Checkpoint with optimizer state in PyTorch - Model Metrics & Evaluation

Metrics & Evaluation - Checkpoint with optimizer state

Which metric matters for this concept and WHY

When saving a checkpoint with optimizer state, the key metric to watch is training loss. This is because the optimizer state helps the model continue learning smoothly from where it left off. If the loss after loading the checkpoint is similar to before saving, it means the optimizer state was restored correctly and training can continue effectively.

Confusion matrix or equivalent visualization (ASCII)

Checkpointing itself does not have a confusion matrix. Instead, we check if the model parameters and optimizer state are restored correctly by comparing training metrics before and after loading.

    Before saving checkpoint:
    Epoch 5 - Loss: 0.45

    After loading checkpoint:
    Epoch 5 - Loss: 0.45
    

If loss values match closely, it means the checkpoint with optimizer state was saved and loaded properly.

Precision vs Recall (or equivalent tradeoff) with concrete examples

Checkpointing with optimizer state is about continuity in training, not classification metrics like precision or recall. The tradeoff here is between saving frequently to avoid losing progress and saving too often which can slow training.

• Saving too rarely risks losing many training steps if interrupted.
• Saving too often wastes time and storage.

Good practice is to save checkpoints at meaningful intervals, including optimizer state, so training can resume exactly where it stopped.

What "good" vs "bad" metric values look like for this use case

Good checkpointing means:

• After loading, training loss continues smoothly without jumps.
• Optimizer state is restored, so learning rate schedules and momentum continue correctly.
• Model accuracy or other metrics improve as expected after resuming.

Bad checkpointing means:

• Loss suddenly increases or training stalls after loading.
• Optimizer state missing causes learning rate resets or momentum loss.
• Training metrics degrade or behave erratically after resuming.

Metrics pitfalls (accuracy paradox, data leakage, overfitting indicators)

Common pitfalls when checkpointing with optimizer state include:

• Not saving optimizer state: Leads to loss of momentum and learning rate info, causing slower or unstable training after resume.
• Partial checkpointing: Saving only model weights but not optimizer state can cause unexpected metric jumps.
• Data leakage: If checkpointing includes validation data accidentally, metrics may be misleading.
• Overfitting: Resuming training without proper early stopping can cause overfitting, seen in metrics worsening.

Your model has 98% accuracy but 12% recall on fraud. Is it good?

No, this is not good for fraud detection. Even if accuracy is high, a recall of 12% means the model misses 88% of fraud cases. For fraud detection, high recall is critical to catch as many frauds as possible. This shows why looking at multiple metrics is important.

Key Result

Saving and restoring optimizer state ensures training loss continues smoothly after resuming, confirming checkpoint correctness.

Practice

1. What is the main reason to save the optimizer state along with the model in a PyTorch checkpoint?

easy

A. To speed up the model's inference time

B. To reduce the model size on disk

C. To resume training with the same learning rate and momentum settings

D. To convert the model to a different format

Solution

1. Step 1: Understand what optimizer state contains

   The optimizer state includes parameters like learning rate, momentum, and other variables that affect training progress.

2. Step 2: Reason why saving optimizer state is important

   Saving the optimizer state allows training to resume exactly where it left off, preserving these settings.

3. Final Answer:

   To resume training with the same learning rate and momentum settings -> Option C

4. Quick Check:

   Optimizer state saves training settings = C [OK]

Hint: Optimizer state saves training progress settings [OK]

Common Mistakes:

• Thinking optimizer state reduces model size
• Confusing optimizer state with model weights
• Believing optimizer state affects inference speed

2. Which of the following is the correct way to save a checkpoint including model and optimizer states in PyTorch?

easy

A. torch.save(model, 'checkpoint.pth')

B. torch.save({'model': model.state_dict(), 'optimizer': optimizer.state_dict()}, 'checkpoint.pth')

C. torch.save(optimizer, 'checkpoint.pth')

D. torch.save({'model': model, 'optimizer': optimizer}, 'checkpoint.pth')

Solution

1. Step 1: Identify correct saving method for states

   PyTorch recommends saving state_dict() of model and optimizer for checkpoints.

