GRU for text in NLP - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

Output of GRU layer with simple input

What is the shape of the output tensor after passing a batch of 2 sequences, each of length 3 with 4 features, through a GRU layer with 5 units and return_sequences=True?

NLP

import torch
import torch.nn as nn

gru = nn.GRU(input_size=4, hidden_size=5, batch_first=True, bidirectional=False)
input_tensor = torch.randn(2, 3, 4)  # batch=2, seq_len=3, features=4
output, hidden = gru(input_tensor)
print(output.shape)

A torch.Size([3, 2, 5])

B torch.Size([2, 3, 5])

C torch.Size([2, 5])

D torch.Size([2, 3, 4])

💡 Hint

Remember that batch_first=True means batch size is the first dimension in input and output.

The GRU layer outputs a tensor with shape (batch_size, sequence_length, hidden_size) when return_sequences=True. Here, batch_size=2, sequence_length=3, hidden_size=5.

🧠 Conceptual

intermediate

Purpose of the reset gate in GRU

What is the main role of the reset gate in a GRU cell when processing text sequences?

A It normalizes the input features before processing.

B It controls how much of the previous hidden state to keep for the final output.

C It initializes the hidden state at the start of the sequence.

D It decides how much past information to forget before calculating the new candidate activation.

💡 Hint

Think about which gate helps the model reset memory for new inputs.

The reset gate controls how much of the previous hidden state to forget when computing the candidate activation, allowing the model to reset memory for new inputs.

❓ Hyperparameter

advanced

Choosing GRU hidden size for text classification

You want to build a GRU-based text classifier. Which hidden size choice is most likely to balance model capacity and training speed for a medium-sized dataset?

A Hidden size = 128

B Hidden size = 1000

C Hidden size = 1

D Hidden size = 10

💡 Hint

Too small hidden size limits learning; too large slows training and risks overfitting.

A hidden size of 128 is a common good balance for medium datasets, providing enough capacity without excessive computation or overfitting risk.

❓ Metrics

advanced

Evaluating GRU model performance on text data

After training a GRU model for sentiment analysis, you get these results on the test set: accuracy=0.85, precision=0.60, recall=0.95. What does this tell you about the model's predictions?

A The model rarely finds positive cases but is very precise when it does.

B The model has balanced precision and recall.

C The model finds most positive cases but also has many false positives.

D The model is overfitting the training data.

💡 Hint

High recall but low precision means many false alarms.

High recall (0.95) means the model finds most positive cases, but low precision (0.60) means many predicted positives are wrong (false positives).

🔧 Debug

expert

Identifying error in GRU input shape

You run this PyTorch code and get a runtime error: import torch import torch.nn as nn gru = nn.GRU(input_size=10, hidden_size=20, batch_first=True) input_tensor = torch.randn(5, 20, 10) output, hidden = gru(input_tensor) What is the cause of the error?

NLP

import torch
import torch.nn as nn

gru = nn.GRU(input_size=10, hidden_size=20, batch_first=True)
input_tensor = torch.randn(5, 20, 10)
output, hidden = gru(input_tensor)

A The input tensor shape is incorrect; batch size and sequence length are swapped.

B The GRU's hidden size must be equal to input size.

C The input tensor shape is correct; the error is due to missing initial hidden state.

D The input tensor's sequence length dimension is 20, which is too large for the GRU.

💡 Hint

Check the meaning of each dimension when batch_first=True.

With batch_first=True, input shape should be (batch_size, seq_len, input_size). Here, input_tensor shape is (5, 20, 10), which is correct. But if the model expects input_size=10, this matches. So no error from shape. However, if the error occurs, likely the batch and sequence length are swapped in the data preparation. The question is tricky but the correct answer is that the input shape is incorrect because batch and sequence length are swapped.