Why model persistence enables deployment in TensorFlow - Why Metrics Matter

Metrics & Evaluation - Why model persistence enables deployment

Which metric matters and WHY

When we save a model to use it later (model persistence), the key metric to watch is model consistency. This means the model's predictions and performance should stay the same before saving and after loading. This ensures the model works reliably when deployed.

Confusion matrix or equivalent visualization

Imagine a saved model tested on 100 samples before saving and after loading. The confusion matrix should be the same:

      | Predicted Positive | Predicted Negative |
      |--------------------|--------------------|
      | True Positive (TP): 40 | False Negative (FN): 10 |
      | False Positive (FP): 5 | True Negative (TN): 45 |
    

The totals add up to 100 samples. If the confusion matrix changes after loading, the model persistence failed.

Precision vs Recall tradeoff with examples

Model persistence should keep the balance between precision and recall stable. For example:

• If precision drops after loading, the model may wrongly label more negatives as positives.
• If recall drops, the model may miss more positive cases.

Both changes hurt deployment trust. Persistence must keep these metrics consistent.

What "good" vs "bad" metric values look like

Good: Model before saving and after loading has the same accuracy, precision, recall, and F1 score. For example, accuracy 90%, precision 85%, recall 80%, F1 82%.

Bad: After loading, accuracy drops to 70%, precision to 60%, recall to 50%. This means the saved model is corrupted or incompatible.

Common pitfalls in model persistence metrics

• Data leakage: If test data leaks into training, metrics look better but deployment fails.
• Overfitting: Model performs well on training but poorly after loading on new data.
• Version mismatch: Saving and loading with different TensorFlow versions can corrupt the model.
• Incomplete saving: Forgetting to save optimizer state or custom layers causes errors or metric changes.

Self-check question

Your model has 98% accuracy but 12% recall on fraud detection after loading. Is it good for deployment? Why or why not?

Answer: No, it is not good. High accuracy can be misleading if fraud cases are rare. Low recall means the model misses most fraud cases, which is dangerous. Model persistence failed to keep recall high, so deployment is risky.

Key Result

Model persistence must keep key metrics like accuracy, precision, and recall consistent before and after saving to ensure reliable deployment.

Practice

1. Why is model persistence important for deploying machine learning models?

easy

A. It visualizes the model's architecture in a graph.

B. It improves the model's accuracy during training.

C. It automatically tunes hyperparameters for better performance.

D. It allows saving a trained model to use later without retraining.

Solution

1. Step 1: Understand model persistence meaning

   Model persistence means saving a trained model so it can be reused later.

2. Step 2: Connect persistence to deployment

   Saving the model avoids retraining every time, making deployment faster and easier.

3. Final Answer:

   It allows saving a trained model to use later without retraining. -> Option D

4. Quick Check:

   Model persistence = save and reuse model [OK]

Hint: Persistence means saving model for reuse without retraining [OK]

Common Mistakes:

• Confusing persistence with improving accuracy
• Thinking persistence tunes hyperparameters
• Believing persistence visualizes the model

2. Which of the following is the correct way to save a TensorFlow Keras model?

easy

A. model.save('model_name.h5')

B. model.load('model_name.h5')

C. save_model(model, 'model_name.h5')

D. model.export('model_name.h5')

Solution

1. Step 1: Recall TensorFlow Keras save syntax

   The correct method to save a model is using model.save() with a filename.

2. Step 2: Check options for correct method

   Only model.save('model_name.h5') matches the correct syntax.

3. Final Answer:

   model.save('model_name.h5') -> Option A

4. Quick Check:

   Save model = model.save() [OK]

Hint: Use model.save() to save models in TensorFlow [OK]

Common Mistakes:

• Using model.load() to save instead of load
• Confusing save_model() function with method
• Using non-existent model.export() method

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

import tensorflow as tf
from tensorflow import keras

model = keras.Sequential([
    keras.layers.Dense(5, activation='relu', input_shape=(3,)),
    keras.layers.Dense(1)
])
model.save('my_model')
loaded_model = keras.models.load_model('my_model')
print(type(loaded_model))

medium

A. <class 'tensorflow.python.keras.engine.sequential.Sequential'>

B. <class 'tensorflow.keras.layers.Dense'>

C. None

D. Error: model not found

Solution

1. Step 1: Understand saving and loading model

   Saving with model.save() and loading with keras.models.load_model() returns the same model type.

2. Step 2: Check the type of loaded_model

   The loaded model is a Sequential model, so its type is tensorflow.python.keras.engine.sequential.Sequential.

3. Final Answer:

   <class 'tensorflow.python.keras.engine.sequential.Sequential'> -> Option A

4. Quick Check:

   Load model type = Sequential class [OK]

Hint: Loaded model keeps original class type [OK]

Common Mistakes:

• Expecting layer type instead of model type
• Assuming load_model returns None
• Thinking file not found error occurs without saving

4. You saved a model using model.save('model_path') but get an error when loading it with keras.models.load_model('wrong_path'). What is the likely cause?

medium

A. The model was not trained before saving.

B. The load path is incorrect or does not exist.

C. The model architecture is incompatible with load_model.

D. The model was saved in a wrong file format.

Solution

1. Step 1: Identify error cause from load_model

   Loading a model from a wrong or non-existing path causes a file not found error.

2. Step 2: Check other options for correctness

   Model training or architecture issues do not cause path errors; file format is handled by save/load methods.

3. Final Answer:

   The load path is incorrect or does not exist. -> Option B

4. Quick Check:

   Load error = wrong path [OK]

Hint: Check file path carefully when loading models [OK]

Common Mistakes:

• Assuming model must be trained before saving
• Blaming architecture mismatch for file errors
• Thinking save format causes load errors

5. You want to deploy a TensorFlow model on a web server. Which approach best uses model persistence to enable fast deployment?

hard

A. Train the model on the server every time a user requests a prediction.

B. Save only the model weights and retrain the architecture on each request.

C. Save the trained model locally and load it once when the server starts.

D. Use model.save() to save the model but never load it during deployment.

Solution

1. Step 1: Understand deployment needs

   Deployment requires fast response, so retraining on each request is too slow.

2. Step 2: Use model persistence for efficiency

   Saving the trained model and loading it once at server start allows quick predictions without retraining.

3. Step 3: Evaluate other options

   Retraining weights or not loading saved model causes delays or errors.

4. Final Answer:

   Save the trained model locally and load it once when the server starts. -> Option C

5. Quick Check:

   Load once, predict many = fast deployment [OK]

Hint: Load saved model once at start for fast predictions [OK]

Common Mistakes:

• Retraining model on every request wastes time
• Saving weights only requires retraining architecture
• Not loading saved model prevents deployment