Transformer decoder in PyTorch - Model Pipeline Trace

Model Pipeline - Transformer decoder

The Transformer decoder takes encoded information and previous outputs to predict the next word in a sequence. It helps machines understand and generate language step-by-step.

Data Flow - 7 Stages

1Input tokens

1 sequence x 5 tokens→Tokenize sentence into word pieces→1 sequence x 5 tokens

[101, 2003, 1037, 2742, 102]

↓

2Embedding lookup

1 sequence x 5 tokens→Convert tokens to vectors→1 sequence x 5 tokens x 512 features

[[0.1, -0.2, ..., 0.05], ..., [0.3, 0.0, ..., -0.1]]

↓

3Positional encoding added

1 sequence x 5 tokens x 512 features→Add position info to embeddings→1 sequence x 5 tokens x 512 features

Embedding vector + position vector for each token

↓

4Masked multi-head self-attention

1 sequence x 5 tokens x 512 features→Focus on previous tokens only→1 sequence x 5 tokens x 512 features

Attention weights mask future tokens

↓

5Encoder-decoder attention

1 sequence x 5 tokens x 512 features (decoder), 1 sequence x 7 tokens x 512 features (encoder output)→Focus on encoder output relevant to decoder tokens→1 sequence x 5 tokens x 512 features

Attention aligns decoder tokens with encoder info

↓

6Feed-forward network

1 sequence x 5 tokens x 512 features→Apply two linear layers with ReLU→1 sequence x 5 tokens x 512 features

Non-linear transformation of features

↓

7Output linear + softmax

1 sequence x 5 tokens x 512 features→Project to vocabulary size and normalize→1 sequence x 5 tokens x 10000 vocabulary

[[0.01, 0.05, ..., 0.0001], ..., [0.02, 0.03, ..., 0.001]]

Training Trace - Epoch by Epoch

Loss
5.2 |*****
4.1 |****
3.3 |***
2.7 |**
2.2 |**
1.8 |*
1.5 |*
1.3 |*
1.1 |*
0.95|*

Epoch	Loss ↓	Accuracy ↑	Observation
1	5.2	0.12	High loss and low accuracy at start
2	4.1	0.25	Loss decreased, accuracy improved
3	3.3	0.38	Model learning meaningful patterns
4	2.7	0.48	Steady improvement in metrics
5	2.2	0.57	Model converging well
6	1.8	0.65	Good balance of loss and accuracy
7	1.5	0.71	Further refinement of predictions
8	1.3	0.76	Model nearing stable performance
9	1.1	0.80	Strong accuracy, low loss
10	0.95	0.83	Training converged well

Prediction Trace - 6 Layers

Layer 1: Input token embedding

Layer 2: Add positional encoding

Layer 3: Masked multi-head self-attention

Layer 4: Encoder-decoder attention

Layer 5: Feed-forward network

Layer 6: Output linear + softmax

Model Quiz - 3 Questions

Why does the decoder use masked self-attention?

A To prevent looking at future tokens during prediction

B To speed up training by ignoring some tokens

C To add position information to tokens

D To combine encoder and decoder outputs

Key Insight

The Transformer decoder learns to predict the next word by focusing only on past words and relevant encoded information. Masking future tokens ensures predictions are made step-by-step, and training shows steady improvement in accuracy as loss decreases.

Practice

1. What is the main role of the Transformer decoder in a sequence-to-sequence model?

easy

A. To reduce the dimensionality of input data

B. To encode input sequences into fixed-size vectors

C. To perform image classification tasks

D. To generate output tokens step-by-step using previous outputs and encoder info

Solution

1. Step 1: Understand the Transformer decoder purpose

   The Transformer decoder generates output sequences by attending to previous outputs and encoder outputs.

2. Step 2: Compare options with decoder role

   Only To generate output tokens step-by-step using previous outputs and encoder info describes generating output tokens step-by-step using previous outputs and encoder info, which matches the decoder's role.

3. Final Answer:

   To generate output tokens step-by-step using previous outputs and encoder info -> Option D

4. Quick Check:

   Decoder generates outputs stepwise = D [OK]

Hint: Decoder creates outputs using past outputs and encoder data [OK]

Common Mistakes:

• Confusing encoder and decoder roles
• Thinking decoder encodes inputs
• Assuming decoder reduces data size

2. Which of the following is the correct way to create a Transformer decoder layer in PyTorch?

easy

A. torch.nn.TransformerDecoderLayer(d_model=512, nhead=8)

B. torch.nn.TransformerEncoderLayer(d_model=512, nhead=8)

C. torch.nn.TransformerDecoder(d_model=512, nhead=8)

D. torch.nn.TransformerLayer(d_model=512, nhead=8)

Solution

1. Step 1: Identify the correct PyTorch class for a decoder layer

   PyTorch provides torch.nn.TransformerDecoderLayer to create a single decoder layer.

