Freezing layers in PyTorch - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

What is the output of this PyTorch code snippet?

Consider the following PyTorch code that freezes some layers of a model. What will be the value of requires_grad for each parameter after running this code?

PyTorch

import torch
import torch.nn as nn

class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc1 = nn.Linear(10, 5)
        self.fc2 = nn.Linear(5, 2)

    def forward(self, x):
        x = self.fc1(x)
        x = self.fc2(x)
        return x

model = SimpleModel()

# Freeze fc1 layer
for param in model.fc1.parameters():
    param.requires_grad = False

requires_grad_list = [param.requires_grad for param in model.parameters()]
print(requires_grad_list)

A [True, False, True, False]

B [True, True, False, False]

C [False, False, True, True]

D [False, True, False, True]

💡 Hint

Remember that each Linear layer has two parameters: weights and biases.

The fc1 layer has two parameters (weight and bias), both set to requires_grad=False. The fc2 layer's parameters remain trainable (requires_grad=True).

❓ Model Choice

intermediate

Which model freezing approach correctly freezes only the convolutional layers?

You want to freeze all convolutional layers in a PyTorch model but keep other layers trainable. Which code snippet correctly achieves this?

A

for name, param in model.named_parameters():
    if 'conv' in name:
        param.requires_grad = False

B

for layer in model.children():
    if 'conv' in str(layer):
        layer.requires_grad = False

C

for param in model.parameters():
    if isinstance(param, nn.Conv2d):
        param.requires_grad = False

D

for module in model.modules():
    if isinstance(module, nn.Conv2d):
        for param in module.parameters():
            param.requires_grad = False

💡 Hint

Parameters themselves are tensors, not layers. You need to access modules to check their type.

Option D correctly iterates over modules, checks if they are convolutional layers, then freezes their parameters. Option D only checks parameter names, which may be unreliable. Option D tries to check parameter type, which is invalid. Option D tries to set requires_grad on layers, which is not a valid attribute.

❓ Hyperparameter

advanced

What is the effect of freezing layers on optimizer hyperparameters?

You freeze some layers in your PyTorch model by setting requires_grad=False for their parameters. What should you do with the optimizer to avoid updating frozen parameters?

A Pass only parameters with <code>requires_grad=True</code> to the optimizer.

B Pass all model parameters to the optimizer; it will automatically skip frozen ones.

C Set the learning rate to zero for frozen parameters in the optimizer.

D Freeze layers after optimizer creation; no changes needed.

💡 Hint

Optimizer updates parameters it receives. Frozen parameters should not be included.

Only parameters with requires_grad=True should be passed to the optimizer. Passing frozen parameters wastes memory and may cause errors.

🔧 Debug

advanced

Why does training still update frozen layers?

You froze some layers by setting requires_grad=False, but after training, those layers' weights changed. What is the most likely cause?

A The loss function does not support frozen layers, causing updates.

B You forgot to filter parameters passed to the optimizer; optimizer still updates frozen layers.

C You used torch.no_grad() during training, which disables freezing.

D The model layers were not set to eval() mode, so gradients were computed anyway.

💡 Hint

Freezing layers disables gradient computation, but optimizer updates parameters it receives.

If frozen parameters are still passed to the optimizer, their values can be updated by optimizer steps even if gradients are zero or None.

🧠 Conceptual

expert

What is the impact of freezing layers on transfer learning performance?

When using transfer learning, freezing early layers of a pretrained model is common. What is the main reason for freezing these layers?

A Early layers capture general features; freezing them prevents overfitting and speeds up training.

B Early layers are specific to the original task; freezing them improves adaptation to new data.

C Freezing early layers increases model capacity by adding more trainable parameters.

D Freezing early layers reduces model size by removing those layers from computation.

💡 Hint

Think about what early layers learn in deep networks and why retraining them might be unnecessary.

Early layers learn general features like edges and textures useful across tasks. Freezing them reduces training time and risk of overfitting on small datasets.