Learning rate for fine-tuning in TensorFlow

Introduction

Learning rate controls how much the model changes with each step. For fine-tuning, a smaller learning rate helps the model adjust gently to new data without forgetting what it learned before.

You have a pre-trained model and want to adapt it to a new but related task.

You want to improve a model's performance on a specific dataset without training from scratch.

You want to avoid large changes that could ruin the useful features learned earlier.

You want to save time by training fewer epochs with careful updates.

You want to prevent overfitting by making small, careful adjustments.

Syntax

TensorFlow

optimizer = tf.keras.optimizers.Adam(learning_rate=0.0001)
model.compile(optimizer=optimizer, loss='sparse_categorical_crossentropy', metrics=['accuracy'])

The learning_rate value is usually smaller for fine-tuning than for training from scratch.

You can use different optimizers like Adam, SGD, or RMSprop with a custom learning rate.

Examples

Using a very small learning rate for sensitive fine-tuning on a classification task.

TensorFlow

optimizer = tf.keras.optimizers.Adam(learning_rate=0.00001)
model.compile(optimizer=optimizer, loss='categorical_crossentropy', metrics=['accuracy'])

Using SGD optimizer with momentum and a small learning rate for fine-tuning.

TensorFlow

optimizer = tf.keras.optimizers.SGD(learning_rate=0.0001, momentum=0.9)
model.compile(optimizer=optimizer, loss='sparse_categorical_crossentropy', metrics=['accuracy'])

Sample Model

This example shows how to load a pre-trained model, freeze it, add a new head, and train with a small learning rate. Then it unfreezes the base model and fine-tunes with an even smaller learning rate.

TensorFlow

import tensorflow as tf

# Load a pre-trained MobileNetV2 model without the top layer
base_model = tf.keras.applications.MobileNetV2(input_shape=(96, 96, 3), include_top=False, weights='imagenet')
base_model.trainable = False  # Freeze base model

# Add new classification head
inputs = tf.keras.Input(shape=(96, 96, 3))
x = base_model(inputs, training=False)
x = tf.keras.layers.GlobalAveragePooling2D()(x)
x = tf.keras.layers.Dense(10)(x)  # 10 classes
outputs = x
model = tf.keras.Model(inputs, outputs)

# Compile with a small learning rate for fine-tuning
optimizer = tf.keras.optimizers.Adam(learning_rate=0.0001)
model.compile(optimizer=optimizer, loss='sparse_categorical_crossentropy', metrics=['accuracy'])

# Create dummy data
import numpy as np
x_train = np.random.random((20, 96, 96, 3))
y_train = np.random.randint(0, 10, 20)

# Train the model
history = model.fit(x_train, y_train, epochs=2, batch_size=5, verbose=2)

# Unfreeze base model for fine-tuning
base_model.trainable = True

# Recompile with even smaller learning rate
optimizer_finetune = tf.keras.optimizers.Adam(learning_rate=0.00001)
model.compile(optimizer=optimizer_finetune, loss='sparse_categorical_crossentropy', metrics=['accuracy'])

# Continue training
history_finetune = model.fit(x_train, y_train, epochs=2, batch_size=5, verbose=2)

Important Notes

Start with a small learning rate to avoid destroying the pre-trained features.

After initial training, unfreeze some layers and use an even smaller learning rate for fine-tuning.

Monitor training metrics to check if the learning rate is too high or too low.

Summary

Use a smaller learning rate when fine-tuning to make gentle updates.

Freeze the base model first, then unfreeze and fine-tune with a smaller learning rate.

Choose learning rates carefully to balance learning speed and stability.

Practice

1. Why is it recommended to use a smaller learning rate when fine-tuning a pre-trained TensorFlow model?

easy

A. To increase the batch size automatically

B. To speed up training drastically

C. To make gentle updates and avoid destroying learned features

D. To skip the need for validation data

Solution

1. Step 1: Understand fine-tuning purpose

   Fine-tuning adjusts a pre-trained model slightly to new data without losing learned knowledge.

2. Step 2: Reason about learning rate effect

   A smaller learning rate makes small changes, preserving useful features learned before.

3. Final Answer:

   To make gentle updates and avoid destroying learned features -> Option C

4. Quick Check:

   Small learning rate = gentle fine-tuning [OK]

Hint: Small learning rate prevents big changes during fine-tuning [OK]

Common Mistakes:

• Using a large learning rate that overwrites learned features
• Ignoring the need to freeze layers before fine-tuning
• Assuming learning rate affects batch size

2. Which of the following TensorFlow code snippets correctly sets a smaller learning rate for fine-tuning?

easy

A. optimizer = tf.keras.optimizers.Adam(learning_rate=0.0001)

B. optimizer = tf.keras.optimizers.Adam(learning_rate=0.001)

C. optimizer = tf.keras.optimizers.Adam(learning_rate=0.01)

D. optimizer = tf.keras.optimizers.Adam(learning_rate=0.1)

Solution

1. Step 1: Identify typical fine-tuning learning rates

   Fine-tuning usually uses smaller rates like 0.0001 to avoid big weight changes.

