Vision Transformer (ViT) in Computer Vision - Model Pipeline Trace

Model Pipeline - Vision Transformer (ViT)

The Vision Transformer (ViT) model splits an image into small patches, turns them into a sequence, and uses a transformer to learn patterns for image classification.

Data Flow - 8 Stages

1Input Image

1 image x 224 height x 224 width x 3 channels→Original color image input→1 image x 224 x 224 x 3

A photo of a cat with RGB colors

↓

2Patch Extraction

1 x 224 x 224 x 3→Split image into 16x16 patches→1 x 196 patches x (16*16*3=768) features

Image split into 196 small square patches, each flattened to 768 numbers

↓

3Linear Projection

1 x 196 x 768→Project each patch to embedding space→1 x 196 x 768

Each patch converted to a 768-length vector

↓

4Add Position Embeddings

1 x 196 x 768→Add position info to each patch embedding→1 x 196 x 768

Patch vectors now include location info

↓

5Classification Token

1 x 196 x 768→Add special classification token to the sequence→1 x 197 x 768

Sequence with classification token prepended

↓

6Transformer Encoder

1 x 197 x 768→Process sequence with multi-head self-attention layers→1 x 197 x 768

Model learns relationships between patches and classification token

↓

7Classification Token Extraction

1 x 197 x 768→Extract classification token output for classification→1 x 768

Single vector representing whole image

↓

8MLP Head

1 x 768→Feedforward network to predict class probabilities→1 x 1000 (for ImageNet classes)

Output probabilities for 1000 classes

Training Trace - Epoch by Epoch

Loss
2.3 |*         
1.5 |  *       
0.9 |    *     
0.6 |      *   
0.45|       *  
    +----------
     1  5 10 15 20 Epochs

Epoch	Loss ↓	Accuracy ↑	Observation
1	2.30	0.12	Starting training, loss high, accuracy low
5	1.50	0.45	Model learning basic features, accuracy improving
10	0.90	0.70	Good progress, model captures complex patterns
15	0.60	0.82	Loss decreasing steadily, accuracy high
20	0.45	0.88	Training converging, model performs well

Prediction Trace - 5 Layers

Layer 1: Patch Extraction

Layer 2: Linear Projection + Position Embedding + Classification Token

Layer 3: Transformer Encoder

Layer 4: Classification Token Extraction

Layer 5: MLP Head

Model Quiz - 3 Questions

What is the purpose of splitting the image into patches in ViT?

A To reduce the image size for faster training

B To increase the number of color channels

C To convert the image into a sequence for the transformer

D To remove noise from the image

Key Insight

Vision Transformer uses a sequence approach by splitting images into patches and applying transformer attention, enabling it to learn complex image patterns effectively, as shown by steady loss decrease and accuracy increase during training.

Practice

1. What is the main purpose of splitting an image into patches in a Vision Transformer (ViT)?

easy

A. To reduce the image size by cropping

B. To convert the image into smaller parts that the transformer can process as tokens

C. To apply convolution filters on each patch separately

D. To increase the image resolution for better detail

Solution

1. Step 1: Understand ViT input processing

   ViT splits images into fixed-size patches to treat each patch like a word token in language models.

2. Step 2: Purpose of patch splitting

   This allows the transformer to process image patches as a sequence, enabling attention mechanisms to learn relationships.

3. Final Answer:

   To convert the image into smaller parts that the transformer can process as tokens -> Option B

4. Quick Check:

   Image patches = tokens for transformer [OK]

Hint: Think of patches as words in a sentence for the transformer [OK]

Common Mistakes:

• Confusing patch splitting with image resizing
• Thinking patches are processed by convolution
• Assuming patches increase image resolution

2. Which of the following is the correct way to add a class token to the patch embeddings in ViT using Python-like pseudocode?

easy

A. patches = torch.cat([class_token, patches], dim=1)

B. patches = torch.cat([patches, class_token], dim=1)

C. patches = torch.cat([patches, class_token], dim=0)

D. patches = torch.cat([class_token, patches], dim=0)

Solution

1. Step 1: Understand tensor concatenation dimension

   Patch embeddings are sequences along dimension 1 (batch, seq, embed); class token must be prepended along this dimension.

