GRU for text in NLP - ML Experiment: Train & Evaluate

Experiment - GRU for text

Problem:We want to classify movie reviews as positive or negative using a GRU (Gated Recurrent Unit) model on text data.

Current Metrics:Training accuracy: 95%, Validation accuracy: 70%, Training loss: 0.15, Validation loss: 0.60

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

Your Task

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

Keep using the GRU architecture.

Do not change the dataset or preprocessing steps.

You can adjust hyperparameters and add regularization techniques.

Hint 1

Hint 2

Hint 3

Hint 4

Solution

NLP

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

# Load data
max_features = 10000
maxlen = 100
(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 dropout and fewer units
model = Sequential([
    Embedding(max_features, 64, input_length=maxlen),
    GRU(32, dropout=0.3, recurrent_dropout=0.3),
    Dropout(0.3),
    Dense(1, activation='sigmoid')
])

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

# Early stopping callback
early_stop = tf.keras.callbacks.EarlyStopping(monitor='val_loss', patience=3, restore_best_weights=True)

# Train model
history = model.fit(
    X_train, y_train,
    epochs=20,
    batch_size=64,
    validation_split=0.2,
    callbacks=[early_stop],
    verbose=2
)

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

print(f'Test accuracy: {test_acc:.2f}', f'Test loss: {test_loss:.2f}')

Reduced GRU units from 64 to 32 to simplify the model.

Added dropout inside GRU and a separate dropout layer to reduce overfitting.

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

Lowered learning rate to 0.001 for smoother training.

Results Interpretation

Before: Training accuracy 95%, Validation accuracy 70%, Validation loss 0.60

After: Training accuracy 90%, Validation accuracy 86%, Validation loss 0.40

Adding dropout and early stopping helps reduce overfitting in GRU models for text. This leads to better validation accuracy and more reliable predictions on new data.

Bonus Experiment

Try replacing the GRU layer with an LSTM layer and compare the validation accuracy and training time.

💡 Hint

LSTM can capture longer dependencies but may train slower. Adjust dropout and units similarly.

Practice

1. What is the main advantage of using a GRU (Gated Recurrent Unit) in text processing tasks?

easy

A. It helps the model remember important information over time while ignoring less important details.

B. It increases the size of the input text automatically.

C. It converts text into images for better analysis.

D. It removes all punctuation from the text before processing.

Solution

1. Step 1: Understand GRU's role in memory

   GRU units are designed to keep important information from previous steps and forget irrelevant data, helping with sequence tasks like text.

2. Step 2: Compare options to GRU function

   Only It helps the model remember important information over time while ignoring less important details. correctly describes this memory feature; others describe unrelated or incorrect functions.

3. Final Answer:

   It helps the model remember important information over time while ignoring less important details. -> Option A

4. Quick Check:

   GRU memory feature = A [OK]

Hint: GRU remembers key info, forgets noise in sequences [OK]

Common Mistakes:

• Thinking GRU changes input size
• Confusing GRU with data preprocessing
• Assuming GRU outputs images

2. Which of the following is the correct way to define a GRU layer in Python using PyTorch for text input with embedding size 100 and hidden size 50?

easy

A. nn.GRU(hidden_size=100, input_size=50)

B. nn.GRU(50, 100)

C. nn.GRU(input_size=100, hidden_size=50)

D. nn.GRU(100)

Solution

1. Step 1: Recall PyTorch GRU parameters

   PyTorch GRU expects input_size first (embedding size), then hidden_size (number of features in hidden state).

2. Step 2: Match parameters to given sizes

   Embedding size is 100, hidden size is 50, so nn.GRU(input_size=100, hidden_size=50) is correct.

3. Final Answer:

   nn.GRU(input_size=100, hidden_size=50) -> Option C

4. Quick Check:

   input_size=100, hidden_size=50 = B [OK]

Hint: Input size first, hidden size second in nn.GRU() [OK]

Common Mistakes:

• Swapping input_size and hidden_size
• Using positional args incorrectly
• Omitting required parameters

3. Given the following PyTorch code snippet, what will be the shape of the output tensor after passing input through the GRU?

import torch
import torch.nn as nn

gru = nn.GRU(input_size=10, hidden_size=20, batch_first=True)
input = torch.randn(5, 7, 10)  # batch=5, seq_len=7, input_size=10
output, hidden = gru(input)
print(output.shape)

medium

A. (7, 5, 20)

B. (5, 7, 20)

C. (5, 20, 7)

D. (5, 7, 10)

Solution

1. Step 1: Understand GRU output shape with batch_first=true

   Output shape is (batch_size, sequence_length, hidden_size) when batch_first=true.

2. Step 2: Match given input sizes

   Input batch=5, seq_len=7, hidden_size=20, so output shape is (5, 7, 20).

3. Final Answer:

   (5, 7, 20) -> Option B

4. Quick Check:

   Output shape = (batch, seq_len, hidden_size) = A [OK]

Hint: With batch_first=true, output shape is (batch, seq_len, hidden) [OK]

Common Mistakes:

• Confusing batch and sequence dimensions
• Ignoring batch_first=true effect
• Assuming output shape equals input shape

4. You wrote this code to create a GRU for text classification but get a runtime error:

gru = nn.GRU(input_size=50, hidden_size=100)
input = torch.randn(32, 10, 100)  # batch=32, seq_len=10, input_size=100
output, hidden = gru(input)

What is the likely cause of the error?

medium

A. Input size 100 does not match GRU input_size 50

B. Batch size 32 is too large for GRU

C. Sequence length 10 is invalid for GRU

D. GRU requires input to be 2D tensor, not 3D

Solution

1. Step 1: Check GRU input_size vs input tensor last dimension

   GRU expects input_size=50, but input tensor last dimension is 100, causing mismatch.

2. Step 2: Understand tensor shape requirements

   GRU input shape should be (batch, seq_len, input_size). Here input_size dimension must match GRU's input_size parameter.

3. Final Answer:

   Input size 100 does not match GRU input_size 50 -> Option A

4. Quick Check:

   Input size mismatch = C [OK]

Hint: Match input tensor last dim to GRU input_size [OK]

Common Mistakes:

• Blaming batch size for error
• Thinking sequence length is invalid
• Assuming GRU only accepts 2D input

5. You want to build a GRU-based model to classify movie reviews as positive or negative. Your dataset has variable-length reviews. Which approach best handles variable-length sequences with a GRU in PyTorch?

hard

A. Convert text to images and use CNN instead of GRU.

B. Truncate all sequences to length 1 and feed to GRU.

C. Feed raw sequences directly without padding or packing.

D. Pad all sequences to the same length and use pack_padded_sequence before GRU.

Solution

1. Step 1: Understand variable-length sequence handling

   GRU requires fixed-length inputs or packed sequences to handle variable lengths efficiently.

2. Step 2: Use padding and packing for variable-length inputs

   Padding sequences to max length and using pack_padded_sequence lets GRU ignore padded parts during processing.

3. Final Answer:

   Pad all sequences to the same length and use pack_padded_sequence before GRU. -> Option D

4. Quick Check:

   Padding + pack_padded_sequence = D [OK]

Hint: Pad sequences and pack before GRU for variable lengths [OK]

Common Mistakes:

• Truncating sequences too short loses info
• Feeding raw variable-length sequences causes errors
• Switching to CNN ignores GRU benefits