Model evaluation best practices in Computer Vision - ML Experiment: Train & Evaluate

Experiment - Model evaluation best practices

Problem:You have trained a computer vision model to classify images into 5 categories. The model shows 95% accuracy on training data but only 70% accuracy on validation data.

Current Metrics:Training accuracy: 95%, Validation accuracy: 70%, Validation loss: 1.2

Issue:The model is overfitting. It performs very well on training data but poorly on validation data, indicating it does not generalize well.

Your Task

Reduce overfitting and improve validation accuracy to at least 85% while keeping training accuracy below 90%.

You can only modify model evaluation and training best practices, not the model architecture.

Do not add new data or change the dataset.

Hint 1

Hint 2

Hint 3

Hint 4

Solution

Computer Vision

import tensorflow as tf
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras.callbacks import EarlyStopping
from sklearn.metrics import classification_report, confusion_matrix

# Assume X_train, y_train, X_val, y_val are preloaded image data and labels

# Data augmentation setup
train_datagen = ImageDataGenerator(
    rotation_range=20,
    width_shift_range=0.2,
    height_shift_range=0.2,
    horizontal_flip=True
)

val_datagen = ImageDataGenerator()

train_generator = train_datagen.flow(X_train, y_train, batch_size=32)
val_generator = val_datagen.flow(X_val, y_val, batch_size=32, shuffle=False)

# Early stopping callback
early_stop = EarlyStopping(monitor='val_loss', patience=5, restore_best_weights=True)

# Assume model is predefined
model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])

history = model.fit(
    train_generator,
    epochs=50,
    validation_data=val_generator,
    callbacks=[early_stop]
)

# Evaluate on validation data
val_loss, val_accuracy = model.evaluate(val_generator)

# Predict on validation data
y_pred_probs = model.predict(val_generator)
y_pred = y_pred_probs.argmax(axis=1)

# Print classification report and confusion matrix
print(classification_report(y_val, y_pred))
print(confusion_matrix(y_val, y_pred))

Added data augmentation to training data to improve model generalization.

Implemented early stopping to stop training when validation loss stops improving.

Used proper validation data generator without augmentation for evaluation.

Evaluated model with classification report and confusion matrix for detailed metrics.

Results Interpretation

Before: Training accuracy: 95%, Validation accuracy: 70%, Validation loss: 1.2

After: Training accuracy: 88%, Validation accuracy: 86%, Validation loss: 0.6

Using best practices like data augmentation and early stopping helps reduce overfitting, improving validation accuracy and model generalization.

Bonus Experiment

Try using k-fold cross-validation to evaluate the model more reliably across different data splits.

💡 Hint

Use sklearn's KFold to split data and train the model multiple times, then average the validation metrics.

Practice

1. Why is it important to use a separate test set when evaluating a computer vision model?

easy

A. To check how well the model performs on new, unseen data

B. To make the training process faster

C. To increase the size of the training data

D. To reduce the number of model parameters

Solution

1. Step 1: Understand the purpose of a test set

   The test set is data the model has never seen before, used to check real-world performance.

2. Step 2: Compare test set role with other options

   Options B, C, and D do not relate to evaluation but to training or model design.

3. Final Answer:

   To check how well the model performs on new, unseen data -> Option A

4. Quick Check:

   Test set = unseen data check [OK]

Hint: Test set = new data to check model accuracy [OK]

Common Mistakes:

• Confusing test set with training set
• Thinking test set speeds up training
• Believing test set changes model size

2. Which of the following is the correct way to split data for model evaluation in Python using scikit-learn?

easy

A. split_train_test(data, 0.2)

B. train_test_split(data, test_size=0.2, random_state=42)

C. train_test(data, 0.2)

D. test_train_split(data, 0.2)

Solution

1. Step 1: Recall the correct function name in scikit-learn

   The function to split data is called train_test_split with parameters like test_size and random_state.

2. Step 2: Check the options for correct syntax

   Only train_test_split(data, test_size=0.2, random_state=42) uses the correct function name and parameters; others are invalid or do not exist.

3. Final Answer:

   train_test_split(data, test_size=0.2, random_state=42) -> Option B

4. Quick Check:

   Correct function = train_test_split [OK]

Hint: Remember scikit-learn function: train_test_split [OK]

Common Mistakes:

• Using wrong function names
• Missing required parameters
• Confusing order of train and test

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

from sklearn.metrics import accuracy_score
true_labels = [1, 0, 1, 1, 0]
pred_labels = [1, 0, 0, 1, 0]
accuracy = accuracy_score(true_labels, pred_labels)
print("{:.2f}".format(round(accuracy, 2)))

medium

A. 0.60

B. 0.40

C. 0.80

D. 1.00

Solution

1. Step 1: Compare true and predicted labels

   True: [1, 0, 1, 1, 0], Predicted: [1, 0, 0, 1, 0]. Matches at positions 0,1,3,4 (4 correct out of 5).

2. Step 2: Calculate accuracy

   Accuracy = correct predictions / total = 4/5 = 0.8. Rounded to 2 decimals is 0.80.

3. Final Answer:

   0.80 -> Option C

4. Quick Check:

   Accuracy = 4/5 = 0.80 [OK]

Hint: Count matches, divide by total labels [OK]

Common Mistakes:

• Counting wrong matches
• Not rounding accuracy
• Confusing accuracy with precision

4. You trained a model but the test accuracy is much higher than the training accuracy. What is the most likely issue?

medium

A. Data leakage between training and test sets

B. Model is underfitting the training data

C. Test set is too small

D. Training data is too large

Solution

1. Step 1: Understand unusual accuracy pattern

   Test accuracy higher than training is unusual and often means test data was seen during training.

2. Step 2: Identify cause from options

   Data leakage means test data accidentally used in training, causing inflated test accuracy.

3. Final Answer:

   Data leakage between training and test sets -> Option A

4. Quick Check:

   High test accuracy > training = data leakage [OK]

Hint: High test accuracy than train? Check data leakage [OK]

Common Mistakes:

• Assuming underfitting causes higher test accuracy
• Ignoring data leakage possibility
• Blaming test set size without evidence

5. You want to evaluate a computer vision model for detecting rare objects in images. Which evaluation metric is best to use and why?

hard

A. Confusion matrix, because it shows training time

B. Accuracy, because it shows overall correct predictions

C. Mean Squared Error, because it measures prediction error

D. F1 score, because it balances precision and recall for imbalanced data

Solution

1. Step 1: Understand the problem of rare object detection

   Rare objects mean data is imbalanced; many negatives, few positives.

2. Step 2: Choose metric suitable for imbalanced data

   F1 score balances precision (correct positive predictions) and recall (finding all positives), ideal for rare classes.

3. Final Answer:

   F1 score, because it balances precision and recall for imbalanced data -> Option D

4. Quick Check:

   Rare class? Use F1 score [OK]

Hint: Rare classes? Use F1 score for balance [OK]

Common Mistakes:

• Using accuracy which hides imbalance
• Confusing regression metrics with classification
• Misunderstanding confusion matrix purpose