Retraining strategies in ML Python - Model Metrics & Evaluation

Metrics & Evaluation - Retraining strategies

Which metric matters for Retraining strategies and WHY

When we retrain a model, we want to see if it learns better from new data. Key metrics to watch are validation loss and validation accuracy. These tell us if the model is improving on unseen data, not just memorizing old data.

Also, watch precision and recall if your task is classification. They show if the model is correctly finding important cases (recall) and avoiding false alarms (precision).

Finally, track training time and resource use to balance model quality with cost.

Confusion matrix example during retraining

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

      Example after retraining:
      TP = 85, FP = 15, TN = 900, FN = 20

      Total samples = 85 + 15 + 900 + 20 = 1020

      Precision = TP / (TP + FP) = 85 / (85 + 15) = 0.85
      Recall = TP / (TP + FN) = 85 / (85 + 20) = 0.81
    

Precision vs Recall tradeoff in retraining

When retraining, improving one metric can lower the other. For example:

• High precision means fewer false alarms. Good if false alarms are costly, like spam filters.
• High recall means fewer missed cases. Important if missing a case is dangerous, like in medical diagnosis.

Retraining can help find a better balance by learning from new examples. But watch if improving recall drops precision too much or vice versa.

Good vs Bad metric values for retraining

Good retraining results:

• Validation accuracy improves or stays stable.
• Precision and recall both increase or maintain a good balance.
• Loss on validation data decreases.
• No big jump in training time or resource use.

Bad retraining results:

• Validation accuracy drops or fluctuates wildly.
• Precision or recall drops significantly.
• Validation loss increases, showing worse generalization.
• Training time grows too long without metric gains.

Common pitfalls in retraining metrics

• Overfitting: Training loss goes down but validation loss goes up. Model memorizes new data but fails on unseen data.
• Data leakage: New training data accidentally includes test data, inflating metrics falsely.
• Ignoring metric drift: Not checking if metrics degrade over time due to changing data patterns.
• Using only accuracy: Accuracy can be misleading if classes are imbalanced. Use precision and recall too.

Self-check question

Your model after retraining 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 means many frauds go undetected, which is risky. You should improve recall before using it in production.

Key Result

Retraining success is shown by improved validation metrics like loss, accuracy, precision, and recall, balanced with training cost.

Practice

1. What is the main purpose of retraining a machine learning model?

easy

A. To remove features from the model permanently

B. To make the model run faster on old data

C. To reduce the size of the training dataset

D. To update the model with new data and keep it accurate

Solution

1. Step 1: Understand retraining purpose

   Retraining updates the model using new data to maintain or improve accuracy.

2. Step 2: Compare options

   Options B, C, and D do not describe retraining goals; only A matches the purpose.

3. Final Answer:

   To update the model with new data and keep it accurate -> Option D

4. Quick Check:

   Retraining = update model with new data [OK]

Hint: Retraining means learning from new data to stay accurate [OK]

Common Mistakes:

• Thinking retraining speeds up old data processing
• Confusing retraining with data reduction
• Believing retraining removes features

2. Which of the following is the correct way to fine-tune a pretrained model in Python using TensorFlow?

easy

A. model.fit(new_data, epochs=5, initial_epoch=0)

B. model.train(new_data, epochs=5)

C. model.evaluate(new_data, epochs=5)

D. model.fit(new_data, epochs=5, initial_epoch=existing_epochs)

Solution

1. Step 1: Understand fine-tuning syntax

   Fine-tuning a pretrained model typically uses model.fit starting from epoch 0 after loading pretrained weights.

2. Step 2: Analyze options

   A uses non-existent method 'train'. C uses evaluate, which doesn't train and lacks epochs param. D risks initial_epoch >= epochs error. B correctly trains with model.fit from epoch 0.

3. Final Answer:

   model.fit(new_data, epochs=5, initial_epoch=0) -> Option A

4. Quick Check:

   Fine-tune = model.fit(new_data, epochs=N, initial_epoch=0) [OK]

Hint: Fine-tune pretrained with model.fit(new_data, epochs=N) [OK]

Common Mistakes:

• Using model.train instead of model.fit
• Setting initial_epoch=existing_epochs causing error
• Confusing evaluation with training

3. Consider this Python code snippet for incremental retraining:

old_accuracy = 0.85
model.fit(new_data, epochs=3)
new_accuracy = 0.88
improvement = new_accuracy - old_accuracy
print(improvement)

What will be the printed output?

medium

A. 0.85

B. 0.88

C. 0.03

D. 3

Solution

1. Step 1: Calculate improvement

   Improvement is new_accuracy minus old_accuracy: 0.88 - 0.85 = 0.03.

2. Step 2: Understand print output

   The print statement outputs the improvement variable, which is 0.03.

3. Final Answer:

   0.03 -> Option C

4. Quick Check:

   0.88 - 0.85 = 0.03 [OK]

Hint: Subtract old accuracy from new to find improvement [OK]

Common Mistakes:

• Printing new_accuracy instead of improvement
• Confusing epochs with accuracy values
• Printing old_accuracy by mistake

4. You have this code snippet for retraining a model incrementally:

model.fit(new_data, epochs=5)
model.fit(new_data, epochs=3, initial_epoch=5)

But the second fit call throws an error. What is the likely cause?

medium

A. initial_epoch should be less than epochs

B. initial_epoch cannot be greater than or equal to epochs

C. You must reset the model before second fit

D. epochs must be the same in both fit calls

Solution

1. Step 1: Understand initial_epoch and epochs relation

   initial_epoch must be less than epochs in the current fit call to define the training range properly.

2. Step 2: Analyze the error cause

   Here, initial_epoch=5 is greater than epochs=3, which is invalid and causes an error.

3. Final Answer:

   initial_epoch cannot be greater than or equal to epochs -> Option B

4. Quick Check:

   initial_epoch < epochs required [OK]

Hint: initial_epoch must be less than epochs in fit call [OK]

Common Mistakes:

• Thinking model reset is needed
• Assuming epochs must match both calls
• Setting initial_epoch greater or equal to epochs

5. You have a model trained on data from 2020. New data from 2023 shows some feature distributions changed. Which retraining strategy best balances training time and model accuracy?

hard

A. Fine-tuning the existing model on new data with a low learning rate

B. Full retraining from scratch with combined old and new data

C. Incremental retraining only on new data without old data

D. No retraining; use the old model as is

Solution

1. Step 1: Consider data changes and retraining options

   Feature distribution changed, so model needs updating but full retraining is costly.

2. Step 2: Evaluate strategies

   Full retraining (A) is accurate but slow. Incremental retraining (C) risks forgetting old data. No retraining (D) ignores changes. Fine-tuning (B) updates model efficiently with new data.

3. Final Answer:

   Fine-tuning the existing model on new data with a low learning rate -> Option A

4. Quick Check:

   Fine-tune balances accuracy and training time [OK]

Hint: Fine-tune to update model efficiently with new data [OK]

Common Mistakes:

• Choosing full retraining despite high cost
• Ignoring old data in incremental retraining
• Skipping retraining when data changes