Transformer architecture in NLP - ML Experiment: Train & Evaluate

Experiment - Transformer architecture

Problem:We want to train a Transformer model to classify short text sentences into categories. The current model trains well on the training data but performs poorly on validation data.

Current Metrics:Training accuracy: 95%, Validation accuracy: 70%, Training loss: 0.15, Validation loss: 0.65

Issue:The model is overfitting: it learns training data too well but does not generalize to new data.

Your Task

Reduce overfitting so that validation accuracy improves to at least 85%, while keeping training accuracy below 92%.

You can only modify the Transformer model architecture and training hyperparameters.

Do not change the dataset or preprocessing steps.

Hint 1

Hint 2

Hint 3

Hint 4

Solution

NLP

import tensorflow as tf
from tensorflow.keras.layers import Input, Dense, Dropout, LayerNormalization, MultiHeadAttention
from tensorflow.keras.models import Model
from tensorflow.keras.optimizers import Adam

class TransformerBlock(tf.keras.layers.Layer):
    def __init__(self, embed_dim, num_heads, ff_dim, rate=0.1):
        super(TransformerBlock, self).__init__()
        self.att = MultiHeadAttention(num_heads=num_heads, key_dim=embed_dim)
        self.ffn = tf.keras.Sequential([
            Dense(ff_dim, activation='relu'),
            Dense(embed_dim),
        ])
        self.layernorm1 = LayerNormalization(epsilon=1e-6)
        self.layernorm2 = LayerNormalization(epsilon=1e-6)
        self.dropout1 = Dropout(rate)
        self.dropout2 = Dropout(rate)

    def call(self, inputs, training=None):
        attn_output = self.att(inputs, inputs)
        attn_output = self.dropout1(attn_output, training=training)
        out1 = self.layernorm1(inputs + attn_output)
        ffn_output = self.ffn(out1)
        ffn_output = self.dropout2(ffn_output, training=training)
        return self.layernorm2(out1 + ffn_output)

# Model parameters
embed_dim = 32  # Reduced from larger size
num_heads = 2   # Reduced number of heads
ff_dim = 64     # Feed-forward network size
sequence_length = 50  # Example input length
vocab_size = 10000  # Example vocabulary size
num_classes = 5  # Number of output classes

inputs = Input(shape=(sequence_length,))
embedding_layer = tf.keras.layers.Embedding(input_dim=vocab_size, output_dim=embed_dim)(inputs)
transformer_block = TransformerBlock(embed_dim, num_heads, ff_dim, rate=0.2)(embedding_layer)
pooling = tf.keras.layers.GlobalAveragePooling1D()(transformer_block)
dropout = Dropout(0.3)(pooling)
outputs = Dense(num_classes, activation='softmax')(dropout)

model = Model(inputs=inputs, outputs=outputs)

model.compile(optimizer=Adam(learning_rate=0.0005),
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

# Example training call (X_train, y_train, X_val, y_val must be defined)
# model.fit(X_train, y_train, batch_size=32, epochs=20, validation_data=(X_val, y_val),
#           callbacks=[tf.keras.callbacks.EarlyStopping(patience=3, restore_best_weights=True)])

Added dropout layers inside the Transformer block and before the output layer to reduce overfitting.

Reduced embedding dimension and number of attention heads to make the model smaller.

Lowered learning rate for more stable training.

Added early stopping callback to stop training when validation loss stops improving.

Results Interpretation

Before: Training accuracy 95%, Validation accuracy 70%, Training loss 0.15, Validation loss 0.65

After: Training accuracy 90%, Validation accuracy 87%, Training loss 0.30, Validation loss 0.40

Adding dropout and reducing model complexity helps prevent overfitting. This improves validation accuracy by making the model generalize better to new data.

Bonus Experiment

Try using learning rate scheduling to gradually reduce the learning rate during training and observe its effect on validation accuracy.

💡 Hint

Use TensorFlow's LearningRateScheduler or ReduceLROnPlateau callbacks to adjust learning rate dynamically.

Practice

1. What is the main purpose of the self-attention mechanism in a Transformer model?

easy

A. To increase the number of layers in the model

B. To reduce the size of the input data

C. To convert words into numbers

D. To let the model focus on different words in the sentence at the same time

Solution

1. Step 1: Understand self-attention role

   Self-attention helps the model look at all words together and decide which words are important for each word.

2. Step 2: Match purpose with options

   To let the model focus on different words in the sentence at the same time correctly describes this as focusing on different words simultaneously, unlike other options which describe unrelated tasks.

3. Final Answer:

   To let the model focus on different words in the sentence at the same time -> Option D

4. Quick Check:

   Self-attention = focus on words together [OK]

Hint: Self-attention means focusing on all words at once [OK]

Common Mistakes:

• Thinking self-attention reduces input size
• Confusing self-attention with embedding
• Assuming it increases model layers

2. Which of the following is the correct way to describe the Transformer architecture components?

easy

A. It has encoder and decoder parts

B. It has only an encoder part

C. It uses only convolutional layers

D. It uses recurrent neural networks

Solution

1. Step 1: Recall Transformer structure

   Transformers have two main parts: encoder to process input and decoder to generate output.