2. Step 2: Correct concatenation syntax

   Using torch.cat with dim=1 adds class_token at the start of the sequence correctly.

3. Final Answer:

   patches = torch.cat([class_token, patches], dim=1) -> Option A

4. Quick Check:

   Class token prepended along sequence dim = patches = torch.cat([class_token, patches], dim=1) [OK]

Hint: Class token goes first, concat along sequence dimension (dim=1) [OK]

Common Mistakes:

• Concatenating along wrong dimension (dim=0)
• Appending class token at the end instead of start
• Mixing order of tensors in concat

3. Given the following simplified ViT patch embedding code, what is the shape of patch_embeddings after processing a batch of 8 images of size 32x32 with patch size 8 and embedding dimension 64?

patch_size = 8
embedding_dim = 64
batch_size = 8
image_size = 32
num_patches = (image_size // patch_size) ** 2
patch_embeddings = torch.randn(batch_size, num_patches, embedding_dim)

medium

A. (16, 8, 64)

B. (8, 64, 16)

C. (8, 8, 64)

D. (8, 16, 64)

Solution

1. Step 1: Calculate number of patches

   Number of patches = (32 / 8)^2 = 4^2 = 16 patches per image.

2. Step 2: Determine patch_embeddings shape

   Shape is (batch_size, num_patches, embedding_dim) = (8, 16, 64).

3. Final Answer:

   (8, 16, 64) -> Option D

4. Quick Check:

   Batch=8, patches=16, embed=64 [OK]

Hint: Calculate patches as (image/patch)^2, then batch x patches x embed [OK]

Common Mistakes:

• Mixing embedding dimension and patch count order
• Calculating patches incorrectly
• Confusing batch size with patch count

4. You have this ViT code snippet that throws an error:

class_token = torch.randn(1, 1, 64)
patches = torch.randn(8, 16, 64)
input_seq = torch.cat([class_token, patches], dim=1)

What is the cause of the error?

medium

A. Embedding dimensions do not match

B. Wrong concatenation dimension; should be dim=0

C. class_token shape should be (8, 1, 64) to match batch size

D. Dimension mismatch because class_token sequence size is 1 but patches sequence size is 16

Solution

1. Step 1: Check batch size compatibility

   class_token has batch size 1, patches have batch size 8; they must match for concatenation.

2. Step 2: Fix class_token shape

   class_token should be repeated or created with shape (8, 1, 64) to match patches batch size.

3. Final Answer:

   class_token shape should be (8, 1, 64) to match batch size -> Option C

4. Quick Check:

   Batch sizes must match for concat [OK]

Hint: Match batch sizes before concatenating tensors [OK]

Common Mistakes:

• Ignoring batch size mismatch
• Changing wrong concat dimension
• Assuming embedding dims cause error

5. In a Vision Transformer model, why is the class token important for image classification tasks?

hard

A. It aggregates information from all patches via attention to produce a final image representation

B. It stores the positional information of patches

C. It applies convolution to patches before transformer layers

D. It normalizes the patch embeddings before feeding to the transformer

Solution

1. Step 1: Understand class token role

   The class token is a special token that attends to all patch tokens and gathers their information.

2. Step 2: Use in classification

   After transformer layers, the class token embedding is used as the image's summary representation for classification.

3. Final Answer:

   It aggregates information from all patches via attention to produce a final image representation -> Option A

4. Quick Check:

   Class token = image summary for classification [OK]

Hint: Class token collects info from patches for final decision [OK]

Common Mistakes:

• Confusing class token with positional encoding
• Thinking class token applies convolution
• Assuming class token normalizes embeddings