#

Vision Transformer (ViT)

This is a PyTorch implementation of the paper An Image Is Worth 16x16 Words: Transformers For Image Recognition At Scale.

Vision transformer applies a pure transformer to images without any convolution layers. They split the image into patches and apply a transformer on patch embeddings. Patch embeddings are generated by applying a simple linear transformation to the flattened pixel values of the patch. Then a standard transformer encoder is fed with the patch embeddings, along with a classification token [CLS]. The encoding on the [CLS] token is used to classify the image with an MLP.

When feeding the transformer with the patches, learned positional embeddings are added to the patch embeddings, because the patch embeddings do not have any information about where that patch is from. The positional embeddings are a set of vectors for each patch location that get trained with gradient descent along with other parameters.

ViTs perform well when they are pre-trained on large datasets. The paper suggests pre-training them with an MLP classification head and then using a single linear layer when fine-tuning. The paper beats SOTA with a ViT pre-trained on a 300 million image dataset. They also use higher resolution images during inference while keeping the patch size the same. The positional embeddings for new patch locations are calculated by interpolating learning positional embeddings.

Here’s an experiment that trains ViT on CIFAR-10. This doesn’t do very well because it’s trained on a small dataset. It’s a simple experiment that anyone can run and play with ViTs.

45import torch
46from torch import nn
47
48from labml_helpers.module import Module
49from labml_nn.transformers import TransformerLayer
50from labml_nn.utils import clone_module_list

#

Get patch embeddings

The paper splits the image into patches of equal size and do a linear transformation on the flattened pixels for each patch.

We implement the same thing through a convolution layer, because it’s simpler to implement.

53class PatchEmbeddings(Module):

#

d_model is the transformer embeddings size
patch_size is the size of the patch
in_channels is the number of channels in the input image (3 for rgb)

65    def __init__(self, d_model: int, patch_size: int, in_channels: int):

#

71        super().__init__()

#

We create a convolution layer with a kernel size and and stride length equal to patch size. This is equivalent to splitting the image into patches and doing a linear transformation on each patch.

76        self.conv = nn.Conv2d(in_channels, d_model, patch_size, stride=patch_size)

#

x is the input image of shape [batch_size, channels, height, width]

78    def forward(self, x: torch.Tensor):

#

Apply convolution layer

83        x = self.conv(x)

#

Get the shape.

85        bs, c, h, w = x.shape

#

Rearrange to shape [patches, batch_size, d_model]

87        x = x.permute(2, 3, 0, 1)
88        x = x.view(h * w, bs, c)

#

Return the patch embeddings

91        return x

#

Add parameterized positional encodings

This adds learned positional embeddings to patch embeddings.

94class LearnedPositionalEmbeddings(Module):

#

d_model is the transformer embeddings size
max_len is the maximum number of patches

103    def __init__(self, d_model: int, max_len: int = 5_000):

#

108        super().__init__()

#

Positional embeddings for each location

110        self.positional_encodings = nn.Parameter(torch.zeros(max_len, 1, d_model), requires_grad=True)

#

x is the patch embeddings of shape [patches, batch_size, d_model]

112    def forward(self, x: torch.Tensor):

#

Get the positional embeddings for the given patches

117        pe = self.positional_encodings[x.shape[0]]

#

Add to patch embeddings and return

119        return x + pe

#

MLP Classification Head

This is the two layer MLP head to classify the image based on [CLS] token embedding.

122class ClassificationHead(Module):

#

d_model is the transformer embedding size
n_hidden is the size of the hidden layer
n_classes is the number of classes in the classification task

130    def __init__(self, d_model: int, n_hidden: int, n_classes: int):

#

136        super().__init__()

#

First layer

138        self.linear1 = nn.Linear(d_model, n_hidden)

#

Activation

140        self.act = nn.ReLU()

#

Second layer

142        self.linear2 = nn.Linear(n_hidden, n_classes)

#

x is the transformer encoding for [CLS] token

144    def forward(self, x: torch.Tensor):

#

First layer and activation

149        x = self.act(self.linear1(x))

#

Second layer

151        x = self.linear2(x)

#

154        return x

#

Vision Transformer

This combines the patch embeddings, positional embeddings, transformer and the classification head.

157class VisionTransformer(Module):

#

transformer_layer is a copy of a single transformer layer. We make copies of it to make the transformer with n_layers.
n_layers is the number of transformer layers.
patch_emb is the patch embeddings layer.
pos_emb is the positional embeddings layer.
classification is the classification head.

165    def __init__(self, transformer_layer: TransformerLayer, n_layers: int,
166                 patch_emb: PatchEmbeddings, pos_emb: LearnedPositionalEmbeddings,
167                 classification: ClassificationHead):

#

176        super().__init__()

#

Patch embeddings

178        self.patch_emb = patch_emb
179        self.pos_emb = pos_emb

#

Classification head

181        self.classification = classification

#

Make copies of the transformer layer

183        self.transformer_layers = clone_module_list(transformer_layer, n_layers)

#

[CLS] token embedding

186        self.cls_token_emb = nn.Parameter(torch.randn(1, 1, transformer_layer.size), requires_grad=True)

#

Final normalization layer

188        self.ln = nn.LayerNorm([transformer_layer.size])

#

x is the input image of shape [batch_size, channels, height, width]

190    def forward(self, x: torch.Tensor):

#

Get patch embeddings. This gives a tensor of shape [patches, batch_size, d_model]

195        x = self.patch_emb(x)

#

Add positional embeddings

197        x = self.pos_emb(x)

#

Concatenate the [CLS] token embeddings before feeding the transformer

199        cls_token_emb = self.cls_token_emb.expand(-1, x.shape[1], -1)
200        x = torch.cat([cls_token_emb, x])

#

Pass through transformer layers with no attention masking

203        for layer in self.transformer_layers:
204            x = layer(x=x, mask=None)

#

Get the transformer output of the [CLS] token (which is the first in the sequence).

207        x = x[0]

#

Layer normalization

210        x = self.ln(x)

#

Classification head, to get logits

213        x = self.classification(x)

#

216        return x