_scale=None,
        drop_rate=0.0,
        attn_drop_rate=0.0,
        drop_path_rate=0.1,
        norm_layer=nn.LayerNorm,
        ape=False,
        patch_norm=True,
        use_checkpoint=False,
        fused_window_process=False,
        **kwargs,
    ):
        super().__init__()

        self.num_classes = num_classes
        self.num_layers = len(depths)
        self.embed_dim = embed_dim
        self.ape = ape
        self.patch_norm = patch_norm
        self.num_features = int(embed_dim * 2 ** (self.num_layers - 1))
        self.mlp_ratio = mlp_ratio

        # split image into non-overlapping patches
        self.patch_embed = PatchEmbed(
            img_size=img_size,
            patch_size=patch_size,
            in_chans=in_chans,
            embed_dim=embed_dim,
            norm_layer=norm_layer if self.patch_norm else None,
        )
        num_patches = self.patch_embed.num_patches
        patches_resolution = self.patch_embed.patches_resolution
        self.patches_resolution = patches_resolution

        # absolute position embedding
        if self.ape:
            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
            trunc_normal_(self.absolute_pos_embed, std=0.02)

        self.pos_drop = nn.Dropout(p=drop_rate)

        # stochastic depth
        dpr = [
            x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))
        ]  # stochastic depth decay rule

        # build layers
        self.layers = nn.ModuleList()
        for i_layer in range(self.num_layers):
            layer = BasicLayer(
                dim=int(embed_dim * 2**i_layer),
                input_resolution=(
                    patches_resolution[0] // (2**i_layer),
                    patches_resolution[1] // (2**i_layer),
                ),
                depth=depths[i_layer],
                num_heads=num_heads[i_layer],
                window_size=window_size,
                mlp_ratio=self.mlp_ratio,
                qkv_bias=qkv_bias,
                qk_scale=qk_scale,
                drop=drop_rate,
                attn_drop=attn_drop_rate,
                drop_path=dpr[sum(depths[:i_layer]) : sum(depths[: i_layer + 1])],
                norm_layer=norm_layer,
                downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
                use_checkpoint=use_checkpoint,
                fused_window_process=fused_window_process,
            )
            self.layers.append(layer)

        self.norm = norm_layer(self.num_features)
        self.avgpool = nn.AdaptiveAvgPool1d(1)
        self.head = (
            nn.Linear(self.num_features, num_classes)
            if num_classes > 0
            else nn.Identity()
        )

        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=0.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)

    @torch.jit.ignore
    def no_weight_decay(self):
        return {"absolute_pos_embed"}

    @torch.jit.ignore
    def no_weight_decay_keywords(self):
        return {"relative_position_bias_table"}

    def forward_features(self, x):
        x = self.patch_embed(x)
        if self.ape:
            x = x + self.absolute_pos_embed
        x = self.pos_drop(x)

        for layer in self.layers:
            x = layer(x)

        x = self.norm(x)  # B L C
        x = self.avgpool(x.transpose(1, 2))  # B C 1
        x = torch.flatten(x, 1)
        return x

    def forward(self, x):
        x = self.forward_features(x)
        x = self.head(x)
        return x

Swin Transformer Block

The SwinTransformerBlock encapsulates the core operations of the Swin Transformer: local windowed attention and subsequent MLP processing. It plays a key role in enabling the Swin Transformer to efficiently handle large images by focusing on local patches while maintaining the ability to learn global representations.

Layer Components:

  • Normalization Layer 1 (self.norm1): Applied before the attention mechanism.
  • Window Attention (self.attn): Computes self-attention within local windows.
  • Drop Path (self.drop_path): Implements stochastic depth for regularization.
  • Normalization Layer 2 (self.norm2): Applied before the MLP layer.
  • MLP (mlp): A multi-layer perceptron for processing features post-attention.
  • Attention Mask (self.register_buffer): The attention mask is used during the self-attention computation to control which elements in the windowed input are allowed to interact (i.e., attend to each other). The shifted window approach helps the model to capture broader contextual information by allowing some cross-window interaction.

Swin Transformer Block’s Initialization

class SwinTransformerBlock(nn.Module):
    r"""Swin Transformer Block.

    Args:
        dim (int): Number of input channels.
        input_resolution (tuple[int]): Input resulotion.
        num_heads (int): Number of attention heads.
        window_size (int): Window size.
        shift_size (int): Shift size for SW-MSA.
        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
        drop (float, optional): Dropout rate. Default: 0.0
        attn_drop (float, optional): Attention dropout rate. Default: 0.0
        drop_path (float, optional): Stochastic depth rate. Default: 0.0
        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
        fused_window_process (bool, optional): If True, use one kernel to fused window shift & window partition for acceleration, similar for the reversed part. Default: False
    """

