Multi-head attention in PyTorch - Model Pipeline Trace

Model Pipeline - Multi-head attention

This pipeline shows how multi-head attention works in a transformer model. It takes input data, splits it into multiple heads to learn different parts of the data, combines the results, and improves predictions by focusing on important features.

Data Flow - 5 Stages

1Input Embeddings

64 rows x 512 features→Input token embeddings representing words in a sentence→64 rows x 512 features

[[0.1, 0.3, ..., 0.2], [0.05, 0.4, ..., 0.1], ...]

↓

2Linear Projections for Q, K, V

64 rows x 512 features→Project input into Query, Key, and Value vectors for each head→64 rows x 8 heads x 64 features

Q shape: (64, 8, 64), K shape: (64, 8, 64), V shape: (64, 8, 64)

↓

3Scaled Dot-Product Attention per Head

Q, K, V each (64 rows x 8 heads x 64 features)→Calculate attention scores, apply softmax, and weight values→64 rows x 8 heads x 64 features

Attention weights sum to 1 per query vector

↓

4Concatenate Heads

64 rows x 8 heads x 64 features→Combine all heads back into one tensor→64 rows x 512 features

Concatenated tensor shape (64, 512)

↓

5Final Linear Layer

64 rows x 512 features→Project concatenated output to final feature space→64 rows x 512 features

Output tensor shape (64, 512)

Training Trace - Epoch by Epoch

Epoch 1: ************ (1.2)
Epoch 2: ******** (0.85)
Epoch 3: ****** (0.65)
Epoch 4: **** (0.50)
Epoch 5: *** (0.40)

Epoch	Loss ↓	Accuracy ↑	Observation
1	1.2	0.45	Model starts learning, loss is high, accuracy low
2	0.85	0.62	Loss decreases, accuracy improves as attention learns
3	0.65	0.75	Model focuses better, attention heads capture useful info
4	0.50	0.82	Loss continues to drop, accuracy rises steadily
5	0.40	0.88	Model converges well, multi-head attention effective

Prediction Trace - 5 Layers

Layer 1: Input Embeddings

Layer 2: Linear Projections to Q, K, V

Layer 3: Scaled Dot-Product Attention

Layer 4: Concatenate Heads

Layer 5: Final Linear Layer

Model Quiz - 3 Questions

Why do we split the input into multiple heads in multi-head attention?

A To make the model run slower

B To reduce the input size drastically

C To learn different parts of the input data separately

D To ignore some parts of the input

Key Insight

Multi-head attention allows the model to look at different parts of the input data at once, improving its ability to understand complex patterns. Training shows steady improvement as the model learns to focus attention effectively.

Practice

1. What is the main purpose of multi-head attention in PyTorch's MultiheadAttention layer?

easy

A. To reduce the size of the input data

B. To perform simple addition of input vectors

C. To increase the number of layers in the model

D. To allow the model to focus on different parts of the input simultaneously

Solution

1. Step 1: Understand the concept of multi-head attention

   Multi-head attention splits the input into multiple parts (heads) to focus on different information simultaneously.

2. Step 2: Relate to PyTorch's MultiheadAttention layer

   This layer implements this idea to improve model understanding by attending to various input features at once.

3. Final Answer:

   To allow the model to focus on different parts of the input simultaneously -> Option D

4. Quick Check:

   Multi-head attention = focus on multiple input parts [OK]

Hint: Multi-head means multiple focus points at once [OK]

Common Mistakes:

• Thinking it reduces input size
• Confusing it with adding layers
• Assuming it just adds vectors

2. Which of the following is the correct way to create a multi-head attention layer with 8 heads and embedding dimension 64 in PyTorch?

easy

A. torch.nn.MultiheadAttention(8, 64)

B. torch.nn.MultiheadAttention(embed_dim=64, num_heads=8)

C. torch.nn.MultiheadAttention(num_heads=64, embed_dim=8)

D. torch.nn.MultiheadAttention(64)

Solution

1. Step 1: Check PyTorch MultiheadAttention parameter order

   The constructor requires embed_dim first, then num_heads as keyword arguments.

