Why Keras simplifies model building in TensorFlow - Why Metrics Matter

Metrics & Evaluation - Why Keras simplifies model building

Which metric matters and WHY

When using Keras to build models, the key metrics depend on your task. For example, in classification, accuracy, precision, recall, and loss are important. Keras simplifies tracking these metrics during training and evaluation, so you can easily see how well your model learns. The main reason these metrics matter is they tell you if your model is improving and if it will work well on new data.

Confusion matrix example

Predicted \ Actual | Positive | Negative
-------------------|----------|---------
Positive           |    80    |   20
Negative           |    10    |   90

Total samples = 80 + 20 + 10 + 90 = 200
Precision = 80 / (80 + 10) = 0.89
Recall = 80 / (80 + 20) = 0.8

Keras helps you compute these metrics easily during training.

Precision vs Recall tradeoff with examples

Keras lets you monitor precision and recall so you can balance them. For example:

• Spam filter: High precision is important to avoid marking good emails as spam.
• Cancer detection: High recall is important to catch all cancer cases, even if some false alarms happen.

Keras makes it easy to add these metrics and adjust your model accordingly.

Good vs Bad metric values

Using Keras, a good model might show:

• Accuracy above 90%
• Precision and recall balanced above 80%
• Loss decreasing steadily during training

A bad model might have:

• Accuracy stuck near 50% (random guessing)
• Precision or recall very low (below 50%)
• Loss not improving or increasing

Keras helps you spot these patterns quickly.

Common pitfalls in metrics

• Accuracy paradox: High accuracy can be misleading if data is imbalanced.
• Data leakage: If test data leaks into training, metrics look falsely good.
• Overfitting: Training metrics improve but test metrics worsen.

Keras tools help detect these issues by showing metrics on training and validation sets.

Self-check question

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

Answer: No, it is not good. The model misses most fraud cases (low recall), which is dangerous. You want high recall to catch fraud, even if accuracy is high.

Key Result

Keras simplifies tracking key metrics like accuracy, precision, and recall, helping you quickly evaluate and improve your model.

Practice

1. Why does Keras simplify building neural networks compared to using raw TensorFlow?

easy

A. Because it requires writing complex low-level code

B. Because it provides a clear, simple way to define layers and train models

C. Because it only works with small datasets

D. Because it does not support training models

Solution

1. Step 1: Understand Keras design goal

   Keras is designed to make neural network building easy by providing simple building blocks like layers.

2. Step 2: Compare with raw TensorFlow

   Raw TensorFlow requires more detailed code for defining models and training, which can be complex for beginners.

3. Final Answer:

   Because it provides a clear, simple way to define layers and train models -> Option B

4. Quick Check:

   Keras simplifies model building = A [OK]

Hint: Keras = simple layers + easy training steps [OK]

Common Mistakes:

• Thinking Keras needs complex code
• Believing Keras only works for small data
• Assuming Keras cannot train models

2. Which of the following is the correct way to start building a model in Keras?

easy

A. model = keras.Sequential()

B. model = keras.compile()

C. model = keras.fit()

D. model = keras.evaluate()

Solution

1. Step 1: Identify model creation method

   In Keras, you create a model by initializing a Sequential or Functional model, commonly with keras.Sequential().

2. Step 2: Understand other methods

   compile(), fit(), and evaluate() are methods called on the model after creation, not for building it.

3. Final Answer:

   model = keras.Sequential() -> Option A

4. Quick Check:

   Start model with Sequential() = B [OK]

Hint: Build model with Sequential(), compile and fit later [OK]

Common Mistakes:

• Using compile() to create model
• Calling fit() before model creation
• Confusing evaluate() with model building

3. What will be the output shape of the model after running this code?

import tensorflow as tf
model = tf.keras.Sequential([
  tf.keras.layers.Dense(10, input_shape=(5,)),
  tf.keras.layers.Dense(3)
])
model.summary()

medium

A. Output shape: (None, 3)

B. Output shape: (None, 10)

C. Output shape: (5, 3)

D. Output shape: (10, 3)

Solution

1. Step 1: Analyze model layers

   The first Dense layer outputs 10 units; the second Dense layer outputs 3 units.

2. Step 2: Determine final output shape

   The model output shape matches the last layer's units, so (None, 3), where None is batch size.

3. Final Answer:

   Output shape: (None, 3) -> Option A

4. Quick Check:

   Last layer units = output shape = A [OK]

Hint: Output shape matches last layer units [OK]

Common Mistakes:

• Confusing input shape with output shape
• Ignoring last layer's units
• Thinking batch size is fixed

4. Identify the error in this Keras model code:

import tensorflow as tf
model = tf.keras.Sequential()
model.add(tf.keras.layers.Dense(8))
model.compile(optimizer='adam', loss='mse')
model.fit(x_train, y_train, epochs=5)

medium

A. fit() missing batch_size argument

B. compile() called before adding layers

C. Missing input shape in first Dense layer

D. Using 'mse' loss is invalid

Solution

1. Step 1: Check model layer definition

   The first Dense layer lacks input_shape, which is required for the first layer in Sequential models.

2. Step 2: Verify other steps

   compile() is correctly called after adding layers; batch_size is optional; 'mse' is a valid loss.

3. Final Answer:

   Missing input shape in first Dense layer -> Option C

4. Quick Check:

   First layer needs input shape = C [OK]

Hint: First layer must have input_shape defined [OK]

Common Mistakes:

• Assuming batch_size is mandatory in fit()
• Thinking compile() order is wrong
• Believing 'mse' is invalid loss

5. You want to build a Keras model that classifies images into 4 categories. Which sequence of steps correctly uses Keras to build, compile, and train this model?

hard

A. Define layers without input shape, fit model, then compile

B. Compile model first, then define layers, then fit with data

C. Fit model first, then define layers, then compile

D. Define layers with input shape, compile with optimizer and loss, then fit with data

Solution

1. Step 1: Build model with layers including input shape

   First, define the model layers specifying input shape so Keras knows input size.

2. Step 2: Compile model with optimizer and loss

   Next, compile the model to set optimizer and loss function before training.

3. Step 3: Train model with fit()

   Finally, train the model using fit() with training data and epochs.

4. Final Answer:

   Define layers with input shape, compile with optimizer and loss, then fit with data -> Option D

5. Quick Check:

   Build -> Compile -> Train = D [OK]

Hint: Build layers, compile, then fit to train [OK]

Common Mistakes:

• Compiling before building layers
• Fitting before compiling
• Skipping input shape in first layer