Prediction and evaluation in TensorFlow - Model Metrics & Evaluation

Metrics & Evaluation - Prediction and evaluation

Which metric matters for Prediction and Evaluation and WHY

When we make predictions with a model, we want to know how well it works. The main metrics to check are accuracy, precision, recall, and F1 score. These tell us if the model guesses right, if it finds all the important cases, and if it avoids false alarms. We choose the metric based on what matters most for the problem.

Confusion Matrix

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

Example:
TP = 50, FP = 10, TN = 30, FN = 10
Total samples = 100
    

Precision vs Recall Tradeoff with Examples

Precision means when the model says "yes", how often is it right? High precision means few false alarms.

Recall means how many actual "yes" cases the model finds. High recall means it misses very few real cases.

Example 1: Spam filter - high precision is important so good emails are not marked as spam.

Example 2: Cancer detection - high recall is important so no cancer cases are missed.

Good vs Bad Metric Values for Prediction and Evaluation

Good values: Accuracy > 90%, Precision and Recall both above 85%, F1 score close to 1.

Bad values: Accuracy around 50% (random guessing), Precision or Recall below 50%, F1 score very low.

Note: High accuracy alone can be misleading if classes are imbalanced.

Common Pitfalls in Metrics

• Accuracy paradox: High accuracy can hide poor performance on rare classes.
• Data leakage: Using future or test data in training inflates metrics falsely.
• Overfitting: Very high training accuracy but low test accuracy means model memorizes data, not learns.

Self Check

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

Answer: No. Even though accuracy is high, the model misses 88% of fraud cases (low recall). This is bad because catching fraud is critical. The model needs improvement to find more fraud cases.

Key Result

Precision, recall, and F1 score are key to understand model prediction quality beyond accuracy.

Practice

1. What does the model.predict() function do in TensorFlow?

easy

A. It saves the model to a file

B. It trains the model on the data

C. It deletes the model from memory

D. It gives the model's guesses on new data

Solution

1. Step 1: Understand the purpose of model.predict()

   This function is used to get the model's output predictions for new input data after training.

2. Step 2: Differentiate from other functions

   Training uses model.fit(), saving uses model.save(), and deleting is manual memory management, none of which are predict().

3. Final Answer:

   It gives the model's guesses on new data -> Option D

4. Quick Check:

   model.predict() = model guesses [OK]

Hint: Predict means guess output for new inputs [OK]

Common Mistakes:

• Confusing predict() with fit() for training
• Thinking predict() saves the model
• Assuming predict() deletes the model

2. Which of the following is the correct way to evaluate a TensorFlow model on test data stored in X_test and y_test?

easy

A. model.score(X_test, y_test)

B. model.evaluate(X_test, y_test)

C. model.fit(X_test, y_test)

D. model.predict(X_test, y_test)

Solution

1. Step 1: Identify the evaluation function

   TensorFlow uses model.evaluate() to measure performance on test data.

2. Step 2: Check other options

   model.predict() makes predictions, model.fit() trains, and model.score() is not a TensorFlow method.

3. Final Answer:

   model.evaluate(X_test, y_test) -> Option B

4. Quick Check:

   Evaluate = measure performance [OK]

Hint: Use evaluate() to check model accuracy on test data [OK]

Common Mistakes:

• Using predict() instead of evaluate() for metrics
• Trying to train with evaluate()
• Using non-existent model.score() method

3. What will be the output of the following code snippet?

import tensorflow as tf
import numpy as np

model = tf.keras.Sequential([
  tf.keras.layers.Dense(1, input_shape=(1,))
])
model.compile(optimizer='sgd', loss='mse')

X = np.array([1, 2, 3, 4], dtype=float)
y = np.array([2, 4, 6, 8], dtype=float)

model.fit(X, y, epochs=10, verbose=0)
predictions = model.predict(np.array([5.0]))
print(predictions)

medium

A. A numpy array close to [[1.0]]

B. A numpy array close to [[5.0]]

C. A numpy array close to [[10.0]]

D. An error because input shape is wrong

Solution

1. Step 1: Understand the model and data

   The model is a simple linear layer trained to learn y = 2*x approximately.

2. Step 2: Predict for input 5.0

   After training, the model should predict close to 2*5 = 10, so output is near [[10.0]].

3. Final Answer:

   A numpy array close to [[10.0]] -> Option C

4. Quick Check:

   Prediction for 5 ≈ 10 [OK]

Hint: Model learns y=2x, predict(5) ≈ 10 [OK]

Common Mistakes:

• Expecting exact 10 instead of approximate
• Confusing input shape causing error
• Thinking prediction returns scalar, not array

4. You run model.evaluate(X_test, y_test) but get a ValueError about mismatched shapes. What is the most likely cause?

medium

A. The shapes of X_test and y_test do not match the model's expected input and output shapes

B. The model was not compiled before evaluation

C. The model.predict() function was called instead of evaluate()

D. The optimizer was set incorrectly

Solution

1. Step 1: Understand the error cause

   A ValueError about shape mismatch usually means input or output data shapes don't match what the model expects.

2. Step 2: Check other options

   Not compiling causes different errors, predict() vs evaluate() is unrelated, and optimizer issues cause training errors, not shape errors.

3. Final Answer:

   The shapes of X_test and y_test do not match the model's expected input and output shapes -> Option A

4. Quick Check:

   Shape mismatch causes ValueError in evaluate() [OK]

Hint: Check input/output shapes match model before evaluate() [OK]

Common Mistakes:

• Ignoring shape mismatch and blaming optimizer
• Confusing predict() with evaluate() errors
• Not compiling model but blaming shape error

5. You trained a model and want to compare its performance on two test sets: X_test1, y_test1 and X_test2, y_test2. Which approach correctly compares their accuracy using TensorFlow?

hard

A. Use model.evaluate() on both test sets separately and compare the returned loss or accuracy values

B. Use model.predict() on both test sets and compare the raw predictions directly

C. Train the model again on X_test2, y_test2 and compare training losses

D. Use model.fit() on both test sets and compare the final epoch losses

Solution

1. Step 1: Understand evaluation for performance

   model.evaluate() returns loss and metrics on test data without training, ideal for comparing performance.

2. Step 2: Why other options are incorrect

   Comparing raw predictions is not a direct accuracy measure; retraining or fitting on test sets changes the model and is not a fair comparison.

3. Final Answer:

   Use model.evaluate() on both test sets separately and compare the returned loss or accuracy values -> Option A

4. Quick Check:

   Evaluate test sets separately for fair comparison [OK]

Hint: Evaluate test sets separately, compare metrics [OK]

Common Mistakes:

• Comparing raw predictions without metrics
• Retraining on test data for comparison
• Using fit() on test data instead of evaluate()