Why transfer learning saves time and data in TensorFlow - Experiment to Prove It

Experiment - Why transfer learning saves time and data

Problem:You want to train a model to recognize new types of images but have only a small dataset and limited time.

Current Metrics:Training accuracy: 95%, Validation accuracy: 60%, Training time: 2 hours

Issue:The model overfits due to small data and takes a long time to train from scratch.

Your Task

Use transfer learning to improve validation accuracy to at least 80% and reduce training time to under 30 minutes.

Use a pre-trained model as a base.

Fine-tune only the last layers.

Do not increase dataset size.

Hint 1

Hint 2

Hint 3

Hint 4

Solution

TensorFlow

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

# Load pre-trained MobileNetV2 without top layer
base_model = tf.keras.applications.MobileNetV2(input_shape=(128,128,3), include_top=False, weights='imagenet')
base_model.trainable = False  # Freeze base

# Add new classification head
model = models.Sequential([
    base_model,
    layers.GlobalAveragePooling2D(),
    layers.Dense(128, activation='relu'),
    layers.Dropout(0.3),
    layers.Dense(5, activation='softmax')  # Assuming 5 classes
])

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

# Assume train_ds and val_ds are prepared tf.data datasets with batch size 32
# train_ds, val_ds = ... (user provides their datasets)

history = model.fit(train_ds, epochs=5, validation_data=val_ds)

# Optionally unfreeze some base layers and fine-tune
base_model.trainable = True
fine_tune_at = 100
for layer in base_model.layers[:fine_tune_at]:
    layer.trainable = False

model.compile(optimizer=tf.keras.optimizers.Adam(1e-5), loss='sparse_categorical_crossentropy', metrics=['accuracy'])
fine_tune_history = model.fit(train_ds, epochs=5, validation_data=val_ds)

Replaced training from scratch with a pre-trained MobileNetV2 base model.

Froze base model layers to keep learned features and reduce training time.

Added a small new classification head to adapt to the new task.

Trained only the new layers initially, then fine-tuned some base layers with a low learning rate.

Results Interpretation

Before Transfer Learning: Training accuracy was very high (95%) but validation accuracy was low (60%), showing overfitting. Training took 2 hours.

After Transfer Learning: Validation accuracy improved to 82%, training accuracy is closer at 85%, showing better generalization. Training time reduced to 25 minutes.

Transfer learning uses knowledge from a model trained on a large dataset to help learn new tasks faster and with less data. Freezing base layers keeps useful features, reducing overfitting and training time.

Bonus Experiment

Try unfreezing more layers of the base model and fine-tune with a smaller learning rate to see if validation accuracy improves further.

💡 Hint

Gradually unfreeze layers starting from the top and use a very low learning rate to avoid losing pre-trained knowledge.

Practice

1. Why does transfer learning save time when training a model in TensorFlow?

easy

A. Because it requires more data to train

B. Because it uses more layers than a normal model

C. Because it trains all layers from the beginning

D. Because it reuses a pre-trained model's knowledge instead of training from scratch

Solution

1. Step 1: Understand transfer learning basics

   Transfer learning uses a model already trained on a large dataset to start learning a new task.

2. Step 2: Identify how time is saved

   Since the model already knows useful features, it needs less training time than starting fresh.

3. Final Answer:

   Because it reuses a pre-trained model's knowledge instead of training from scratch -> Option D

4. Quick Check:

   Transfer learning saves time by reusing knowledge [OK]

Hint: Transfer learning reuses knowledge, so less training time needed [OK]

Common Mistakes:

• Thinking transfer learning trains all layers from scratch
• Assuming transfer learning needs more data
• Confusing transfer learning with training a bigger model

2. Which TensorFlow code snippet correctly freezes layers in a pre-trained model for transfer learning?

easy

A. for layer in base_model.layers:
layer.trainable = True

B. base_model.trainable = True

C. for layer in base_model.layers:
layer.trainable = False

D. base_model.freeze_layers()

Solution

1. Step 1: Understand freezing layers in TensorFlow

   To freeze layers, set each layer's trainable property to False.

