Checkpoint with optimizer state in PyTorch - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

What is the output of this PyTorch checkpoint loading code?

Consider the following PyTorch code that saves and loads a model checkpoint including the optimizer state. What will be printed after loading?

PyTorch

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

model = nn.Linear(2, 1)
optimizer = optim.SGD(model.parameters(), lr=0.1)

# Simulate one optimizer step
optimizer.zero_grad()
output = model(torch.tensor([[1.0, 2.0]]))
loss = output.sum()
loss.backward()
optimizer.step()

# Save checkpoint
checkpoint = {'model_state': model.state_dict(), 'optimizer_state': optimizer.state_dict()}
torch.save(checkpoint, 'checkpoint.pth')

# Create new model and optimizer
model2 = nn.Linear(2, 1)
optimizer2 = optim.SGD(model2.parameters(), lr=0.1)

# Load checkpoint
loaded = torch.load('checkpoint.pth')
model2.load_state_dict(loaded['model_state'])
optimizer2.load_state_dict(loaded['optimizer_state'])

# Check optimizer state keys
print(sorted(optimizer2.state_dict().keys()))

A ['param_groups', 'state']

B ['model_state', 'optimizer_state']

C ['weights', 'biases']

D ['learning_rate', 'momentum']

💡 Hint

Look at what keys are stored inside optimizer.state_dict() in PyTorch.

The optimizer.state_dict() returns a dictionary with two keys: 'state' which holds the internal state for each parameter (like momentum buffers), and 'param_groups' which holds parameter group info including learning rates.

❓ Model Choice

intermediate

Which optimizer state is necessary to save for resuming training exactly?

When saving a checkpoint to resume training later without losing optimizer progress, which part of the optimizer must be saved?

A Only the model's parameters

B Only the gradients of the model parameters

C Only the optimizer's hyperparameters like learning rate and momentum

D The optimizer's internal state (momentum buffers, etc.) and parameter groups

💡 Hint

Think about what the optimizer uses internally to update parameters beyond just hyperparameters.

To resume training exactly, you must save the optimizer's internal state (like momentum buffers) and parameter groups. This allows the optimizer to continue updating parameters as if training was never interrupted.

❓ Hyperparameter

advanced

What happens if you load an optimizer state with a different learning rate than the current optimizer?

Suppose you saved an optimizer state with learning rate 0.01 but now you create a new optimizer with learning rate 0.001 and load the saved state. What learning rate will the optimizer use after loading?

A It will raise an error due to mismatch

B It will keep 0.001 from the new optimizer ignoring the loaded state

C It will use 0.01 from the loaded state, overriding the new optimizer's setting

D It will average 0.01 and 0.001 and use 0.0055

💡 Hint

Loading optimizer state_dict overwrites all parameter groups including learning rates.

When loading optimizer state_dict, the parameter groups including learning rates are restored from the saved state, overriding any new settings.

🔧 Debug

advanced

Why does this checkpoint loading code cause a runtime error?

Given this code snippet, why does loading the optimizer state cause a runtime error? ```python model = nn.Linear(3, 2) optimizer = optim.Adam(model.parameters(), lr=0.01) checkpoint = torch.load('checkpoint.pth') model.load_state_dict(checkpoint['model_state']) optimizer.load_state_dict(checkpoint['optimizer_state']) ``` Assume the checkpoint was saved from a model with 2 input features instead of 3.

A The model state dict keys do not match, causing optimizer load to fail

B The optimizer state keys do not match because model parameters changed shape

C The checkpoint file is corrupted

D The optimizer type must be the same but it is different

💡 Hint

Check if model parameter shapes match between saved and current model.

If the model parameters have different shapes, the optimizer state dict cannot map saved states to current parameters, causing a runtime error when loading optimizer state.

🧠 Conceptual

expert

Why is saving optimizer state important for training with adaptive optimizers?

Adaptive optimizers like Adam keep internal statistics (e.g., running averages of gradients). Why is saving and restoring the optimizer state critical when resuming training with such optimizers?

A Because internal statistics affect parameter updates and losing them changes training dynamics

B Because optimizer state controls the learning rate scheduler

C Because optimizer state contains the training data used so far

D Because without optimizer state, the model weights cannot be restored

💡 Hint

Think about what adaptive optimizers use internally to adjust updates.

Adaptive optimizers rely on internal running averages of gradients and squared gradients to adjust updates. Losing this state resets these statistics, changing training behavior and possibly harming convergence.