2. Step 2: Check each option

   torch.save({'model': model.state_dict(), 'optimizer': optimizer.state_dict()}, 'checkpoint.pth') saves state_dict() of both model and optimizer in a dictionary, which is correct.

3. Final Answer:

   torch.save({'model': model.state_dict(), 'optimizer': optimizer.state_dict()}, 'checkpoint.pth') -> Option B

4. Quick Check:

   Save state_dict() for model and optimizer = B [OK]

Hint: Save state_dict() of model and optimizer in dict [OK]

Common Mistakes:

• Saving full model object instead of state_dict
• Saving optimizer object directly
• Not saving optimizer state at all

3. Given this code snippet, what will be printed?

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)

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

# Load checkpoint
loaded = torch.load('cp.pth')
optimizer.load_state_dict(loaded['optimizer'])
print(optimizer.param_groups[0]['lr'])

medium

A. 0.1

B. 0.01

C. 1.0

D. Error: optimizer state not loaded

Solution

1. Step 1: Understand optimizer initialization

   Optimizer is created with learning rate 0.1 and saved in checkpoint.

2. Step 2: Loading optimizer state restores learning rate

   Loading optimizer state_dict sets learning rate back to 0.1.

3. Final Answer:

   0.1 -> Option A

4. Quick Check:

   Loaded optimizer lr = 0.1 [OK]

Hint: Loaded optimizer keeps saved learning rate [OK]

Common Mistakes:

• Assuming learning rate resets to default
• Forgetting to load optimizer state
• Confusing model and optimizer states

4. You saved a checkpoint with model and optimizer states but when loading, training behaves as if optimizer settings are lost. What is the most likely mistake?

medium

A. Not calling optimizer.load_state_dict() after loading checkpoint

B. Saving model.state_dict() instead of model

C. Using torch.save() instead of torch.load()

D. Not setting model.eval() before saving

Solution

1. Step 1: Identify cause of lost optimizer settings

   If optimizer state is not loaded, training uses default optimizer settings.

2. Step 2: Check common mistakes

   Not calling optimizer.load_state_dict() after loading checkpoint causes this issue.

3. Final Answer:

   Not calling optimizer.load_state_dict() after loading checkpoint -> Option A

4. Quick Check:

   Load optimizer state to keep settings = D [OK]

Hint: Always load optimizer state after loading checkpoint [OK]

Common Mistakes:

• Saving full model instead of state_dict
• Confusing torch.save and torch.load usage
• Setting model.eval() affects inference, not optimizer

5. You want to save a checkpoint that allows resuming training exactly, including epoch number and best loss so far. Which is the best way to structure the checkpoint dictionary?

hard

A. {'epoch': epoch, 'model': model.state_dict()}

B. {'model': model, 'optimizer': optimizer, 'epoch': epoch}

C. {'model_state': model.state_dict(), 'loss': best_loss}

D. {'epoch': epoch, 'model': model.state_dict(), 'optimizer': optimizer.state_dict(), 'best_loss': best_loss}

Solution

1. Step 1: Identify required checkpoint components

   To resume training exactly, save epoch, model state, optimizer state, and best loss.

2. Step 2: Evaluate options

   {'epoch': epoch, 'model': model.state_dict(), 'optimizer': optimizer.state_dict(), 'best_loss': best_loss} includes all required keys with correct state_dict() usage.

3. Final Answer:

   {'epoch': epoch, 'model': model.state_dict(), 'optimizer': optimizer.state_dict(), 'best_loss': best_loss} -> Option D

4. Quick Check:

   Save epoch, model, optimizer, loss in checkpoint = A [OK]

Hint: Include epoch, model, optimizer, and loss in checkpoint dict [OK]

Common Mistakes:

• Saving full model or optimizer objects
• Omitting optimizer state
• Not saving epoch or loss for training resume