K-fold cross-validation in TensorFlow - Model Metrics & Evaluation

Metrics & Evaluation - K-fold cross-validation

Which metric matters for K-fold cross-validation and WHY

K-fold cross-validation helps us check how well a model will perform on new data. It splits data into parts (folds), trains on some, tests on others, and repeats. The key metrics to watch are average accuracy, average loss, or other scores like F1-score across all folds. This shows if the model is stable and reliable, not just lucky on one split.

Confusion matrix or equivalent visualization

For each fold, we get a confusion matrix like this (example for binary classification):

      Fold 1 Confusion Matrix:
      -------------------------
      | TP=40 | FP=10 |
      | FN=5  | TN=45 |
      -------------------------

      Fold 2 Confusion Matrix:
      -------------------------
      | TP=38 | FP=12 |
      | FN=7  | TN=43 |
      -------------------------

      ...

      We average metrics from all folds to get a final performance estimate.
    

Precision vs Recall tradeoff with concrete examples

K-fold cross-validation helps us see if precision and recall are consistent across data splits. For example:

• If precision is high but recall varies a lot between folds, the model might miss some important cases sometimes.
• If recall is high but precision drops in some folds, the model might give many false alarms on some data.

By checking these tradeoffs in each fold, we can choose a model that balances precision and recall well on all data parts.

What "good" vs "bad" metric values look like for K-fold cross-validation

Good: Metrics like accuracy, precision, recall, or F1-score are similar across all folds (small difference). For example, accuracy around 85% ± 2% across folds means stable performance.

Bad: Metrics vary a lot between folds, like accuracy 90% in one fold but 60% in another. This means the model is unstable and might not work well on new data.

Common pitfalls when using metrics with K-fold cross-validation

• Data leakage: If data from the test fold leaks into training folds, metrics will be too optimistic.
• Ignoring variance: Reporting only average metric hides if model is unstable across folds.
• Overfitting folds: Tuning model too much on cross-validation data can cause overfitting, making metrics misleading.
• Imbalanced data: If classes are uneven, accuracy can be misleading; use precision, recall, or F1 instead.

Self-check question

Your model has 98% accuracy but only 12% recall on fraud cases in K-fold cross-validation. Is it good for production? Why or why not?

Answer: No, it is not good. The low recall means the model misses most fraud cases, which is dangerous. Even with high accuracy, the model fails to catch fraud, so it should be improved before use.

Key Result

K-fold cross-validation averages metrics across folds to ensure model stability and reliable performance estimates.

Practice

1. What is the main purpose of using K-fold cross-validation in TensorFlow model training?

easy

A. To increase the size of the training dataset

B. To evaluate model performance on different parts of the data

C. To speed up the training process

D. To reduce the number of model parameters

Solution

1. Step 1: Understand K-fold cross-validation concept

   K-fold cross-validation splits the dataset into K parts and trains the model K times, each time using a different part as test data.

2. Step 2: Identify the purpose of this process

   This method helps evaluate how well the model performs on unseen data by testing it on different subsets.

3. Final Answer:

   To evaluate model performance on different parts of the data -> Option B

4. Quick Check:

   Purpose = Evaluate model on different data parts [OK]

Hint: K-fold tests model on all data parts for better evaluation [OK]

Common Mistakes:

• Thinking it increases training data size
• Believing it speeds up training
• Confusing it with parameter tuning

2. Which of the following is the correct way to split data into 5 folds using TensorFlow's tf.data API for K-fold cross-validation?

easy

A. Use tf.data.Dataset.shard(num_shards=5, index=i) inside a loop over i from 0 to 4

B. Use tf.data.Dataset.batch(5) to create 5 folds

C. Use tf.data.Dataset.repeat(5) to repeat data 5 times

D. Use tf.data.Dataset.shuffle(5) to shuffle into 5 folds

Solution

1. Step 1: Recall TensorFlow data splitting method

   TensorFlow's shard method splits a dataset into equal parts by specifying total shards and the index of the shard.

2. Step 2: Match method to K-fold splitting

   For 5 folds, use num_shards=5 and loop over index from 0 to 4 to get each fold.

