Training history and visualization in TensorFlow - Model Metrics & Evaluation

Metrics & Evaluation - Training history and visualization

Which metric matters for Training History and Visualization and WHY

When training a model, we watch metrics like loss and accuracy over time (epochs). Loss shows how well the model fits the data; lower is better. Accuracy shows how many predictions are correct; higher is better. Visualizing these helps us see if the model is learning or stuck.

Confusion Matrix or Equivalent Visualization

Training history visualization usually shows line charts of loss and accuracy for both training and validation sets over epochs.

Epoch | Train Loss | Val Loss | Train Acc | Val Acc
-----------------------------------------------
1     | 0.65       | 0.70     | 0.60      | 0.58
2     | 0.50       | 0.55     | 0.75      | 0.70
3     | 0.40       | 0.45     | 0.82      | 0.78
...   | ...        | ...      | ...       | ...
    

This table is often shown as a line graph with epochs on the x-axis and metric values on the y-axis.

Precision vs Recall Tradeoff (Related to Training History)

While training history focuses on loss and accuracy, precision and recall are also important metrics to track, especially for imbalanced data. Sometimes improving precision lowers recall and vice versa. Watching training history helps us decide if the model is improving overall or just memorizing.

For example, if validation loss stops improving but training loss keeps dropping, the model might be overfitting, hurting recall or precision on new data.

What "Good" vs "Bad" Training History Looks Like

Good: Training and validation loss both decrease smoothly and stabilize close together. Accuracy improves steadily on both sets.

Bad: Training loss keeps dropping but validation loss rises (overfitting). Accuracy on validation stays low or fluctuates wildly (underfitting or data issues).

Common Pitfalls in Training History and Visualization

• Ignoring validation metrics: Only watching training loss can hide overfitting.
• Misinterpreting fluctuations: Small ups and downs are normal; don't panic early.
• Not using early stopping: Without it, model may overfit after many epochs.
• Data leakage: If validation data leaks into training, metrics look too good but model fails in real use.

Self Check: Your model has 98% accuracy but validation loss is rising after epoch 10. Is it good?

No, this suggests overfitting. The model fits training data well (high accuracy) but performs worse on new data (rising validation loss). You should stop training earlier or use regularization.

Key Result

Training history visualization helps detect if the model is learning well or overfitting by comparing training and validation loss and accuracy over epochs.

Practice

1. What does the history.history object store after training a TensorFlow model?

easy

A. The dataset used for training

B. The model's architecture details

C. Loss and accuracy values for each epoch during training

D. The optimizer's internal state

Solution

1. Step 1: Understand what history.history contains

   After training, TensorFlow's model.fit() returns a history object that stores metrics like loss and accuracy for each epoch.

2. Step 2: Identify the correct stored data

   The history.history dictionary holds lists of loss and accuracy values recorded at each epoch for training and validation.

3. Final Answer:

   Loss and accuracy values for each epoch during training -> Option C

4. Quick Check:

   Training metrics stored in history.history = Loss and accuracy values for each epoch during training [OK]

Hint: Remember: history stores metrics per epoch, not model or data [OK]

Common Mistakes:

• Confusing history with model architecture
• Thinking history stores the dataset
• Assuming history holds optimizer state

2. Which of the following is the correct way to plot training and validation accuracy from a TensorFlow history object using matplotlib?

easy

A. plt.plot(history.history['accuracy']); plt.plot(history.history['val_accuracy'])

B. plt.plot(history['accuracy']); plt.plot(history['val_accuracy'])

C. plt.plot(history.accuracy); plt.plot(history.val_accuracy)

D. plt.plot(history.accuracy()); plt.plot(history.val_accuracy())

Solution

1. Step 1: Recall how to access metrics in history object

   The history object stores metrics in a dictionary under history.history. Access keys like 'accuracy' and 'val_accuracy' as dictionary keys.

2. Step 2: Use matplotlib to plot lists from the dictionary

   Use plt.plot() with history.history['accuracy'] and history.history['val_accuracy'] to plot training and validation accuracy.

3. Final Answer:

   plt.plot(history.history['accuracy']); plt.plot(history.history['val_accuracy']) -> Option A

4. Quick Check:

   Access metrics via history.history['key'] for plotting [OK]

Hint: Always access metrics with history.history['metric_name'] [OK]

Common Mistakes:

• Using dot notation instead of dictionary keys
• Calling metrics as functions
• Accessing history directly without .history

3. Given the following code snippet, what will be the output of print(history.history['loss']) after training for 3 epochs?

model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])
history = model.fit(x_train, y_train, epochs=3, validation_data=(x_val, y_val))
print(history.history['loss'])

medium

A. A list of 3 loss values, one per epoch

B. An error because 'loss' key does not exist

C. A single float value of final loss

D. [0.8, 0.6, 0.4]

Solution

1. Step 1: Understand what history.history['loss'] contains

   It stores the loss values recorded at the end of each epoch during training as a list.

2. Step 2: Predict the output after 3 epochs

   Since training runs for 3 epochs, the list will have 3 float values representing loss per epoch, not just one or a fixed list.

3. Final Answer:

   A list of 3 loss values, one per epoch -> Option A

4. Quick Check:

   Loss per epoch stored as list = A list of 3 loss values, one per epoch [OK]

Hint: Loss history is a list with one value per epoch [OK]

Common Mistakes:

• Expecting a single float instead of a list
• Assuming fixed loss values without training
• Thinking 'loss' key is missing

4. Identify the error in this code snippet that tries to plot training and validation loss:

import matplotlib.pyplot as plt
plt.plot(history['loss'])
plt.plot(history['val_loss'])
plt.show()

medium

A. plt.plot() cannot plot lists

B. history should be accessed as history.history, not directly

C. Missing plt.title() causes error

D. No error, code runs fine

Solution

1. Step 1: Check how history metrics are accessed

   The history object stores metrics inside the history attribute, so direct access like history['loss'] is incorrect.

2. Step 2: Correct the access to history.history['loss']

   To fix, use history.history['loss'] and history.history['val_loss'] for plotting.

3. Final Answer:

   history should be accessed as history.history, not directly -> Option B

4. Quick Check:

   Access metrics via history.history, not history [OK]

Hint: Use history.history to access metrics, not history alone [OK]

Common Mistakes:

• Accessing history metrics directly
• Assuming plt.plot can't plot lists
• Thinking missing title causes error

5. You trained a model for 10 epochs but notice the validation loss increases after epoch 5 while training loss decreases. How can visualizing the training history help you decide the next step?

hard

A. It suggests increasing the learning rate to fix validation loss

B. It confirms the model is perfect, so no changes needed

C. It means the training data is incorrect and should be discarded

D. It shows overfitting, so you might stop training early or add regularization

Solution

1. Step 1: Interpret the training and validation loss curves

   When training loss decreases but validation loss increases, it indicates the model is overfitting the training data.

2. Step 2: Decide actions based on visualization

   Visualizing history helps identify overfitting, suggesting to stop early, add dropout, or use regularization to improve generalization.

3. Final Answer:

   It shows overfitting, so you might stop training early or add regularization -> Option D

4. Quick Check:

   Increasing validation loss with decreasing training loss = overfitting [OK]

Hint: Watch for validation loss rising while training loss falls [OK]

Common Mistakes:

• Ignoring validation loss trends
• Increasing learning rate without reason
• Assuming data is wrong without checking