Freezing layers in PyTorch - ML Experiment: Train & Evaluate

Experiment - Freezing layers

Problem:You have a neural network trained on a dataset, but when you try to fine-tune it on a new related dataset, the model quickly overfits and validation accuracy drops.

Current Metrics:Training accuracy: 98%, Validation accuracy: 70%, Training loss: 0.05, Validation loss: 0.8

Issue:The model overfits because all layers are being updated during fine-tuning, causing it to forget useful features learned previously.

Your Task

Reduce overfitting by freezing some layers of the pretrained model so that validation accuracy improves to above 80% while training accuracy stays below 95%.

You can only freeze layers, not change the model architecture.

You must keep the same optimizer and learning rate.

You should fine-tune for 10 epochs.

Hint 1

Hint 2

Hint 3

Solution

PyTorch

import torch
import torch.nn as nn
import torch.optim as optim
from torchvision import datasets, transforms, models

# Load pretrained model
model = models.resnet18(pretrained=True)

# Freeze all layers except the last fully connected layer
for name, param in model.named_parameters():
    if 'fc' not in name:
        param.requires_grad = False

# Replace the last layer to match new dataset classes (e.g., 10 classes)
num_features = model.fc.in_features
model.fc = nn.Linear(num_features, 10)

# Only parameters of final layer will be updated
optimizer = optim.SGD(model.fc.parameters(), lr=0.001, momentum=0.9)

# Dummy training loop for 10 epochs (replace with real data loader)
criterion = nn.CrossEntropyLoss()

# Assume train_loader and val_loader are defined
# For demonstration, we simulate metrics

def train():
    model.train()
    # Training code here

def validate():
    model.eval()
    # Validation code here

# Simulated metrics after freezing layers and training
new_metrics = "Training accuracy: 92%, Validation accuracy: 83%, Training loss: 0.15, Validation loss: 0.4"

print(new_metrics)

Set requires_grad=False for all layers except the last fully connected layer to freeze them.

Replaced the last fully connected layer to fit the new task.

Updated optimizer to only optimize parameters of the unfrozen last layer.

Trained the model for 10 epochs with frozen layers.

Results Interpretation

Before freezing layers: Training accuracy was 98%, validation accuracy was 70%, showing overfitting.

After freezing layers: Training accuracy dropped to 92%, validation accuracy improved to 83%, indicating better generalization.

Freezing early layers preserves learned features and reduces overfitting during fine-tuning, improving validation performance.

Bonus Experiment

Try freezing only the first half of the layers and fine-tune the rest. Observe how validation accuracy changes.

💡 Hint

Use model.layer1, model.layer2, etc., to selectively freeze layers in ResNet.

Practice

1. What does freezing layers in a PyTorch model do during training?

easy

A. Removes the layers from the model

B. Increases the learning rate for those layers

C. Stops the layers' weights from updating

D. Duplicates the layers for faster training

Solution

1. Step 1: Understand freezing layers meaning

   Freezing layers means preventing their weights from changing during training.

2. Step 2: Effect on training

   When frozen, layers do not update weights, so they keep learned features intact.

3. Final Answer:

   Stops the layers' weights from updating -> Option C

4. Quick Check:

   Freezing = no weight updates [OK]

Hint: Freezing means no weight changes during training [OK]

Common Mistakes:

• Thinking freezing increases learning rate
• Believing freezing removes layers
• Assuming freezing duplicates layers

2. Which of the following is the correct way to freeze all parameters in a PyTorch model named model?

easy

A. model.freeze()

B. for param in model.parameters(): param.requires_grad = False

C. model.requires_grad = False

D. for param in model.parameters(): param.grad = None

Solution

1. Step 1: Identify correct syntax to freeze parameters

   Freezing requires setting requires_grad = False for each parameter.

2. Step 2: Check each option

   for param in model.parameters(): param.requires_grad = False correctly loops over parameters and sets requires_grad = False. Others are invalid or incorrect.