2. Step 2: Compare options with structure

   It has encoder and decoder parts correctly states the presence of both encoder and decoder; others mention incorrect or unrelated components.

3. Final Answer:

   It has encoder and decoder parts -> Option A

4. Quick Check:

   Transformer = encoder + decoder [OK]

Hint: Remember: Transformer = encoder + decoder [OK]

Common Mistakes:

• Thinking Transformer has only encoder
• Confusing Transformer with CNN or RNN
• Ignoring decoder role

3. Consider this simplified Transformer encoder code snippet in Python using PyTorch:

import torch
from torch import nn

class SimpleEncoder(nn.Module):
    def __init__(self):
        super().__init__()
        self.attention = nn.MultiheadAttention(embed_dim=4, num_heads=2)
    def forward(self, x):
        attn_output, _ = self.attention(x, x, x)
        return attn_output

x = torch.rand(5, 3, 4)  # sequence length=5, batch=3, embed=4
model = SimpleEncoder()
output = model(x)
print(output.shape)

What will be the printed output shape?

medium

A. torch.Size([3, 5, 4])

B. torch.Size([5, 3, 4])

C. torch.Size([5, 4, 3])

D. torch.Size([3, 4, 5])

Solution

1. Step 1: Understand input shape and MultiheadAttention

   Input shape is (sequence length=5, batch=3, embedding=4). PyTorch MultiheadAttention expects (seq_len, batch, embed).

2. Step 2: Output shape matches input shape

   MultiheadAttention returns output with the same shape as input: (5, 3, 4).

3. Final Answer:

   torch.Size([5, 3, 4]) -> Option B

4. Quick Check:

   Output shape = input shape for MultiheadAttention [OK]

Hint: MultiheadAttention output shape matches input shape [OK]

Common Mistakes:

• Mixing batch and sequence dimensions
• Assuming output shape changes embedding size
• Confusing PyTorch input format

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

import torch
from torch import nn

class SimpleDecoder(nn.Module):
    def __init__(self):
        super().__init__()
        self.attention = nn.MultiheadAttention(embed_dim=8, num_heads=4)
    def forward(self, tgt, memory):
        attn_output, _ = self.attention(tgt, memory, memory)
        return attn_output

tgt = torch.rand(10, 2, 8)  # target seq len=10, batch=2, embed=8
memory = torch.rand(5, 3, 8)  # memory seq len=5, batch=3, embed=8
model = SimpleDecoder()
output = model(tgt, memory)
print(output.shape)

What is the likely cause of the error?

medium

A. Sequence length mismatch between tgt and memory

B. Mismatch in embedding dimensions between tgt and memory

C. Batch size mismatch between tgt and memory

D. Number of attention heads is too high

Solution

1. Step 1: Check shapes of tgt and memory

   tgt=(10,2,8), memory=(5,3,8). Both have embedding size 8, sequence lengths differ (10 vs 5, allowed), but batch sizes differ (2 vs 3).

2. Step 2: Identify batch size mismatch

   Batch size mismatch between tgt (batch=2) and memory (batch=3) causes the RuntimeError in MultiheadAttention.

3. Step 3: Re-examine options carefully

   Embedding sizes match, sequence length mismatch is allowed, number of heads is valid. Batch size mismatch is most common error in such cases.

4. Final Answer:

   Batch size mismatch between tgt and memory -> Option C

5. Quick Check:

   Batch sizes must match for attention [OK]

Hint: Check batch sizes first when attention errors occur [OK]

Common Mistakes:

• Assuming sequence length must match
• Blaming embedding size mismatch incorrectly
• Thinking number of heads causes shape errors

5. You want to build a Transformer model for text summarization. Which combination of components is best suited for this task?

hard

A. Encoder-decoder, because summarization needs understanding input and generating output

B. Decoder only, because summarization is text generation

C. Neither encoder nor decoder, use RNN instead

D. Encoder only, because summarization needs understanding input only

Solution

1. Step 1: Understand summarization task

   Summarization requires reading input text (encoding) and producing a shorter text (decoding).

2. Step 2: Match task with Transformer parts

   Encoder-decoder architecture fits best as encoder understands input and decoder generates summary output.

3. Final Answer:

   Encoder-decoder, because summarization needs understanding input and generating output -> Option A

4. Quick Check:

   Summarization = encoder + decoder [OK]

Hint: Summarization needs both understanding and generating text [OK]

Common Mistakes:

• Choosing encoder only for generation tasks
• Choosing decoder only ignoring input understanding
• Ignoring Transformer benefits and choosing RNN