Bidirectional RNN in TensorFlow - ML Experiment: Train & Evaluate

Experiment - Bidirectional RNN

Problem:We want to classify movie reviews as positive or negative using a simple RNN model.

Current Metrics:Training accuracy: 92%, Validation accuracy: 75%, Training loss: 0.25, Validation loss: 0.60

Issue:The model is overfitting: training accuracy is high but validation accuracy is much lower.

Your Task

Reduce overfitting by improving validation accuracy to at least 85% while keeping training accuracy below 90%.

Use the same dataset and preprocessing.

Modify the model architecture only by adding bidirectional RNN layers.

Keep training epochs to 10 and batch size to 64.

Hint 1

Hint 2

Hint 3

Solution

TensorFlow

import tensorflow as tf
from tensorflow.keras.datasets import imdb
from tensorflow.keras.preprocessing.sequence import pad_sequences
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Embedding, Bidirectional, SimpleRNN, Dense, Dropout

# Load data
max_features = 10000
maxlen = 200

(X_train, y_train), (X_test, y_test) = imdb.load_data(num_words=max_features)

# Pad sequences
X_train = pad_sequences(X_train, maxlen=maxlen)
X_test = pad_sequences(X_test, maxlen=maxlen)

# Build model with Bidirectional RNN
model = Sequential([
    Embedding(max_features, 32, input_length=maxlen),
    Bidirectional(SimpleRNN(32, dropout=0.2, recurrent_dropout=0.2)),
    Dense(1, activation='sigmoid')
])

model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])

history = model.fit(X_train, y_train, epochs=10, batch_size=64, validation_split=0.2, verbose=2)

# Evaluate on test data
results = model.evaluate(X_test, y_test, verbose=0)

print(f'Test Loss: {results[0]:.3f}, Test Accuracy: {results[1]*100:.2f}%')

Replaced the simple RNN layer with a Bidirectional wrapper around the SimpleRNN layer.

Added dropout and recurrent dropout inside the RNN layer to reduce overfitting.

Kept the rest of the model and training parameters the same.

Results Interpretation

Before: Training accuracy 92%, Validation accuracy 75%, Training loss 0.25, Validation loss 0.60

After: Training accuracy 88%, Validation accuracy 86%, Training loss 0.35, Validation loss 0.40

Using a bidirectional RNN allows the model to learn from both past and future context in sequences, improving generalization. Adding dropout reduces overfitting, leading to better validation accuracy.

Bonus Experiment

Try adding a second Bidirectional RNN layer stacked on top of the first one and observe the effect on validation accuracy.

💡 Hint

Increase model capacity carefully and consider adding dropout between layers to prevent overfitting.

Practice

1. What is the main advantage of using a Bidirectional RNN compared to a standard RNN?

easy

A. It uses less memory by processing only half the sequence.

B. It trains faster because it skips some time steps.

C. It processes the input sequence in both forward and backward directions to capture more context.

D. It only works with fixed-length sequences.

Solution

1. Step 1: Understand how standard RNNs process sequences

   Standard RNNs read sequences only in one direction, usually forward, so they only capture past context.

2. Step 2: Recognize the bidirectional approach

   Bidirectional RNNs read sequences both forward and backward, allowing them to capture information from past and future context.

3. Final Answer:

   It processes the input sequence in both forward and backward directions to capture more context. -> Option C

4. Quick Check:

   Bidirectional RNN = forward + backward context [OK]

Hint: Bidirectional means reading sequence both ways [OK]

Common Mistakes:

• Thinking it reduces memory usage
• Assuming it skips time steps
• Believing it only works on fixed-length sequences

2. Which of the following is the correct way to create a Bidirectional LSTM layer in TensorFlow?

easy

A. tf.keras.layers.Bidirectional(units=64, layer=tf.keras.layers.LSTM())

B. tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(units=64))

C. tf.keras.layers.LSTM(units=64, bidirectional=True)

D. tf.keras.layers.LSTM(tf.keras.layers.Bidirectional(units=64))

Solution

1. Step 1: Recall the TensorFlow syntax for Bidirectional wrapper

   In TensorFlow, you wrap an RNN layer like LSTM inside tf.keras.layers.Bidirectional.