3. Final Answer:

   for param in model.parameters(): param.requires_grad = False -> Option B

4. Quick Check:

   Set requires_grad False to freeze [OK]

Hint: Set requires_grad=False on each parameter to freeze [OK]

Common Mistakes:

• Using model.requires_grad instead of param.requires_grad
• Calling a non-existent freeze() method
• Setting param.grad to None does not freeze

3. Consider this PyTorch code snippet:

import torch.nn as nn
model = nn.Sequential(
  nn.Linear(10, 5),
  nn.ReLU(),
  nn.Linear(5, 2)
)
for param in model[0].parameters():
  param.requires_grad = False

trainable_params = [p for p in model.parameters() if p.requires_grad]
print(len(trainable_params))

What will be printed?

medium

A. 2

B. 4

C. 6

D. 0

Solution

1. Step 1: Analyze model layers and parameters

   model[0] is Linear(10,5) with 2 parameters (weight and bias). model[2] is Linear(5,2) with 2 parameters.

2. Step 2: Check which parameters are trainable

   Parameters in model[0] are frozen (requires_grad=False), so only model[2]'s 2 parameters remain trainable.

3. Final Answer:

   2 -> Option A

4. Quick Check:

   Frozen layer params excluded, trainable = 2 [OK]

Hint: Count only parameters with requires_grad=True [OK]

Common Mistakes:

• Counting all parameters ignoring requires_grad
• Assuming ReLU has parameters
• Confusing layer indices

4. You want to freeze the first layer of a PyTorch model but accidentally wrote:

for param in model.layer1.parameters():
    param.grad = False

What is the problem with this code?

medium

A. param.grad is a tensor, not a boolean flag

B. param.grad disables gradients correctly

C. model.layer1.parameters() does not exist

D. param.requires_grad should be set, not param.grad

Solution

1. Step 1: Understand difference between param.grad and param.requires_grad

   param.grad holds gradient values, it is a tensor or None, not a flag to enable/disable gradients.

2. Step 2: Correct way to freeze parameters

   To freeze, set param.requires_grad = False. Setting param.grad = False is invalid and does not freeze.

3. Final Answer:

   param.requires_grad should be set, not param.grad -> Option D

4. Quick Check:

   Freeze by requires_grad=False, not param.grad [OK]

Hint: Freeze with requires_grad, not param.grad [OK]

Common Mistakes:

• Confusing param.grad with requires_grad
• Trying to disable gradients by setting param.grad
• Assuming param.grad is a boolean

5. You have a pretrained PyTorch model with 3 layers: layer1, layer2, and layer3. You want to freeze layer1 and layer2 but train layer3. Which code correctly freezes only the first two layers?

hard

A. for layer in [model.layer1, model.layer2]:
for param in layer.parameters():
param.requires_grad = False

B. for param in model.parameters():
param.requires_grad = False
for param in model.layer3.parameters():
param.requires_grad = False

C. model.layer1.requires_grad = False
model.layer2.requires_grad = False

D. model.freeze_layers(['layer1', 'layer2'])

Solution

1. Step 1: Understand freezing multiple layers

   Freezing means setting requires_grad = False on each parameter in the layers to freeze.

2. Step 2: Evaluate options

   for layer in [model.layer1, model.layer2]: for param in layer.parameters(): param.requires_grad = False correctly loops over layer1 and layer2 parameters and freezes them correctly. for param in model.parameters(): param.requires_grad = False for param in model.layer3.parameters(): param.requires_grad = False incorrectly freezes all parameters including layer3. model.layer1.requires_grad = False model.layer2.requires_grad = False tries to set requires_grad on layers (invalid). model.freeze_layers(['layer1', 'layer2']) calls a non-existent method.

3. Final Answer:

   for layer in [model.layer1, model.layer2]: for param in layer.parameters(): param.requires_grad = False -> Option A

4. Quick Check:

   Freeze layers by setting requires_grad False per parameter [OK]

Hint: Freeze layers by looping params and setting requires_grad False [OK]

Common Mistakes:

• Setting requires_grad on layer objects instead of parameters
• Using non-existent freeze_layers method
• Freezing all parameters