Freezing and unfreezing layers in TensorFlow - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

Output of model training with frozen layers

Consider a TensorFlow model where the first two layers are frozen before training. What will be the effect on the training loss after one epoch?

TensorFlow

import tensorflow as tf
from tensorflow.keras import layers, models

model = models.Sequential([
    layers.Dense(10, activation='relu', input_shape=(5,)),
    layers.Dense(10, activation='relu'),
    layers.Dense(1)
])

# Freeze first two layers
for layer in model.layers[:2]:
    layer.trainable = False

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

import numpy as np
x = np.random.random((100, 5))
y = np.random.random((100, 1))

history = model.fit(x, y, epochs=1, verbose=0)

print(round(history.history['loss'][0], 3))

A An error due to missing input shape specification

B A float value less than 1.0 representing the training loss

C A TypeError because frozen layers cannot be trained

D A float value greater than 10.0 representing the training loss

💡 Hint

Freezing layers means their weights do not update, but the model can still train other layers.

Freezing layers sets their trainable attribute to False, so their weights stay fixed. The model still trains the unfrozen layers, so training loss decreases and is a float value less than 1.0 after one epoch on random data.

❓ Model Choice

intermediate

Choosing layers to unfreeze for fine-tuning

You have a pretrained convolutional neural network for image classification. To fine-tune it on a new dataset, which layers should you unfreeze?

A No layers; keep all frozen

B All layers including the input layer

C Only the last few layers to adapt high-level features

D Only the first convolutional layer

💡 Hint

High-level features are learned in deeper layers.

Unfreezing only the last few layers allows the model to adapt to new data while keeping basic features intact. Unfreezing all layers may cause overfitting or require more data and time.

❓ Hyperparameter

advanced

Effect of freezing layers on learning rate choice

When you freeze most layers of a pretrained model and only train a few layers, how should you adjust the learning rate?

A Use a larger learning rate to speed up training

B Learning rate does not matter when layers are frozen

C Keep the learning rate the same as training from scratch

D Use a smaller learning rate to avoid large updates on few trainable layers

💡 Hint

Few trainable layers need careful updates to avoid destroying pretrained weights.

A smaller learning rate helps fine-tune the few trainable layers gently, preserving learned features and improving stability.

🔧 Debug

advanced

Why does unfreezing layers not update weights?

You unfreeze layers in a TensorFlow model by setting layer.trainable = True, but after training, weights do not change. What is the likely cause?

A The model was not recompiled after changing trainable flags

B The optimizer does not support training unfrozen layers

C The dataset is empty, so no training occurs

D TensorFlow does not allow unfreezing layers after compilation

💡 Hint

Changing trainable flags requires recompiling the model.

In TensorFlow, after changing layer.trainable, you must recompile the model to update the trainable variables list. Otherwise, training ignores the change.

🧠 Conceptual

expert

Impact of freezing layers on model capacity and generalization

Freezing layers in a pretrained model affects its capacity and generalization. Which statement best describes this impact?

A Freezing layers reduces model capacity but can improve generalization on small datasets

B Freezing layers increases model capacity and always improves generalization

C Freezing layers has no effect on capacity or generalization

D Freezing layers reduces generalization by limiting learning

💡 Hint

Think about how freezing limits learning but can prevent overfitting.

Freezing layers reduces the number of trainable parameters, lowering capacity. This can help generalization on small datasets by preventing overfitting.