3. Final Answer:

   Use tf.data.Dataset.shard(num_shards=5, index=i) inside a loop over i from 0 to 4 -> Option A

4. Quick Check:

   Sharding splits data into folds correctly [OK]

Hint: Use shard with num_shards = K for K-fold splits [OK]

Common Mistakes:

• Using batch instead of shard for splitting
• Repeating data instead of splitting
• Shuffling does not create folds

3. Consider this code snippet for 3-fold cross-validation in TensorFlow:

for i in range(3):
    test_data = dataset.shard(num_shards=3, index=i)
    train_data = dataset.shard(num_shards=3, index=(i+1) % 3)
    # model training code here
    print(f'Fold {i} train size:', len(list(train_data)))
    print(f'Fold {i} test size:', len(list(test_data)))

What will be the output sizes for train and test data in each fold if the original dataset has 9 samples?

medium

A. Train size: 9, Test size: 0 for all folds

B. Train size: 6, Test size: 3 for all folds

C. Train size: 3, Test size: 3 for all folds

D. Train size: 3, Test size: 6 for all folds

Solution

1. Step 1: Understand shard usage in code

   Each fold uses dataset.shard(num_shards=3, index=i) for test data and dataset.shard(num_shards=3, index=(i+1) % 3) for train data. This means train and test data are both single shards, not complementary sets.

2. Step 2: Calculate sizes based on shards

   Each shard has 3 samples (9/3). So test_data has 3 samples. But train_data is only one shard (3 samples), not the other two shards combined (6 samples). So train and test overlap and total is 6 samples, missing 3 samples.

3. Step 3: Check code behavior

   Since train_data and test_data are both shards, the rest of the data is ignored. The code prints sizes of 3 for train and 3 for test, but the rest 3 samples are missing from both.

4. Final Answer:

   Train size: 3, Test size: 3 for all folds -> Option C

5. Quick Check:

   Shard selects one part only, so train = 3, test = 3 [OK]

Hint: Shard picks one part; train must combine other shards [OK]

Common Mistakes:

• Assuming train is all except test shard
• Counting total dataset size for train
• Ignoring shard overlap

4. You wrote this code for 4-fold cross-validation but get an error:

for i in range(4):
    test_data = dataset.shard(num_shards=4, index=i)
    train_data = dataset.shard(num_shards=4, index!=i)
    # train model here

What is the error and how to fix it?

medium

A. Syntax error in index!=i; use a loop to combine other shards

B. Wrong number of shards; should be 5 instead of 4

C. Missing shuffle before sharding causes error

D. Using shard twice causes conflict; use split instead

Solution

1. Step 1: Identify syntax error in shard index

   The expression index!=i is invalid syntax for the index parameter, which expects an integer, not a boolean expression.

2. Step 2: Correct approach to get train data

   To get train data (all shards except test shard), you must combine shards with indices other than i using a loop or filtering, not a single shard call with index!=i.

3. Final Answer:

   Syntax error in index!=i; use a loop to combine other shards -> Option A

4. Quick Check:

   Shard index must be integer, not condition [OK]

Hint: Shard index needs integer; combine others in loop [OK]

Common Mistakes:

• Using boolean condition as shard index
• Assuming shard auto-excludes test fold
• Not combining shards for train data

5. You want to implement 5-fold cross-validation in TensorFlow to train a neural network on a dataset with 1000 samples. Which approach correctly creates train and test datasets for each fold to avoid data leakage?

hard

A. For each fold i, use test_data = dataset.shard(5, i) and train_data = dataset.filter(lambda x: x not in test_data)

B. For each fold i, use test_data = dataset.shard(5, i) and train_data = dataset.shard(5, (i+1) % 5)

C. For each fold i, use dataset.shuffle(1000) and split 80% train, 20% test randomly

D. For each fold i, use test_data = dataset.shard(5, i) and combine all other shards except i for train_data

Solution

1. Step 1: Understand correct train-test split in K-fold

   Test data is one shard. Train data must be all other shards combined to avoid overlap and data leakage.

2. Step 2: Evaluate options

   For each fold i, use test_data = dataset.shard(5, i) and train_data = dataset.filter(lambda x: x not in test_data) tries to filter dataset by membership, which is not supported directly in TensorFlow datasets. For each fold i, use test_data = dataset.shard(5, i) and train_data = dataset.shard(5, (i+1) % 5) uses only one shard for train, missing other data. For each fold i, use dataset.shuffle(1000) and split 80% train, 20% test randomly shuffles and splits randomly, not K-fold. For each fold i, use test_data = dataset.shard(5, i) and combine all other shards except i for train_data correctly combines all shards except the test shard for training.

3. Final Answer:

   For each fold i, use test_data = dataset.shard(5, i) and combine all other shards except i for train_data -> Option D

4. Quick Check:

   Train = all shards except test shard [OK]

Hint: Train = all shards except test shard combined [OK]

Common Mistakes:

• Using only one shard for train data
• Filtering datasets by membership incorrectly
• Random split instead of K-fold