Training and validation loss tracking in PyTorch - Model Metrics & Evaluation

Metrics & Evaluation - Training and validation loss tracking

Which metric matters for training and validation loss tracking and WHY

Loss shows how well the model is learning. Training loss tells us how well the model fits the training data. Validation loss tells us how well the model works on new, unseen data. We want both losses to go down. If training loss goes down but validation loss goes up, the model is overfitting. So, tracking both helps us know if the model is learning well and generalizing.

Confusion matrix or equivalent visualization

For loss tracking, we use line plots instead of confusion matrices. Here is an example of loss values over epochs:

Epoch | Training Loss | Validation Loss
---------------------------------------
  1   |     0.85     |      0.90
  2   |     0.60     |      0.65
  3   |     0.45     |      0.50
  4   |     0.30     |      0.55
  5   |     0.20     |      0.70

This shows training loss steadily decreasing, but validation loss starts increasing after epoch 3, indicating overfitting.

Precision vs Recall tradeoff analogy for loss tracking

Think of training loss as how well you practice at home, and validation loss as how well you perform in a real game. If you only focus on practice (training loss), you might get very good there but fail in the game (validation loss). Balancing both is like balancing practice and real performance. Lower training loss with stable or lower validation loss means good learning. If validation loss rises, the model is memorizing practice but not learning to play well in real games.

What good vs bad loss values look like

Good: Both training and validation loss decrease and stay close. For example, training loss 0.2 and validation loss 0.25 after many epochs.

Bad: Training loss very low (e.g., 0.1) but validation loss high or increasing (e.g., 0.7). This means overfitting.

Also, if both losses stay high and do not decrease, the model is underfitting and not learning well.

Common pitfalls in loss tracking

• Ignoring validation loss: Only watching training loss can hide overfitting.
• Data leakage: If validation data leaks into training, validation loss looks unrealistically low.
• Overfitting signs: Training loss keeps dropping but validation loss rises.
• Underfitting signs: Both losses stay high and flat.
• Not enough epochs: Loss may not have stabilized yet, so early stopping too soon can mislead.

Self-check question

Your model has training loss 0.1 but validation loss 0.6 after many epochs. Is it good for production? Why or why not?

Answer: No, this is not good. The low training loss means the model fits training data well, but the high validation loss means it does not generalize to new data. This is overfitting. The model may perform poorly on real-world data.

Key Result

Tracking both training and validation loss helps detect overfitting and ensures the model learns well and generalizes.

Practice

1. Why do we track both training and validation loss during model training in PyTorch?

easy

A. To speed up the training process

B. To increase the batch size automatically

C. To monitor if the model is learning and generalizing well to new data

D. To reduce the number of model parameters

Solution

1. Step 1: Understand the purpose of training loss

   Training loss shows how well the model fits the training data during learning.

2. Step 2: Understand the purpose of validation loss

   Validation loss shows how well the model performs on unseen data, indicating generalization.

3. Final Answer:

   To monitor if the model is learning and generalizing well to new data -> Option C

4. Quick Check:

   Training and validation loss track learning and generalization [OK]

Hint: Training loss = fit, validation loss = generalization [OK]

Common Mistakes:

• Thinking validation loss speeds up training
• Confusing loss tracking with batch size changes
• Assuming loss tracking reduces parameters

2. Which PyTorch method should you call before starting the validation phase to ensure correct behavior?

easy

A. model.train()

B. model.eval()

C. model.validate()

D. model.freeze()

Solution

1. Step 1: Recall PyTorch modes

   model.train() sets the model to training mode, enabling dropout and batch norm updates.

2. Step 2: Identify validation mode

   model.eval() sets the model to evaluation mode, disabling dropout and batch norm updates for stable validation.

3. Final Answer:

   model.eval() -> Option B

4. Quick Check:

   Use model.eval() before validation [OK]

Hint: Use model.eval() to disable training-only layers [OK]

Common Mistakes:

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

3. Given the code below, what will be the printed average training loss after one epoch?

train_losses = []
for inputs, targets in train_loader:
    optimizer.zero_grad()
    outputs = model(inputs)
    loss = criterion(outputs, targets)
    loss.backward()
    optimizer.step()
    train_losses.append(loss.item())
avg_train_loss = sum(train_losses) / len(train_losses)
print(f"Average training loss: {avg_train_loss:.4f}")

Assuming losses per batch are [0.8, 0.6, 0.4], what is the output?

medium

A. Average training loss: 0.4000

B. Average training loss: 1.8000

C. Average training loss: 0.0000

D. Average training loss: 0.6000

Solution

1. Step 1: Sum the batch losses

   Sum = 0.8 + 0.6 + 0.4 = 1.8

2. Step 2: Calculate average loss

   Average = 1.8 / 3 = 0.6

3. Final Answer:

   Average training loss: 0.6000 -> Option D

4. Quick Check:

   Sum 1.8 / 3 batches = 0.6 [OK]

Hint: Average = sum of losses ÷ number of batches [OK]

Common Mistakes:

• Printing sum instead of average
• Dividing by wrong number of batches
• Using loss tensor instead of loss.item()

4. What is the bug in this code snippet for calculating average validation loss?

model.eval()
val_losses = []
for inputs, targets in val_loader:
    outputs = model(inputs)
    loss = criterion(outputs, targets)
    val_losses.append(loss)
avg_val_loss = sum(val_losses) / len(val_losses)
print(f"Validation loss: {avg_val_loss:.4f}")

medium

A. Appending loss tensor instead of loss.item()

B. Not calling model.train() before validation

C. Using model.eval() disables loss calculation

D. Forgetting to zero gradients before validation

Solution

1. Step 1: Check what is appended to val_losses

   The code appends loss tensors, not scalar values.

2. Step 2: Understand sum behavior on tensors list

   Summing loss tensors retains computation graphs from each forward pass, preventing memory release and causing high memory usage over multiple epochs.

3. Final Answer:

   Appending loss tensor instead of loss.item() -> Option A

4. Quick Check:

   Use loss.item() to get scalar loss [OK]

Hint: Append loss.item(), not loss tensor, for correct averaging [OK]

Common Mistakes:

• Appending loss tensor directly
• Confusing model.train() and model.eval() roles
• Expecting loss calculation to stop in eval mode

5. You want to track training and validation loss per epoch and plot them to check overfitting. Which approach correctly implements this in PyTorch?

hard

A. Use separate lists for training and validation losses, append average loss each epoch, then plot both lists

B. Append batch losses directly to one list for both training and validation, then plot

C. Only track training loss since validation loss is not useful

D. Call model.train() during validation to get accurate loss

Solution

1. Step 1: Understand loss tracking per epoch

   We need to store average losses per epoch separately for training and validation.

2. Step 2: Use separate lists and append averages

   Appending average loss per epoch to separate lists allows clear plotting and comparison.

3. Final Answer:

   Use separate lists for training and validation losses, append average loss each epoch, then plot both lists -> Option A

4. Quick Check:

   Track average losses per epoch separately [OK]

Hint: Track average losses per epoch separately for clear plots [OK]

Common Mistakes:

• Mixing batch losses in one list
• Ignoring validation loss tracking
• Using model.train() during validation