    def __init__(
        self,
        dim,
        input_resolution,
        num_heads,
        window_size=7,
        shift_size=0,
        mlp_ratio=4.0,
        qkv_bias=True,
        qk_scale=None,
        drop=0.0,
        attn_drop=0.0,
        drop_path=0.0,
        act_layer=nn.GELU,
        norm_layer=nn.LayerNorm,
        fused_window_process=False,
    ):
        super().__init__()
        self.dim = dim
        self.input_resolution = input_resolution
        self.num_heads = num_heads
        self.window_size = window_size
        self.shift_size = shift_size
        self.mlp_ratio = mlp_ratio
        if min(self.input_resolution) <= self.window_size:
            # if window size is larger than input resolution, we don't partition windows
            self.shift_size = 0
            self.window_size = min(self.input_resolution)
        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"

        self.norm1 = norm_layer(dim)
        self.attn = WindowAttention(
            dim,
            window_size=to_2tuple(self.window_size),
            num_heads=num_heads,
            qkv_bias=qkv_bias,
            qk_scale=qk_scale,
            attn_drop=attn_drop,
            proj_drop=drop,
        )

        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(
            in_features=dim,
            hidden_features=mlp_hidden_dim,
            act_layer=act_layer,
            drop=drop,
        )

        if self.shift_size > 0:
            # calculate attention mask for SW-MSA
            H, W = self.input_resolution
            img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
            h_slices = (
                slice(0, -self.window_size),
                slice(-self.window_size, -self.shift_size),
                slice(-self.shift_size, None),
            )
            w_slices = (
                slice(0, -self.window_size),
                slice(-self.window_size, -self.shift_size),
                slice(-self.shift_size, None),
            )
            cnt = 0
            for h in h_slices:
                for w in w_slices:
                    img_mask[:, h, w, :] = cnt
                    cnt += 1

            mask_windows = window_partition(
                img_mask, self.window_size
            )  # nW, window_size, window_size, 1
            mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
            attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
            attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(
                attn_mask == 0, float(0.0)
            )
        else:
            attn_mask = None

        self.register_buffer("attn_mask", attn_mask)
        self.fused_window_process = fused_window_process

    ### New cell ###
    def forward(self, x):
        H, W = self.input_resolution
        B, L, C = x.shape
        assert L == H * W, "input feature has wrong size"

        shortcut = x
        x = self.norm1(x)
        x = x.view(B, H, W, C)

        # cyclic shift
        if self.shift_size > 0:
            if not self.fused_window_process:
                shifted_x = torch.roll(
                    x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)
                )
                # partition windows
                x_windows = window_partition(
                    shifted_x, self.window_size
                )  # nW*B, window_size, window_size, C
            else:
                x_windows = WindowProcess.apply(
                    x, B, H, W, C, -self.shift_size, self.window_size
                )
        else:
            shifted_x = x
            # partition windows
            x_windows = window_partition(
                shifted_x, self.window_size
            )  # nW*B, window_size, window_size, C

        x_windows = x_windows.view(
            -1, self.window_size * self.window_size, C
        )  # nW*B, window_size*window_size, C

        # W-MSA/SW-MSA
        attn_windows = self.attn(
            x_windows, mask=self.attn_mask
        )  # nW*B, window_size*window_size, C

        # merge windows
        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)

        # reverse cyclic shift
        if self.shift_size > 0:
            if not self.fused_window_process:
                shifted_x = window_reverse(
                    attn_windows, self.window_size, H, W
                )  # B H' W' C
                x = torch.roll(
                    shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2)
                )
            else:
                x = WindowProcessReverse.apply(
                    attn_windows, B, H, W, C, self.shift_size, self.window_size
                )
        else:
            shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C
            x = shifted_x
        x = x.view(B, H * W, C)
        x = shortcut + self.drop_path(x)

        # Feed-forward network (FFN)
        x = x + self.drop_path(self.mlp(self.norm2(x)))

        return x

Swin Transformer Block’s Forward Pass

There are 4 key steps:

  1. Cyclic shift: The feature map is partitioned into windows via window_partition. A cyclic shift is then applied to the partitions. Cyclic shift is done by moving elements (in this case, partitions) in a sequence to the left or right, and wrapping around the elements that go off one end back to the other end. This process changes the positions of the elements relative to each other but keeps the sequence otherwise intact. For example, if you cyclically shift the sequence A, B, C, D to the right by one position, it becomes D, A, B, C.

Cyclic shift allows the model to capture relationships between adjacent windows, enhancing its ability to learn spatial contexts beyond the local scope of individual windows.