2. Step 2: Check options for correct class and parameters

   torch.nn.TransformerDecoderLayer(d_model=512, nhead=8) uses torch.nn.TransformerDecoderLayer with correct parameters. Others are either encoder layers or invalid classes.

3. Final Answer:

   torch.nn.TransformerDecoderLayer(d_model=512, nhead=8) -> Option A

4. Quick Check:

   Decoder layer class is TransformerDecoderLayer = A [OK]

Hint: Decoder layer class ends with 'DecoderLayer' in PyTorch [OK]

Common Mistakes:

• Using EncoderLayer instead of DecoderLayer
• Trying to instantiate TransformerDecoder directly
• Using non-existent class names

3. Given the code below, what is the shape of the output tensor?

import torch
from torch.nn import TransformerDecoder, TransformerDecoderLayer

layer = TransformerDecoderLayer(d_model=256, nhead=8)
decoder = TransformerDecoder(layer, num_layers=3)

memory = torch.rand(10, 32, 256)  # (source_seq_len, batch_size, d_model)
tgt = torch.rand(20, 32, 256)     # (target_seq_len, batch_size, d_model)

output = decoder(tgt, memory)

medium

A. (32, 20, 256)

B. (10, 32, 256)

C. (20, 32, 256)

D. (20, 256, 32)

Solution

1. Step 1: Understand input and output shapes for TransformerDecoder

   The decoder input shape is (target_seq_len, batch_size, d_model). The output shape matches the target input shape.

2. Step 2: Check given shapes and output

   Target input 'tgt' has shape (20, 32, 256), so output will have the same shape.

3. Final Answer:

   (20, 32, 256) -> Option C

4. Quick Check:

   Output shape matches target input shape = C [OK]

Hint: Output shape matches target input shape in decoder [OK]

Common Mistakes:

• Confusing source and target sequence lengths
• Mixing batch size and feature dimensions
• Assuming output shape changes after decoding

4. What is wrong with this PyTorch Transformer decoder usage?

import torch
from torch.nn import TransformerDecoder, TransformerDecoderLayer

layer = TransformerDecoderLayer(d_model=128, nhead=4)
decoder = TransformerDecoder(layer, num_layers=2)

memory = torch.rand(15, 16, 128)
tgt = torch.rand(15, 16, 64)  # Incorrect feature size

output = decoder(tgt, memory)

medium

A. The memory tensor has wrong batch size

B. The target tensor 'tgt' has wrong feature dimension

C. The number of decoder layers is too low

D. The nhead parameter should be equal to batch size

Solution

1. Step 1: Check feature dimensions of inputs

   The decoder expects target and memory tensors to have the same feature dimension (d_model=128).

2. Step 2: Identify mismatch in target tensor

   Target tensor 'tgt' has feature size 64, which does not match d_model=128, causing an error.

3. Final Answer:

   The target tensor 'tgt' has wrong feature dimension -> Option B

4. Quick Check:

   Feature size mismatch in target tensor = A [OK]

Hint: Check if target and memory feature sizes match d_model [OK]

Common Mistakes:

• Ignoring feature dimension mismatch
• Assuming batch size must match nhead
• Thinking number of layers causes this error

5. You want to build a Transformer decoder that generates text step-by-step using previous outputs and encoder memory. Which of the following is the best approach to implement the decoder's forward pass in PyTorch?

hard

A. Use a TransformerDecoder with masked target input to prevent attending to future tokens

B. Use a TransformerDecoder without masking so it sees all target tokens

C. Use only a TransformerEncoder for generation

D. Use a TransformerDecoder but feed the entire output sequence at once without masking

Solution

1. Step 1: Understand the need for masking in decoding

   During generation, the decoder must not look at future tokens to avoid cheating, so masking is required.

2. Step 2: Identify correct usage in PyTorch

   Using TransformerDecoder with a mask on target input prevents attending to future tokens, enabling step-by-step generation.

3. Final Answer:

   Use a TransformerDecoder with masked target input to prevent attending to future tokens -> Option A

4. Quick Check:

   Mask future tokens in decoder input = B [OK]

Hint: Mask future tokens in decoder input for stepwise generation [OK]

Common Mistakes:

• Not using masks and allowing future token attention
• Using encoder only for generation
• Feeding full output sequence without masking