Text generation with RNN in TensorFlow - ML Experiment: Train & Evaluate

Experiment - Text generation with RNN

Problem:Generate text sequences using a Recurrent Neural Network (RNN) trained on a small dataset of Shakespeare-like text.

Current Metrics:Training loss: 0.15, Validation loss: 0.45, Training accuracy: 92%, Validation accuracy: 70%

Issue:The model is overfitting: training accuracy is high but validation accuracy is much lower, indicating poor generalization.

Your Task

Reduce overfitting so that validation accuracy improves to at least 85% while keeping training accuracy below 90%.

You can only modify the model architecture and training hyperparameters.

Do not change the dataset or preprocessing steps.

Hint 1

Hint 2

Hint 3

Hint 4

Solution

TensorFlow

import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Embedding, SimpleRNN, Dense, Dropout
from tensorflow.keras.callbacks import EarlyStopping

# Assume text data is preprocessed into sequences X_train, y_train, X_val, y_val

vocab_size = 5000  # example vocabulary size
embedding_dim = 64
rnn_units = 64

model = Sequential([
    Embedding(vocab_size, embedding_dim),
    SimpleRNN(rnn_units, return_sequences=False),
    Dropout(0.3),
    Dense(vocab_size, activation='softmax')
])

model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

early_stop = EarlyStopping(monitor='val_loss', patience=3, restore_best_weights=True)

history = model.fit(X_train, y_train,
                    epochs=30,
                    batch_size=64,
                    validation_data=(X_val, y_val),
                    callbacks=[early_stop])

Added a Dropout layer with rate 0.3 after the RNN layer to reduce overfitting.

Reduced RNN units from a higher number to 64 to simplify the model.

Added EarlyStopping callback to stop training when validation loss stops improving.

Set learning rate to 0.001 for stable training.

Results Interpretation

Before: Training accuracy 92%, Validation accuracy 70%, Validation loss 0.45

After: Training accuracy 88%, Validation accuracy 86%, Validation loss 0.30

Adding dropout and early stopping helps reduce overfitting by preventing the model from memorizing training data, improving validation accuracy and generalization.

Bonus Experiment

Try replacing the SimpleRNN layer with an LSTM layer and observe how it affects text generation quality and metrics.

💡 Hint

LSTM layers can capture longer dependencies in text, which may improve generation but can also increase training time.

Practice

1. What is the main purpose of using an RNN in text generation tasks?

easy

A. To sort words alphabetically

B. To count the number of words in a sentence

C. To learn patterns in sequences of words and predict the next word

D. To translate text into images

Solution

1. Step 1: Understand RNN function in text

   RNNs process sequences and learn how words follow each other.

2. Step 2: Identify the goal of text generation

   The goal is to predict the next word based on previous words.

3. Final Answer:

   To learn patterns in sequences of words and predict the next word -> Option C

4. Quick Check:

   RNN predicts next word = B [OK]

Hint: RNNs handle sequences and predict next items [OK]

Common Mistakes:

• Thinking RNNs sort or count words
• Confusing RNN with image processing
• Assuming RNNs translate text directly

2. Which of the following is the correct way to add an embedding layer in a TensorFlow RNN model for text generation?

easy

A. tf.keras.layers.MaxPooling1D(pool_size=2)

B. tf.keras.layers.Dense(units=embedding_dim)

C. tf.keras.layers.Conv2D(filters=embedding_dim, kernel_size=3)

D. tf.keras.layers.Embedding(input_dim=vocab_size, output_dim=embedding_dim)

Solution

1. Step 1: Identify embedding layer syntax

   Embedding layers use input_dim and output_dim to map words to vectors.

2. Step 2: Match correct TensorFlow layer

   tf.keras.layers.Embedding is the correct layer for word embeddings.