2. Step 2: Check code correctness

   for layer in base_model.layers: layer.trainable = False loops through layers and sets trainable to False, which is correct syntax.

3. Final Answer:

   for layer in base_model.layers: layer.trainable = False -> Option C

4. Quick Check:

   Freeze layers by setting trainable = False [OK]

Hint: Freeze layers by setting layer.trainable = False in a loop [OK]

Common Mistakes:

• Using base_model.freeze_layers() which is not a TensorFlow method
• Setting base_model.trainable = True to freeze layers
• Setting trainable = True instead of False to freeze

3. Given this TensorFlow code using transfer learning, what will be the output shape of the model's predictions?

base_model = tf.keras.applications.MobileNetV2(input_shape=(128,128,3), include_top=False, pooling='avg')
base_model.trainable = False
model = tf.keras.Sequential([
    base_model,
    tf.keras.layers.Dense(10, activation='softmax')
])
output = model(tf.random.uniform([1,128,128,3]))
print(output.shape)

medium

A. (1, 10)

B. (128, 10)

C. (1, 128, 128, 10)

D. (10,)

Solution

1. Step 1: Analyze model architecture

   The base MobileNetV2 outputs a pooled vector per input image, shape (batch_size, features). Then Dense(10) outputs 10 class scores.

2. Step 2: Determine output shape

   Input batch size is 1, so output shape is (1, 10) representing 10 class probabilities for one image.

3. Final Answer:

   (1, 10) -> Option A

4. Quick Check:

   Output shape = (batch_size, number_of_classes) = (1, 10) [OK]

Hint: Output shape matches batch size and final Dense units [OK]

Common Mistakes:

• Confusing spatial input shape with output shape
• Ignoring pooling layer effect on output shape
• Assuming output shape includes input image dimensions

4. You tried transfer learning but your model overfits quickly. Which fix below correctly addresses this issue in TensorFlow?

medium

A. Freeze more layers and add dropout layers to reduce overfitting

B. Unfreeze all layers and train with a very high learning rate

C. Remove the pre-trained base model and train from scratch

D. Use fewer training examples to speed up training

Solution

1. Step 1: Identify overfitting causes

   Overfitting happens when model learns noise; freezing layers and dropout help reduce it.

2. Step 2: Evaluate options

   Freeze more layers and add dropout layers to reduce overfitting freezes more layers and adds dropout, both good to reduce overfitting.

3. Final Answer:

   Freeze more layers and add dropout layers to reduce overfitting -> Option A

4. Quick Check:

   Freeze layers + dropout reduces overfitting [OK]

Hint: Freeze layers and add dropout to fight overfitting [OK]

Common Mistakes:

• Unfreezing all layers with high learning rate causes more overfitting
• Removing pre-trained model loses transfer learning benefits
• Using fewer examples worsens overfitting

5. You have a small dataset of 500 images for a new classification task. How does transfer learning help you save time and data compared to training a model from scratch in TensorFlow?

hard

A. By training a large model from scratch with all layers trainable

B. By using a pre-trained model and freezing most layers, you need fewer images and less training time

C. By ignoring pre-trained weights and only using the architecture

D. By increasing the dataset size artificially without changing the model

Solution

1. Step 1: Understand small dataset challenges

   Small datasets cause overfitting and long training if model is trained from scratch.

2. Step 2: Explain transfer learning benefits

   Using a pre-trained model with frozen layers reuses learned features, needing fewer images and less time.

3. Final Answer:

   By using a pre-trained model and freezing most layers, you need fewer images and less training time -> Option B

4. Quick Check:

   Transfer learning saves time and data by reusing knowledge [OK]

Hint: Freeze layers in pre-trained model to save data and time [OK]

Common Mistakes:

• Training large model from scratch with small data
• Ignoring pre-trained weights loses transfer learning advantage
• Assuming data size increase alone solves training time