2. Step 2: Compare given learning rates

   0.0001 is smaller than 0.001, 0.01, and 0.1, making it suitable for fine-tuning.

3. Final Answer:

   optimizer = tf.keras.optimizers.Adam(learning_rate=0.0001) -> Option A

4. Quick Check:

   Smaller learning rate = 0.0001 [OK]

Hint: Pick the smallest learning rate value for fine-tuning [OK]

Common Mistakes:

• Choosing a learning rate too large for fine-tuning
• Confusing learning rate with batch size
• Not using tf.keras.optimizers.Adam correctly

3. Consider this TensorFlow code for fine-tuning a model:

base_model.trainable = True
optimizer = tf.keras.optimizers.Adam(learning_rate=0.01)
model.compile(optimizer=optimizer, loss='sparse_categorical_crossentropy', metrics=['accuracy'])
history = model.fit(train_data, epochs=3)

What is the likely outcome of using this learning rate during fine-tuning?

medium

A. The model will fine-tune smoothly with stable accuracy improvements

B. The model may have unstable training and lose pre-trained knowledge

C. The training will be very slow but stable

D. The model will skip training and keep old weights

Solution

1. Step 1: Analyze learning rate size

   Learning rate 0.01 is relatively large for fine-tuning, which risks big weight changes.

2. Step 2: Predict training behavior

   Large learning rate can cause unstable training and overwrite useful pre-trained features.

3. Final Answer:

   The model may have unstable training and lose pre-trained knowledge -> Option B

4. Quick Check:

   Large learning rate = unstable fine-tuning [OK]

Hint: Large learning rate can harm fine-tuning stability [OK]

Common Mistakes:

• Assuming higher learning rate always improves training speed
• Ignoring the effect of unfreezing the base model
• Thinking loss function controls learning rate impact

4. You wrote this code to fine-tune a TensorFlow model but notice training loss is not decreasing:

base_model.trainable = True
optimizer = tf.keras.optimizers.Adam(learning_rate=0.1)
model.compile(optimizer=optimizer, loss='categorical_crossentropy')
history = model.fit(train_data, epochs=5)

What is the main issue and how to fix it?

medium

A. Learning rate is too high; reduce it to 0.0001

B. Loss function is wrong; use 'sparse_categorical_crossentropy'

C. Base model should be frozen; set trainable = False

D. Epochs are too few; increase to 50

Solution

1. Step 1: Identify learning rate problem

   Learning rate 0.1 is very high for fine-tuning, causing unstable or no learning.

2. Step 2: Suggest fix

   Lowering learning rate to 0.0001 allows gentle updates and loss decrease.

3. Final Answer:

   Learning rate is too high; reduce it to 0.0001 -> Option A

4. Quick Check:

   High learning rate blocks loss decrease [OK]

Hint: Lower learning rate if loss doesn't improve during fine-tuning [OK]

Common Mistakes:

• Changing loss function unnecessarily
• Freezing base model after unfreezing
• Increasing epochs without fixing learning rate

5. You want to fine-tune a pre-trained TensorFlow model on a small dataset. Which learning rate strategy is best to balance learning speed and stability?

hard

A. Start with a small learning rate (e.g., 0.0001), then gradually increase it

B. Use a large learning rate (e.g., 0.1) from the start for faster training

C. Train all layers from scratch with a medium learning rate (e.g., 0.01)

D. Freeze the base model, train the new layers with a higher learning rate, then unfreeze and fine-tune with a smaller learning rate

Solution

1. Step 1: Understand fine-tuning best practice

   Freeze base model first to train new layers safely with higher learning rate.

2. Step 2: Fine-tune with smaller learning rate after unfreezing

   Unfreeze base model and use smaller learning rate to gently adjust pre-trained weights.

3. Final Answer:

   Freeze the base model, train the new layers with a higher learning rate, then unfreeze and fine-tune with a smaller learning rate -> Option D

4. Quick Check:

   Freeze first, then fine-tune with small learning rate [OK]

Hint: Freeze base, train new layers high LR, then fine-tune low LR [OK]

Common Mistakes:

• Using large learning rate from start causing unstable training
• Training all layers from scratch wasting pre-trained knowledge
• Not freezing base model before fine-tuning