3. Final Answer:

   tf.keras.layers.Embedding(input_dim=vocab_size, output_dim=embedding_dim) -> Option D

4. Quick Check:

   Embedding layer uses input_dim and output_dim = A [OK]

Hint: Embedding layer uses input_dim and output_dim parameters [OK]

Common Mistakes:

• Using Dense instead of Embedding for word vectors
• Confusing Conv2D with text embedding
• Applying MaxPooling1D incorrectly for embeddings

3. Given the following code snippet, what will be the shape of the output from the RNN layer?

model = tf.keras.Sequential([
  tf.keras.layers.Embedding(input_dim=1000, output_dim=64),
  tf.keras.layers.SimpleRNN(128, return_sequences=False),
  tf.keras.layers.Dense(1000, activation='softmax')
])

input_shape = (32, 10)  # batch size 32, sequence length 10
output = model(tf.random.uniform(input_shape, maxval=1000, dtype=tf.int32))
print(output.shape)

medium

A. (32, 1000)

B. (32, 10, 128)

C. (10, 128)

D. (32, 128)

Solution

1. Step 1: Analyze RNN layer output shape

   SimpleRNN with return_sequences=False outputs (batch_size, units) = (32, 128).

2. Step 2: Check Dense layer output shape

   Dense layer with 1000 units outputs (32, 1000).

3. Final Answer:

   (32, 1000) -> Option A

4. Quick Check:

   Final output shape = (batch_size, vocab_size) = C [OK]

Hint: return_sequences=False means RNN outputs last step only [OK]

Common Mistakes:

• Confusing return_sequences=True output shape
• Ignoring batch size in output shape
• Mixing RNN units with Dense output units

4. Which of the following statements is CORRECT about this TensorFlow RNN model code for text generation:

model = tf.keras.Sequential([
  tf.keras.layers.Embedding(input_dim=5000, output_dim=64),
  tf.keras.layers.SimpleRNN(128, return_sequences=True),
  tf.keras.layers.Dense(5000, activation='softmax')
])

model.compile(optimizer='adam', loss='sparse_categorical_crossentropy')

# Training data: input sequences of shape (batch_size, seq_length)
# Labels: next word indices of shape (batch_size, seq_length)

model.fit(train_data, train_labels, epochs=5)

medium

A. Loss function sparse_categorical_crossentropy is incorrect for this task

B. Using return_sequences=True but labels shape is (batch_size, seq_length) which is correct

C. Labels shape should be (batch_size,) not (batch_size, seq_length)

D. Embedding layer input_dim is too small

Solution

1. Step 1: Check RNN output and labels shape

   return_sequences=True outputs sequences matching label shape (batch_size, seq_length).

2. Step 2: Verify loss function suitability

   sparse_categorical_crossentropy is correct for integer labels in classification.

3. Final Answer:

   Using return_sequences=True but labels shape is (batch_size, seq_length) which is correct -> Option B

4. Quick Check:

   return_sequences=True matches label shape = A [OK]

Hint: return_sequences=True means labels must be sequences too [OK]

Common Mistakes:

• Assuming labels must be 1D when output is sequence
• Choosing wrong loss function for integer labels
• Changing embedding input_dim unnecessarily

5. You want to generate text using a trained RNN model. Which approach correctly generates new text step-by-step?

hard

A. Feed initial seed text, predict next word, append it, then feed updated text repeatedly

B. Feed entire training dataset at once to get all predictions

C. Use the model to predict all words in one call without updating input

D. Randomly select words without using the model predictions

Solution

1. Step 1: Understand text generation process

   Text generation uses seed text and predicts one word at a time.

2. Step 2: Describe iterative prediction

   After predicting a word, append it to input and predict next word repeatedly.

3. Final Answer:

   Feed initial seed text, predict next word, append it, then feed updated text repeatedly -> Option A

4. Quick Check:

   Stepwise prediction with updated input = D [OK]

Hint: Generate text word-by-word, updating input each time [OK]

Common Mistakes:

• Trying to predict all words at once without feedback
• Ignoring seed text and using random words
• Feeding entire dataset instead of seed for generation