Why checkpointing preserves progress in PyTorch - Why Metrics Matter

Metrics & Evaluation - Why checkpointing preserves progress

Which metric matters for this concept and WHY

Checkpointing itself is about saving the model's state during training. The key metric to watch is training loss or validation loss over time. This shows if the model is improving. Checkpointing preserves progress by saving these states, so if training stops, you can restart without losing improvements.

Confusion matrix or equivalent visualization (ASCII)

Checkpointing does not directly involve confusion matrices.
Instead, think of it as saving snapshots of training:

Training steps: 1 2 3 4 5 6 7 8 9 10
Loss:          0.9 0.8 0.7 0.6 0.5 0.4 0.35 0.3 0.28 0.25

Checkpoint saved at step 5 (loss 0.5)
If training stops at step 7, you can reload checkpoint from step 5
and continue training from there, not from step 1.
    

Precision vs Recall (or equivalent tradeoff) with concrete examples

Checkpointing trades off time saved vs storage used. Saving checkpoints often means more storage but less lost work if interrupted. Saving less often saves space but risks losing more progress.

Example: If you save checkpoints every 10 minutes, you lose at most 10 minutes of work on failure. If you save every hour, you risk losing up to an hour of training.

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

Good checkpointing means:

• Checkpoints saved frequently enough to avoid losing much progress.
• Checkpoints correctly restore model and optimizer states.
• Training loss continues to decrease after resuming from checkpoint.

Bad checkpointing means:

• Checkpoints saved too rarely, causing large loss of training time on failure.
• Checkpoints missing optimizer state, causing training to restart badly.
• Loss jumps or training stalls after resuming, indicating corrupted or incomplete checkpoint.

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

Common pitfalls with checkpointing include:

• Not saving optimizer state: causes learning rate and momentum to reset, hurting training.
• Overwriting checkpoints without backups: losing all progress if checkpoint is corrupted.
• Confusing checkpoint saving with model evaluation metrics: checkpointing only saves state, it does not improve metrics by itself.
• Not verifying checkpoint integrity before resuming: can cause silent errors.

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

No, it is not good for fraud detection. The low recall (12%) means the model misses most fraud cases, which is dangerous. Checkpointing helps preserve training progress but does not fix poor model performance. You need to improve the model or data, not just rely on checkpointing.

Key Result

Checkpointing preserves training progress by saving model and optimizer states, allowing training to resume without losing improvements.

Practice

1. What is the main reason for using checkpointing during PyTorch model training?

easy

A. To save the model's current state so training can resume later without loss

B. To speed up the training by skipping some layers

C. To reduce the size of the training dataset

D. To automatically tune hyperparameters during training

Solution

1. Step 1: Understand checkpointing purpose

   Checkpointing saves the model's current state including weights and optimizer info.

2. Step 2: Connect checkpointing to training progress

   This allows training to stop and resume later without losing progress.

3. Final Answer:

   To save the model's current state so training can resume later without loss -> Option A

4. Quick Check:

   Checkpointing = Save progress [OK]

Hint: Checkpointing means saving progress to continue later [OK]

Common Mistakes:

• Thinking checkpointing speeds up training
• Confusing checkpointing with data reduction
• Assuming checkpointing tunes hyperparameters

2. Which of the following is the correct PyTorch code snippet to save a checkpoint?

easy

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

B. torch.save(model.state_dict(), 'checkpoint.pth')

C. torch.load('checkpoint.pth')

D. optimizer.save('checkpoint.pth')

Solution

1. Step 1: Identify saving function

   torch.save() is used to save objects like model weights to a file.

2. Step 2: Check correct usage for saving model state

   model.state_dict() returns model weights; saving it with torch.save() is correct.

3. Final Answer:

   torch.save(model.state_dict(), 'checkpoint.pth') -> Option B

4. Quick Check:

   Save model weights = torch.save(state_dict) [OK]

Hint: Use torch.save with model.state_dict() to save checkpoint [OK]

Common Mistakes:

• Using torch.load instead of torch.save to save
• Trying to save optimizer with wrong method
• Confusing load_state_dict with saving

3. Given this code snippet, what will be printed after loading the checkpoint?

model = MyModel()
optimizer = torch.optim.Adam(model.parameters())
checkpoint = torch.load('checkpoint.pth')
model.load_state_dict(checkpoint['model_state'])
optimizer.load_state_dict(checkpoint['optimizer_state'])
epoch = checkpoint['epoch']
print(epoch)

medium

A. An error because checkpoint keys are missing

B. The total number of model parameters

C. The optimizer learning rate

D. The epoch number saved in the checkpoint

Solution

1. Step 1: Understand checkpoint contents

   The checkpoint dictionary contains keys 'model_state', 'optimizer_state', and 'epoch'.

2. Step 2: Identify printed value

   Variable 'epoch' is assigned checkpoint['epoch'], so print(epoch) outputs the saved epoch number.

3. Final Answer:

   The epoch number saved in the checkpoint -> Option D

4. Quick Check:

   Print epoch from checkpoint = epoch number [OK]

Hint: Print shows saved epoch from checkpoint dictionary [OK]

Common Mistakes:

• Thinking print shows model parameters count
• Confusing optimizer state with epoch
• Assuming missing keys cause error here

4. You tried to resume training but got an error: RuntimeError: Error(s) in loading state_dict. What is the most likely cause related to checkpointing?

medium

A. The training data was modified after checkpointing

B. The checkpoint file was saved with torch.load instead of torch.save

C. The model architecture changed after saving the checkpoint

D. The optimizer state was not saved in the checkpoint

Solution

1. Step 1: Understand error meaning

   Loading state_dict errors usually happen if model layers differ from saved checkpoint.

2. Step 2: Connect error to checkpoint cause

   If model architecture changed after saving, weights won't match, causing this error.

3. Final Answer:

   The model architecture changed after saving the checkpoint -> Option C

4. Quick Check:

   State_dict error = architecture mismatch [OK]

Hint: Mismatch model layers cause state_dict loading errors [OK]

Common Mistakes:

• Confusing save/load functions causing error
• Assuming missing optimizer state causes this error
• Blaming training data changes for state_dict error

5. You want to checkpoint your training every 5 epochs to avoid losing progress. Which approach best preserves training progress including optimizer state and epoch count?

hard

A. Save a dictionary with model.state_dict(), optimizer.state_dict(), and current epoch number

B. Save only model.state_dict() every 5 epochs

C. Save optimizer.state_dict() and epoch number but not model weights

D. Save the training data batch every 5 epochs

Solution

1. Step 1: Identify what preserves full training state

   Saving model weights, optimizer state, and epoch number allows full resume.

2. Step 2: Compare options

   Only saving model weights misses optimizer info; saving optimizer and epoch without model is incomplete; saving data batch doesn't preserve progress.

3. Final Answer:

   Save a dictionary with model.state_dict(), optimizer.state_dict(), and current epoch number -> Option A

4. Quick Check:

   Checkpoint = model + optimizer + epoch [OK]

Hint: Checkpoint all: model, optimizer, and epoch for full resume [OK]

Common Mistakes:

• Saving only model weights loses optimizer progress
• Ignoring epoch number causes restart from zero
• Saving training data batch does not preserve model state