  1. Windowed attention: Perform attention using window-based multi-head self attention (W-MSA) module

  2. Merge Patches: Patches are merged via PatchMerging

  3. Reverse cyclic shift: After attention is done, the window partitioning is undone via reverse_window, and the cyclic shift operation is reversed, so that the feature map retains its original form.

class WindowAttention(nn.Module):
    """
    Args:
        dim (int): Number of input channels.
        window_size (tuple[int]): The height and width of the window.
        num_heads (int): Number of attention heads.
        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
    """

    def __init__(
        self,
        dim,
        window_size,
        num_heads,
        qkv_bias=True,
        qk_scale=None,
        attn_drop=0.0,
        proj_drop=0.0,
    ):

        super().__init__()
        self.dim = dim
        self.window_size = window_size  # Wh, Ww
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = qk_scale or head_dim**-0.5

        # define a parameter table of relative position bias
        self.relative_position_bias_table = nn.Parameter(
            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads)
        )  # 2*Wh-1 * 2*Ww-1, nH

        # get pair-wise relative position index for each token inside the window
        coords_h = torch.arange(self.window_size[0])
        coords_w = torch.arange(self.window_size[1])
        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
        relative_coords = (
            coords_flatten[:, :, None] - coords_flatten[:, None, :]
        )  # 2, Wh*Ww, Wh*Ww
        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
        relative_coords[:, :, 1] += self.window_size[1] - 1
        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
        self.register_buffer("relative_position_index", relative_position_index)

        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)

        trunc_normal_(self.relative_position_bias_table, std=0.02)
        self.softmax = nn.Softmax(dim=-1)

    def forward(self, x, mask=None):
        """
        Args:
            x: input features with shape of (num_windows*B, N, C)
            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
        """
        B_, N, C = x.shape
        qkv = (
            self.qkv(x)
            .reshape(B_, N, 3, self.num_heads, C // self.num_heads)
            .permute(2, 0, 3, 1, 4)
        )
        q, k, v = (
            qkv[0],
            qkv[1],
            qkv[2],
        )  # make torchscript happy (cannot use tensor as tuple)

        q = q * self.scale
        attn = q @ k.transpose(-2, -1)

        relative_position_bias = self.relative_position_bias_table[
            self.relative_position_index.view(-1)
        ].view(
            self.window_size[0] * self.window_size[1],
            self.window_size[0] * self.window_size[1],
            -1,
        )  # Wh*Ww,Wh*Ww,nH
        relative_position_bias = relative_position_bias.permute(
            2, 0, 1
        ).contiguous()  # nH, Wh*Ww, Wh*Ww
        attn = attn + relative_position_bias.unsqueeze(0)

        if mask is not None:
            nW = mask.shape[0]
            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(
                1
            ).unsqueeze(0)
            attn = attn.view(-1, self.num_heads, N, N)
            attn = self.softmax(attn)
        else:
            attn = self.softmax(attn)

        attn = self.attn_drop(attn)

        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x

Window Attention

WindowAttention is a window-based multi-head self attention (W-MSA) module with relative position bias. This can be used for both shifted and non-shifted windows.

class PatchMerging(nn.Module):
    r"""Patch Merging Layer.

    Args:
        input_resolution (tuple[int]): Resolution of input feature.
        dim (int): Number of input channels.
        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
    """

    def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
        super().__init__()
        self.input_resolution = input_resolution
        self.dim = dim
        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
        self.norm = norm_layer(4 * dim)

    def forward(self, x):
        """
        x: B, H*W, C
        """
        H, W = self.input_resolution
        B, L, C = x.shape
        assert L == H * W, "input feature has wrong size"
        assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even."

        x = x.view(B, H, W, C)

        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C

        x = self.norm(x)
        x = self.reduction(x)

        return x

Patch Merging Layer

Patch merging method is used for downsampling. It is used to reduce the spatial dimensions of the feature map, similar to pooling in traditional convolutional neural networks (CNNs). It helps in building hierarchical feature representations by progressively increasing the receptive field and reducing the spatial resolution.

from datasets import load_dataset
from transformers import AutoImageProcessor, SwinForImageClassification
import torch

model = SwinForImageClassification.from_pretrained(
    "microsoft/swin-tiny-patch4-window7-224"
)
image_processor = AutoImageProcessor.from_pretrained(
    "microsoft/swin-tiny-patch4-window7-224"
)

dataset = load_dataset("huggingface/cats-image")
image = dataset["test"]["image"][0]

inputs = image_processor(image, return_tensors="pt")

with torch.no_grad():
    logits = model(**inputs).logits

predicted_label_id = logits.argmax(-1).item()
predicted_label_text = model.config.id2label[predicted_label_id]

print(predicted_label_text)

Try it out

You can find the 🤗 documentation for Swin here.

Usage of pretrained Swin model for classification

Here is how to use Swin model to classify a cat image into one of the 1,000 ImageNet classes:

from datasets import load_dataset
from transformers import AutoImageProcessor, SwinForImageClassification
import torch

model = SwinForImageClassification.from_pretrained(
    "microsoft/swin-tiny-patch4-window7-224"
)
image_processor = AutoImageProcessor.from_pretrained(
    "microsoft/swin-tiny-patch4-window7-224"
)

dataset = load_dataset("huggingface/cats-image")
image = dataset["test"]["image"][0]

inputs = image_processor(image, return_tensors="pt")

with torch.no_grad():
    logits = model(**inputs).logits

predicted_label_id = logits.argmax(-1).item()
predicted_label_text = model.config.id2label[predicted_label_id]

print(predicted_label_text)
< > Update on GitHub