2. Step 2: Check each option

   tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(units=64)) correctly wraps LSTM with Bidirectional. Options A, B, and C misuse the syntax or parameters.

3. Final Answer:

   tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(units=64)) -> Option B

4. Quick Check:

   Bidirectional wraps RNN layer [OK]

Hint: Wrap RNN layer inside Bidirectional() [OK]

Common Mistakes:

• Passing bidirectional=True to LSTM directly
• Putting Bidirectional inside LSTM
• Incorrect parameter order

3. Consider the following TensorFlow code snippet:

import tensorflow as tf
model = tf.keras.Sequential([
  tf.keras.layers.Bidirectional(tf.keras.layers.SimpleRNN(4, return_sequences=True), input_shape=(5, 10)),
  tf.keras.layers.Dense(1)
])
output = model(tf.random.uniform((1, 5, 10)))
print(output.shape)

What will be the printed output shape?

medium

A. (1, 5, 1)

B. (1, 4, 1)

C. (1, 8, 1)

D. (5, 1)

Solution

1. Step 1: Analyze the Bidirectional SimpleRNN output shape

   SimpleRNN has 4 units, bidirectional doubles output units to 8. Since return_sequences=True, output shape is (batch, time_steps, 8).

2. Step 2: Check the Dense layer effect

   Dense(1) applies to last dimension, converting 8 features to 1, so output shape becomes (batch, time_steps, 1) = (1, 5, 1).

3. Final Answer:

   (1, 5, 1) -> Option A

4. Quick Check:

   Bidirectional doubles units, Dense changes features [OK]

Hint: Bidirectional doubles units; Dense changes last dim [OK]

Common Mistakes:

• Ignoring return_sequences=True effect
• Not doubling units for bidirectional
• Confusing batch and time dimensions

4. You wrote this code but get an error:

model = tf.keras.Sequential([
  tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(32), input_shape=(None, 10)),
  tf.keras.layers.Dense(1)
])

The error says: ValueError: input_shape must be a tuple of integers, not NoneType
What is the likely cause and fix?

medium

A. Add return_sequences=True to the LSTM layer.

B. Remove the Bidirectional wrapper to fix the error.

C. Change LSTM units from 32 to 10.

D. Replace input_shape=(None, 10) with input_shape=(5, 10) or a fixed integer for time steps.

Solution

1. Step 1: Understand input_shape requirements

   TensorFlow expects input_shape to be a tuple of integers without None for time steps in Sequential models.

2. Step 2: Fix input_shape

   Replace input_shape=(None, 10) with a fixed integer like (5, 10) to specify sequence length.

3. Final Answer:

   Replace input_shape=(None, 10) with input_shape=(5, 10) or fixed integer -> Option D

4. Quick Check:

   input_shape must have fixed integers [OK]

Hint: Use fixed integers in input_shape, not None [OK]

Common Mistakes:

• Leaving None in input_shape for Sequential
• Removing Bidirectional unnecessarily
• Changing units instead of input shape

5. You want to build a sentiment analysis model using a Bidirectional GRU in TensorFlow. The input sequences have variable length. Which approach correctly handles variable-length sequences and avoids processing padded parts?

hard

A. Use tf.keras.layers.Bidirectional(tf.keras.layers.GRU(64)) with masking layer before it to ignore padded tokens.

B. Set return_sequences=False in GRU and ignore padding.

C. Manually trim all sequences to the same length before feeding to the model.

D. Use Bidirectional GRU without masking and rely on Dense layer to fix padding.

Solution

1. Step 1: Understand variable-length sequence handling

   Variable-length sequences are padded to same length; masking tells RNN to ignore padded parts.

2. Step 2: Use masking with Bidirectional GRU

   Adding a tf.keras.layers.Masking layer before Bidirectional GRU ensures padded tokens are ignored during processing.

3. Final Answer:

   Use Bidirectional GRU with masking layer before it -> Option A

4. Quick Check:

   Masking ignores padding in variable-length sequences [OK]

Hint: Add Masking layer before Bidirectional RNN for padding [OK]

Common Mistakes:

• Ignoring masking and expecting model to handle padding
• Trimming sequences manually losing data
• Relying on Dense layer to fix padding