Validation loop in PyTorch - Model Metrics & Evaluation

Metrics & Evaluation - Validation loop

Which metric matters for Validation Loop and WHY

The validation loop helps us check how well our model works on new data it hasn't seen before. The key metrics here are validation loss and validation accuracy. Validation loss tells us how far off the model's predictions are from the true answers on the validation set. Validation accuracy shows the percentage of correct predictions. These metrics matter because they help us see if the model is learning well or just memorizing training data.

Confusion Matrix Example

      | Predicted Positive | Predicted Negative |
      |--------------------|--------------------|
      | True Positive (TP)  | False Negative (FN) |
      | False Positive (FP) | True Negative (TN)  |

      Example:
      TP = 40, FP = 10
      FN = 5,  TN = 45
      Total samples = 100
    

This matrix helps calculate precision, recall, and accuracy during validation.

Precision vs Recall Tradeoff in Validation

During validation, precision and recall help us understand different errors:

• Precision = TP / (TP + FP): How many predicted positives are actually correct?
• Recall = TP / (TP + FN): How many actual positives did the model find?

For example, in a spam filter, high precision means fewer good emails marked as spam (important to avoid annoying users). High recall means catching most spam emails. Depending on the task, we choose which metric to prioritize during validation.

Good vs Bad Metric Values in Validation Loop

Good validation metrics:

• Validation loss steadily decreases or stays low.
• Validation accuracy is close to training accuracy (no big gap).
• Precision and recall are balanced and high for the task.

Bad validation metrics:

• Validation loss stops improving or increases (sign of overfitting).
• Validation accuracy much lower than training accuracy.
• Very low precision or recall indicating poor predictions.

Common Pitfalls in Validation Metrics

• Accuracy paradox: High accuracy can be misleading if classes are imbalanced.
• Data leakage: Validation data accidentally used in training leads to overly optimistic metrics.
• Overfitting indicators: Validation loss increasing while training loss decreases.
• Ignoring metric choice: Using accuracy alone when precision or recall matters more.

Self Check

Your model has 98% accuracy but only 12% recall on fraud cases during validation. Is it good for production? Why or why not?

Answer: No, it is not good. The low recall means the model misses most fraud cases, which is dangerous. Even with high accuracy, the model fails to catch fraud effectively. For fraud detection, high recall is critical to catch as many frauds as possible.

Key Result

Validation loss and accuracy show if the model generalizes well; precision and recall guide error tradeoffs.

Practice

1. What is the main purpose of a validation loop in PyTorch?

easy

A. To update the model weights during training

B. To check how well the model performs on unseen data

C. To load the training dataset

D. To save the model to disk

Solution

1. Step 1: Understand the role of validation

   The validation loop is used to test the model on new data that it has not seen during training.

2. Step 2: Differentiate from training

   Unlike training, validation does not update model weights but measures performance like accuracy or loss.

3. Final Answer:

   To check how well the model performs on unseen data -> Option B

4. Quick Check:

   Validation loop = test model on new data [OK]

Hint: Validation loop tests model without changing weights [OK]

Common Mistakes:

• Confusing validation with training
• Thinking validation updates model weights
• Mixing validation with data loading

2. Which PyTorch code snippet correctly sets the model for validation?

easy

A. model.eval()

B. model.train()

C. model.validate()

D. model.test()

Solution

1. Step 1: Recall PyTorch model modes

   In PyTorch, model.eval() sets the model to evaluation mode, disabling dropout and batch norm updates.

2. Step 2: Identify correct method for validation

   Validation requires evaluation mode, so model.eval() is correct; others do not exist or are for training.

3. Final Answer:

   model.eval() -> Option A

4. Quick Check:

   Use model.eval() for validation [OK]

Hint: Use model.eval() to disable training behaviors [OK]

Common Mistakes:

• Using model.train() during validation
• Assuming model.validate() exists
• Confusing test() with eval()

3. What will be the output of the following validation loop snippet?

model.eval()
correct = 0
with torch.no_grad():
    for inputs, labels in val_loader:
        outputs = model(inputs)
        predicted = outputs.argmax(dim=1)
        correct += (predicted == labels).sum().item()
accuracy = correct / len(val_loader.dataset)
print(f"Accuracy: {accuracy:.2f}")

medium

A. Prints loss value instead of accuracy

B. Raises an error because of missing optimizer

C. Prints the accuracy of the model on validation data

D. Prints training accuracy instead of validation accuracy

Solution

1. Step 1: Analyze the code logic

   The code sets model to eval mode, disables gradient calculation, and counts correct predictions over validation data.

2. Step 2: Understand the output calculation

   Accuracy is computed as correct predictions divided by total validation samples, then printed.

3. Final Answer:

   Prints the accuracy of the model on validation data -> Option C

4. Quick Check:

   Validation loop code prints accuracy [OK]

Hint: Check if code counts correct predictions for accuracy [OK]

Common Mistakes:

• Expecting loss instead of accuracy
• Thinking optimizer is needed in validation
• Confusing training and validation accuracy

4. Identify the error in this validation loop code:

model.eval()
for inputs, labels in val_loader:
    outputs = model(inputs)
    loss = criterion(outputs, labels)
    loss.backward()
    total_loss += loss.item()

medium

A. Not zeroing gradients before backward

B. Not using model.train() before loop

C. Missing optimizer.step() after loss.backward()

D. Calling loss.backward() during validation

Solution

1. Step 1: Understand validation loop purpose

   Validation should not update model weights or compute gradients, so backward() is not needed.

2. Step 2: Identify the error

   Calling loss.backward() computes gradients, which is incorrect in validation; it wastes resources and may cause errors.

3. Final Answer:

   Calling loss.backward() during validation -> Option D

4. Quick Check:

   No backward() in validation loop [OK]

Hint: No backward() in validation, only in training [OK]

Common Mistakes:

• Running backward() in validation
• Confusing train and eval modes
• Forgetting torch.no_grad() in validation

5. You want to add accuracy tracking in your validation loop. Which code snippet correctly computes and updates accuracy inside the loop?

model.eval()
total_correct = 0
total_samples = 0
with torch.no_grad():
    for inputs, labels in val_loader:
        outputs = model(inputs)
        predicted = outputs.argmax(dim=1)
        # What goes here?
accuracy = total_correct / total_samples
print(f"Validation Accuracy: {accuracy:.2f}")

hard

A. total_correct += (predicted == labels).sum().item()
total_samples += labels.size(0)

B. total_correct += predicted.sum().item()
total_samples += labels.size(0)

C. total_correct += (predicted != labels).sum().item()
total_samples += labels.size(0)

D. total_correct += (predicted == labels).sum()
total_samples += len(val_loader)

Solution

1. Step 1: Calculate correct predictions

   Use (predicted == labels).sum().item() to count correct predictions as a Python number.

2. Step 2: Update total samples count

   Use labels.size(0) to add the batch size to total samples for accurate accuracy calculation.

3. Final Answer:

   total_correct += (predicted == labels).sum().item() total_samples += labels.size(0) -> Option A

4. Quick Check:

   Sum correct matches and batch size for accuracy [OK]

Hint: Sum correct predictions and batch size for accuracy [OK]

Common Mistakes:

• Summing predicted values instead of correct matches
• Using != instead of == for correct count
• Adding batch count incorrectly as len(val_loader)