LSTM layer in TensorFlow - ML Experiment: Train & Evaluate

Experiment - LSTM layer

Problem:We want to classify whether the sum of numbers in a sequence exceeds 5 using an LSTM model. The current model trains well but performs poorly on new data.

Current Metrics:Training accuracy: 98%, Validation accuracy: 65%, Training loss: 0.05, Validation loss: 0.9

Issue:The model is overfitting: it learns the training data too well but does not generalize to validation data.

Your Task

Reduce overfitting so that validation accuracy improves to at least 85% while keeping training accuracy below 95%.

You can only change the model architecture and training parameters.

Do not change the dataset or preprocessing steps.

Hint 1

Hint 2

Hint 3

Hint 4

Solution

TensorFlow

import numpy as np
import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import LSTM, Dense, Dropout
from tensorflow.keras.callbacks import EarlyStopping

# Generate dummy sequential data
np.random.seed(42)
X_train = np.random.rand(1000, 10, 1)  # 1000 samples, 10 time steps, 1 feature
y_train = (np.sum(X_train, axis=1) > 5).astype(int)  # Binary target
X_val = np.random.rand(200, 10, 1)
y_val = (np.sum(X_val, axis=1) > 5).astype(int)

# Build model with dropout and fewer units
model = Sequential([
    LSTM(32, input_shape=(10,1), return_sequences=False),
    Dropout(0.3),
    Dense(1, activation='sigmoid')
])

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

# Early stopping callback
early_stop = EarlyStopping(monitor='val_loss', patience=5, restore_best_weights=True)

history = model.fit(
    X_train, y_train,
    epochs=50,
    batch_size=32,
    validation_data=(X_val, y_val),
    callbacks=[early_stop],
    verbose=0
)

# Evaluate model
train_loss, train_acc = model.evaluate(X_train, y_train, verbose=0)
val_loss, val_acc = model.evaluate(X_val, y_val, verbose=0)

print(f'Training accuracy: {train_acc*100:.2f}%, Validation accuracy: {val_acc*100:.2f}%')

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

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

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

Kept batch size at 32 and used Adam optimizer for stable training.

Results Interpretation

Before: Training accuracy 98%, Validation accuracy 65%, Training loss 0.05, Validation loss 0.9

After: Training accuracy 92%, Validation accuracy 87%, Training loss 0.2, Validation loss 0.3

Adding dropout and reducing model complexity helps prevent overfitting. Early stopping stops training before the model memorizes training data, improving validation accuracy.

Bonus Experiment

Try using a Bidirectional LSTM layer instead of a single LSTM layer and observe how it affects accuracy and overfitting.

💡 Hint

Wrap the LSTM layer with tf.keras.layers.Bidirectional and keep dropout and early stopping.

Practice

1. What is the main purpose of an LSTM layer in a neural network?

easy

A. To remember important information from past sequence data

B. To perform image classification

C. To reduce the size of input data

D. To generate random numbers

Solution

1. Step 1: Understand LSTM function

   LSTM layers are designed to handle sequence data by remembering important past information, unlike regular layers.

2. Step 2: Compare options

   Options A, B, and C describe unrelated tasks. Only To remember important information from past sequence data correctly describes LSTM's purpose.

3. Final Answer:

   To remember important information from past sequence data -> Option A

4. Quick Check:

   LSTM purpose = Remember past info [OK]

Hint: LSTM = memory for sequences, not general tasks [OK]

Common Mistakes:

• Confusing LSTM with convolutional layers
• Thinking LSTM reduces data size
• Assuming LSTM generates random outputs

2. Which of the following is the correct way to add an LSTM layer with 50 units using the `units` keyword in TensorFlow Keras?

easy

A. tf.keras.layers.LSTM(50)

B. tf.keras.layers.LSTM(units=50)

C. tf.keras.layers.LSTM(50, return_sequences=True)

D. tf.keras.layers.LSTM(50, dropout=0.2)

Solution

1. Step 1: Recall LSTM syntax

   The LSTM layer requires the number of units as a named argument 'units=50' for clarity and correctness.

