Optimizers (SGD, Adam) in PyTorch - Practice Problems & Coding Challenges

Challenge - 5 Problems

🧠 Conceptual

intermediate

Difference between SGD and Adam optimizers

Which statement correctly describes a key difference between SGD and Adam optimizers?

A SGD uses adaptive learning rates based on gradient variance, while Adam uses a constant learning rate.

B SGD uses momentum by default, while Adam does not use momentum at all.

C Adam requires manual tuning of learning rate decay, but SGD automatically adjusts learning rate during training.

D Adam adapts the learning rate for each parameter individually using estimates of first and second moments, while SGD uses a fixed learning rate for all parameters.

💡 Hint

Think about how each optimizer handles learning rates for different parameters.

Adam optimizer calculates adaptive learning rates for each parameter using estimates of first (mean) and second (variance) moments of gradients. SGD applies the same learning rate to all parameters unless manually changed.

❓ Predict Output

intermediate

Output of training loss with SGD vs Adam

Given the following PyTorch training loop snippet, what will be the printed loss values after one update step using SGD and Adam optimizers respectively?

PyTorch

import torch
import torch.nn as nn
import torch.optim as optim

model = nn.Linear(2, 1)
criterion = nn.MSELoss()

x = torch.tensor([[1.0, 2.0]])
y = torch.tensor([[1.0]])

# Using SGD optimizer
optimizer_sgd = optim.SGD(model.parameters(), lr=0.1)

# Forward pass
output = model(x)
loss = criterion(output, y)

# Backward and optimize
optimizer_sgd.zero_grad()
loss.backward()
optimizer_sgd.step()

loss_sgd = criterion(model(x), y).item()

# Reset model weights
model = nn.Linear(2, 1)
optimizer_adam = optim.Adam(model.parameters(), lr=0.1)

output = model(x)
loss = criterion(output, y)
optimizer_adam.zero_grad()
loss.backward()
optimizer_adam.step()
loss_adam = criterion(model(x), y).item()

print(round(loss_sgd, 3), round(loss_adam, 3))

A Both losses decrease but Adam's loss decreases more after one step.

B Both losses remain exactly the same after one step.

C SGD loss decreases, but Adam loss increases after one step.

D Adam loss decreases, but SGD loss increases after one step.

💡 Hint

Adam usually converges faster due to adaptive learning rates.

After one update step, both optimizers reduce the loss, but Adam typically reduces it more because it adapts learning rates per parameter.

❓ Hyperparameter

advanced

Choosing learning rate for Adam optimizer

Which learning rate value is generally recommended as a good starting point for the Adam optimizer in PyTorch?

A 0.01

B 0.001

C 0.1

D 1.0

💡 Hint

Adam usually requires smaller learning rates than SGD.

Adam typically works well with a learning rate around 0.001 as a default starting point.

❓ Metrics

advanced

Effect of optimizer on training accuracy

You train the same neural network on a classification task using SGD and Adam optimizers with the same learning rate and number of epochs. Which outcome is most likely regarding training accuracy?

A SGD always achieves higher training accuracy than Adam.

B Both optimizers achieve exactly the same training accuracy at all times.

C Adam achieves higher training accuracy faster than SGD.

D Adam achieves lower training accuracy than SGD due to overfitting.

💡 Hint

Consider how adaptive learning rates affect convergence speed.

Adam usually converges faster and reaches higher accuracy earlier because it adapts learning rates per parameter.

🔧 Debug

expert

Identifying error in optimizer usage

What error will this PyTorch code raise when trying to train a model with Adam optimizer?

PyTorch

import torch
import torch.nn as nn
import torch.optim as optim

model = nn.Linear(3, 1)
optimizer = optim.Adam(model.parameters(), lr=0.01)

x = torch.randn(5, 3)
y = torch.randn(5, 1)

criterion = nn.MSELoss()

optimizer.zero_grad()
output = model(x)
loss = criterion(output, y)

loss.backward()
optimizer.step()

A No error; code runs correctly.

B RuntimeError: optimizer.zero_grad() called after loss.backward() instead of before.

C RuntimeError: Trying to backward through the graph a second time without retaining it.

D TypeError: optimizer.step() missing required positional argument.

💡 Hint

Check the order of zero_grad and backward calls.

optimizer.zero_grad() should be called before loss.backward() to clear old gradients. Calling zero_grad after backward causes incorrect gradient accumulation and runtime error.