2. Step 2: Verify correct syntax

   torch.nn.MultiheadAttention(embed_dim=64, num_heads=8) correctly uses embed_dim=64 and num_heads=8. torch.nn.MultiheadAttention(num_heads=64, embed_dim=8) swaps parameters incorrectly, A and D use positional arguments wrongly.

3. Final Answer:

   torch.nn.MultiheadAttention(embed_dim=64, num_heads=8) -> Option B

4. Quick Check:

   embed_dim first, num_heads second [OK]

Hint: Remember: embed_dim first, num_heads second in constructor [OK]

Common Mistakes:

• Swapping embed_dim and num_heads
• Using positional args incorrectly
• Missing one of the required parameters

3. Given the following PyTorch code, what is the shape of the output tensor attn_output?

import torch
mha = torch.nn.MultiheadAttention(embed_dim=16, num_heads=4)
query = torch.rand(5, 3, 16)  # (sequence_length, batch_size, embed_dim)
key = torch.rand(5, 3, 16)
value = torch.rand(5, 3, 16)
attn_output, attn_weights = mha(query, key, value)

medium

A. (5, 3, 16)

B. (3, 5, 16)

C. (5, 16, 3)

D. (3, 16, 5)

Solution

1. Step 1: Understand input shapes for MultiheadAttention

   Inputs are (sequence_length, batch_size, embed_dim). Here, query shape is (5, 3, 16).

2. Step 2: Check output shape of attn_output

   Output shape matches query shape: (5, 3, 16).

3. Final Answer:

   (5, 3, 16) -> Option A

4. Quick Check:

   Output shape = query shape = (5, 3, 16) [OK]

Hint: Output shape matches query input shape [OK]

Common Mistakes:

• Mixing batch and sequence dimensions
• Confusing embed_dim position
• Assuming output shape changes with heads

4. Identify the error in this PyTorch multi-head attention code snippet:

import torch
mha = torch.nn.MultiheadAttention(embed_dim=32, num_heads=4)
query = torch.rand(10, 2, 32)
key = torch.rand(10, 2, 16)
value = torch.rand(10, 2, 32)
output, weights = mha(query, key, value)

medium

A. The embed_dim of key does not match embed_dim of query and value

B. The batch size is incorrect

C. The sequence length is too large

D. The number of heads is too high

Solution

1. Step 1: Check embed_dim consistency

   MultiheadAttention requires query, key, and value to have the same embedding dimension. Here, key has embed_dim 16, while query and value have 32.

2. Step 2: Identify the error cause

   This mismatch causes a runtime error because the attention mechanism cannot align different embedding sizes.

3. Final Answer:

   The embed_dim of key does not match embed_dim of query and value -> Option A

4. Quick Check:

   Embed dims must match for query, key, value [OK]

Hint: Check all embed_dims match before calling attention [OK]

Common Mistakes:

• Ignoring embed_dim mismatch
• Assuming batch size causes error
• Thinking sequence length is the problem

5. You want to apply multi-head attention to a batch of sentences with varying lengths. Which PyTorch feature helps you handle padding tokens correctly during attention computation?

hard

A. Increasing the number of heads to cover padding

B. Setting batch_first=True in MultiheadAttention

C. Using the key_padding_mask argument in MultiheadAttention

D. Manually removing padding tokens before attention

Solution

1. Step 1: Understand padding in variable-length sequences

   Sentences have different lengths, so padding tokens are added to make batch tensors uniform.

2. Step 2: Use key_padding_mask to ignore padding

   PyTorch's MultiheadAttention accepts key_padding_mask to mask out padded positions during attention, preventing them from affecting results.

3. Final Answer:

   Using the key_padding_mask argument in MultiheadAttention -> Option C

4. Quick Check:

   Mask padding tokens with key_padding_mask [OK]

Hint: Use key_padding_mask to ignore padded tokens [OK]

Common Mistakes:

• Removing padding manually (inefficient)
• Confusing batch_first with padding handling
• Increasing heads does not fix padding