2. Step 2: Evaluate options

   tf.keras.layers.LSTM(50) is allowed but less clear. Options C and D add extra parameters but the question asks for the correct basic syntax using the 'units' keyword.

3. Final Answer:

   tf.keras.layers.LSTM(units=50) -> Option B

4. Quick Check:

   LSTM syntax = units=number [OK]

Hint: Use named argument 'units' for clarity in LSTM layer [OK]

Common Mistakes:

• Omitting 'units=' keyword
• Confusing dropout with units parameter
• Adding unnecessary parameters when not asked

3. What will be the output shape of the following LSTM layer if the input shape is (batch_size, 10, 8)?

tf.keras.layers.LSTM(16, return_sequences=False)

medium

A. (batch_size, 8)

B. (batch_size, 10, 16)

C. (batch_size, 16)

D. (batch_size, 10, 8)

Solution

1. Step 1: Understand return_sequences effect

   When return_sequences=False, the LSTM returns only the last output for the sequence, so output shape is (batch_size, units).

2. Step 2: Apply to given parameters

   Units=16, so output shape is (batch_size, 16). Options B and D include sequence length, which is incorrect here.

3. Final Answer:

   (batch_size, 16) -> Option C

4. Quick Check:

   return_sequences=False output = (batch_size, units) [OK]

Hint: return_sequences=False means output shape excludes time steps [OK]

Common Mistakes:

• Assuming output keeps sequence length when return_sequences=False
• Confusing input shape with output shape
• Mixing units and input feature dimensions

4. Identify the error in this LSTM layer code snippet:

model = tf.keras.Sequential()
model.add(tf.keras.layers.LSTM(32, input_shape=(20, 5), return_sequences=False))
model.add(tf.keras.layers.LSTM(16))
model.add(tf.keras.layers.Dense(1))

medium

A. The first LSTM layer must have return_sequences=True

B. The input_shape should be (5, 20) instead of (20, 5)

C. The second LSTM layer should have return_sequences=True

D. Dense layer cannot follow LSTM layers

Solution

1. Step 1: Check LSTM stacking rules

   When stacking LSTM layers, all except the last must have return_sequences=True to pass full sequence outputs to the next layer.

2. Step 2: Analyze given code

   The first LSTM has return_sequences=False, outputting only (batch_size, 32). The second LSTM expects 3D sequence input but receives 2D, causing a shape incompatibility error.

3. Final Answer:

   The first LSTM layer must have return_sequences=True -> Option A

4. Quick Check:

   Stacked LSTM earlier layers need return_sequences=True [OK]

Hint: Only last LSTM can have return_sequences=False [OK]

Common Mistakes:

• Forgetting return_sequences=True in middle LSTM layers
• Mixing input_shape dimensions
• Thinking Dense can't follow LSTM

5. You want to build an LSTM model that outputs a sequence of the same length as input (10 time steps) with 20 features each. Which configuration achieves this?

hard

A. tf.keras.layers.LSTM(20, return_sequences=False, input_shape=(10, 20))

B. tf.keras.layers.LSTM(20, return_sequences=True, input_shape=(20, 10))

C. tf.keras.layers.LSTM(10, return_sequences=True, input_shape=(20, 20))

D. tf.keras.layers.LSTM(20, return_sequences=True, input_shape=(10, 20))

Solution

1. Step 1: Understand input and output shapes

   Input shape is (10, 20) meaning 10 time steps, 20 features. To output sequence of same length, return_sequences must be True.

2. Step 2: Check options

   tf.keras.layers.LSTM(20, return_sequences=True, input_shape=(10, 20)) matches input shape and sets return_sequences=True with units=20, so output shape is (batch_size, 10, 20). Options A, C and D have wrong return_sequences or input shapes.

3. Final Answer:

   tf.keras.layers.LSTM(20, return_sequences=True, input_shape=(10, 20)) -> Option D

4. Quick Check:

   return_sequences=True + correct input shape = desired output [OK]

Hint: Use return_sequences=True to keep sequence length [OK]

Common Mistakes:

• Setting return_sequences=False loses sequence length
• Mixing time steps and feature dimensions in input_shape
• Confusing units with input feature size