Understanding the structure of **encoder** and **decoder blocks** is key to mastering how Transformers process and generate text. Each **encoder block** in a Transformer is designed to transform input sequences into context-rich representations, while each **decoder block** generates output sequences by attending to both previous outputs and the encoder’s representations. In sequence-to-sequence text tasks, such as translation or summarization, the encoder takes the input text and encodes it into a series of hidden states. The decoder then uses these hidden states, along with its own self-attention, to generate the target sequence step by step. This interaction between encoder and decoder blocks enables the model to capture complex dependencies in text, making Transformers highly effective for a wide range of natural language processing tasks.

The following table summarizes the sequence of operations in a Transformer encoder block and highlights their importance for text data:



| Step | Operation                      | Purpose for Text Data                                   |
|------|-------------------------------|---------------------------------------------------------|
| 1    | **Multi-head self-attention**     | Captures relationships between all tokens in the input. |
| 2    | **Add & Normalize**               | Stabilizes training and preserves information.          |
| 3    | **Feed-forward network**          | Applies non-linear transformations to each token.       |
| 4    | **Add & Normalize**               | Further stabilizes and enables deep stacking.           |


Each operation ensures that the encoder builds increasingly abstract and context-aware representations of the input text, which are essential for downstream sequence-to-sequence tasks.


import unittest
import user_code
import ast
import re   
import importlib
import csv
import unittest
import importlib

class TestTask(unittest.TestCase):
    def test_attention_and_ffn_instances(self):
        import user_code
        importlib.reload(user_code)
        block = user_code.TransformerEncoderBlock(64)
        attn = getattr(block, 'attention', None)
        ffn = getattr(block, 'ffn', None)
        _dynamic_test(
            self,
            attn is not None and hasattr(attn, '__call__') and attn.__class__.__name__ == 'MultiHeadAttention',
            "Self-attention layer is correctly initialized as MultiHeadAttention.",
            f"Expected 'MultiHeadAttention', got '{attn.__class__.__name__ if attn else attn}'",
        )
        _dynamic_test(
            self,
            ffn is not None and hasattr(ffn, '__call__') and ffn.__class__.__name__ == 'FeedForward',
            "Feed-forward layer is correctly initialized as FeedForward.",
            f"Expected 'FeedForward', got '{ffn.__class__.__name__ if ffn else ffn}'",
        )

    def test_forward_pass_identity(self):
        import user_code
        importlib.reload(user_code)
        block = user_code.TransformerEncoderBlock(32)
        input_data = [[1,2,3],[4,5,6]]
        output = block.forward(input_data)
        _dynamic_test(
            self,
            output == input_data,
            "Forward method returns correct output when dummy layers are used.",
            f"Expected output {input_data}, got {output}",
        )

def _dynamic_test(test_case, condition, success_message, failure_message):
    if condition:
        test_case._testMethodName = success_message
        test_case.assertTrue(True, success_message)
    else:
        test_case._testMethodName = failure_message
        test_case.fail(failure_message)

def normalize_text(text):
    text = text.lower()
    text = re.sub(r"\\s{2,}", " ", text)
    text = re.sub(r"\\s*([,:?])\\s*", r"\\1 ", text)
    return text.strip()

def change_var(code: str, var_name: str, value: str) -> str:
    tree = ast.parse(code)
    lines = code.splitlines()
    changed = False
    # Collect all assignment nodes to modify
    assign_nodes = [
        (i, node)
        for i, node in enumerate(tree.body)
        if isinstance(node, ast.Assign)
        and any(isinstance(target, ast.Name) and target.id == var_name for target in node.targets)
    ]

    # If nothing to change, return unmodified code
    if not assign_nodes:
        return code

    # Perform replacements for all matching assignments (from last to first to not break line offsets)
    for i, node in reversed(assign_nodes):
        start_line = node.lineno - 1
        line = lines[start_line]
        indent = ' ' * (len(line) - len(line.lstrip()))
        lines[start_line] = f"{indent}{var_name} = {value}"
        next_line = len(lines)
        for next_node in tree.body[i+1:]:
            if hasattr(next_node, 'lineno'):
                next_line = next_node.lineno - 1
                break
        if next_line > start_line + 1:
            lines[start_line+1:next_line] = []
        changed = True

    return '\\n'.join(lines) if changed else code

if __name__ == "__main__":
    unittest.main()


test_main.py

Master the essentials of Transformer models in Python for natural language processing. Discover how to build, interpret, and apply Transformers to real-world text data, focusing on practical skills and model understanding.

Explore the essentials of Transformer models, including self-attention, positional encoding, and architecture. Build a strong conceptual and practical base for advanced NLP applications.

Master the skills needed to construct core Transformer building blocks, including multi-head attention, feed-forward layers, and normalization, for effective text processing.

Discover how to use Transformers for real-world NLP tasks, visualize attention, and interpret model predictions for better text understanding.

Challenge: Structuring Encoder and Decoder Blocks

Solution