Classification reports in TensorFlow - ML Experiment: Train & Evaluate

Experiment - Classification reports

Problem:You have trained a neural network to classify images into 3 categories. The model shows good accuracy but you want to understand how well it performs on each class.

Current Metrics:Overall accuracy: 85%, no detailed class-wise metrics available.

Issue:The model's overall accuracy is known but you lack detailed insights like precision, recall, and F1-score per class to evaluate performance properly.

Your Task

Generate and interpret a classification report showing precision, recall, and F1-score for each class using TensorFlow and sklearn.

Use TensorFlow for model prediction.

Use sklearn's classification_report function for metrics.

Do not retrain the model; use existing predictions.

Hint 1

Hint 2

Hint 3

Solution

TensorFlow

import numpy as np
from sklearn.metrics import classification_report
import tensorflow as tf

# Simulate loading test data and true labels
num_samples = 100
num_classes = 3
np.random.seed(42)
X_test = np.random.random((num_samples, 20))  # example test features
true_labels = np.random.randint(0, num_classes, size=num_samples)

# Simulate a trained model
model = tf.keras.Sequential([
    tf.keras.layers.Dense(32, activation='relu', input_shape=(20,)),
    tf.keras.layers.Dense(num_classes, activation='softmax')
])

# Normally model would be trained, here we just compile and predict for demonstration
model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])

# Predict probabilities
pred_probs = model.predict(X_test)

# Convert probabilities to predicted class labels
pred_labels = np.argmax(pred_probs, axis=1)

# Generate classification report
report = classification_report(true_labels, pred_labels, target_names=[f'Class {i}' for i in range(num_classes)])
print(report)

Added code to predict class probabilities using the TensorFlow model.

Converted predicted probabilities to class labels using argmax.

Used sklearn's classification_report to compute precision, recall, and F1-score per class.

Printed the classification report for detailed evaluation.

Results Interpretation

Before: Only overall accuracy was known (85%).

After: Detailed classification report shows precision, recall, and F1-score for each class, revealing strengths and weaknesses per category.

Classification reports provide detailed insights beyond overall accuracy, helping to understand model performance on each class and guiding improvements.

Bonus Experiment

Try generating a classification report with a confusion matrix heatmap visualization.

💡 Hint

Use sklearn.metrics.confusion_matrix and matplotlib's heatmap to visualize class prediction errors.

Practice

1. What does a classification report in TensorFlow typically show?

easy

A. Precision, recall, and F1-score for each class

B. The model's training time

C. The number of layers in the model

D. The size of the input data

Solution

1. Step 1: Understand classification report contents

   A classification report summarizes key metrics like precision, recall, and F1-score for each class predicted by the model.

2. Step 2: Identify what is NOT included

   Training time, model layers, and input size are not part of the classification report.

3. Final Answer:

   Precision, recall, and F1-score for each class -> Option A

4. Quick Check:

   Classification report = precision, recall, F1-score [OK]

Hint: Look for metrics per class, not model structure or time [OK]

Common Mistakes:

• Confusing training time with performance metrics
• Thinking model architecture details are in the report
• Assuming input data size is shown

2. Which code snippet correctly imports the function to generate a classification report in TensorFlow's ecosystem?

easy

A. import classification_report from tensorflow

B. import tensorflow.classification_report as cr

C. from sklearn.metrics import classification_report

D. from tensorflow.metrics import classification_report

Solution

1. Step 1: Recall the source of classification_report

   The classification_report function is part of scikit-learn, not TensorFlow directly.

2. Step 2: Check import syntax correctness

   from sklearn.metrics import classification_report uses correct Python import syntax from sklearn.metrics. Other options are invalid or incorrect.

3. Final Answer:

   from sklearn.metrics import classification_report -> Option C

4. Quick Check:

   Import classification_report from sklearn.metrics [OK]

Hint: classification_report is from sklearn, not tensorflow [OK]

Common Mistakes:

• Trying to import classification_report directly from tensorflow
• Using incorrect import syntax
• Confusing TensorFlow metrics with sklearn metrics

3. Given the following code, what will be the printed output?

from sklearn.metrics import classification_report
true_labels = [0, 1, 0, 1]
pred_labels = [0, 0, 0, 1]
print(classification_report(true_labels, pred_labels))

medium

A. Empty output with no metrics

B. A report showing precision, recall, and F1-score for classes 0 and 1

C. An error because labels are integers

D. Only accuracy score printed

Solution

1. Step 1: Understand input labels and predictions

   Labels are valid integers representing two classes: 0 and 1. Predictions have one mistake at index 1.

2. Step 2: What classification_report prints

   The function prints precision, recall, F1-score, and support for each class, not just accuracy or errors.

3. Final Answer:

   A report showing precision, recall, and F1-score for classes 0 and 1 -> Option B

4. Quick Check:

   classification_report outputs metrics per class [OK]

Hint: classification_report prints detailed metrics, not just accuracy [OK]

Common Mistakes:

• Expecting an error due to integer labels
• Thinking only accuracy is printed
• Assuming no output if predictions differ

4. Identify the error in this code snippet that tries to print a classification report:

from sklearn.metrics import classification_report
true = [1, 0, 1]
pred = [1, 0]
print(classification_report(true, pred))

medium

A. print statement syntax is wrong

B. classification_report is not imported correctly

C. Labels must be strings, not integers

D. The true and pred lists have different lengths

Solution

1. Step 1: Check input list lengths

   True labels list has 3 elements, predicted labels list has 2 elements. They must be the same length.

2. Step 2: Verify other code parts

   Import and print syntax are correct. Labels can be integers.

3. Final Answer:

   The true and pred lists have different lengths -> Option D

4. Quick Check:

   Input lists must match in length [OK]

Hint: Check if true and predicted label lists have same length [OK]

Common Mistakes:

• Ignoring length mismatch
• Thinking labels must be strings
• Assuming import or print is wrong

5. You trained a TensorFlow model for 3 classes: cat, dog, and rabbit. After predictions, you get this classification report:

              precision    recall  f1-score   support

         cat       0.90      0.95      0.92       100
         dog       0.70      0.60      0.65       100
      rabbit       0.50      0.40      0.44       100

    accuracy                           0.68       300
   macro avg       0.70      0.65      0.67       300
weighted avg       0.70      0.68      0.68       300

What should you focus on to improve your model?

hard

A. Improve the model's performance on the rabbit class

B. Increase the support for the cat class

C. Reduce the number of classes

D. Ignore the dog class since it has moderate scores

Solution

1. Step 1: Analyze class-wise metrics

   Rabbit class has the lowest precision (0.50), recall (0.40), and F1-score (0.44), indicating poor performance.

2. Step 2: Decide improvement focus

   Improving rabbit predictions will raise overall accuracy and balance. Increasing support or ignoring classes is not helpful.

3. Final Answer:

   Improve the model's performance on the rabbit class -> Option A

4. Quick Check:

   Lowest metrics class needs improvement [OK]

Hint: Focus on class with lowest F1-score to improve model [OK]

Common Mistakes:

• Trying to increase support instead of improving metrics
• Ignoring poorly performing classes
• Reducing classes without reason