Pre-trained models (VGG, ResNet, MobileNet) in TensorFlow - Model Metrics & Evaluation

Metrics & Evaluation - Pre-trained models (VGG, ResNet, MobileNet)

Which metric matters for Pre-trained models and WHY

When using pre-trained models like VGG, ResNet, or MobileNet, the key metrics to watch are accuracy and loss during training and validation. Accuracy tells us how often the model predicts correctly. Loss shows how far off the model's predictions are from the true answers. Since these models are often used for image classification, accuracy is a simple and clear way to see if the model is learning well.

Additionally, if the task is imbalanced (some classes appear more than others), metrics like precision, recall, and F1 score become important to understand how well the model handles each class.

Confusion matrix example

Suppose we have a 3-class image classifier using a pre-trained model. Here is a confusion matrix showing predictions:

      | Predicted Class A | Predicted Class B | Predicted Class C |
      |-------------------|-------------------|-------------------|
      | True Class A: 50  | 5                 | 0                 |
      | True Class B: 3   | 45                | 2                 |
      | True Class C: 1   | 4                 | 40                |
    

This means out of 55 true Class A images, 50 were correctly predicted (True Positives for A), 5 were wrongly predicted as B (False Negatives for A, False Positives for B). Similar logic applies for other classes.

Precision vs Recall tradeoff with examples

Using pre-trained models, sometimes you want to tune for precision or recall depending on the task:

• High Precision: If you use MobileNet for a face recognition app that unlocks a phone, you want high precision. This means when the model says "this is you," it should be very sure to avoid unlocking for the wrong person.
• High Recall: If you use ResNet to detect rare diseases in medical images, high recall is important. You want to catch as many disease cases as possible, even if some false alarms happen.

Pre-trained models can be fine-tuned to balance precision and recall based on these needs.

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

For pre-trained models on common image tasks:

• Good: Accuracy above 85% on validation data, low loss, balanced precision and recall above 80% for all classes.
• Bad: Accuracy below 60%, high loss, or very low precision/recall for some classes (e.g., 30%). This means the model is not learning well or is biased.

Remember, a high accuracy alone can be misleading if the data is imbalanced.

Common pitfalls with metrics for pre-trained models

• Accuracy paradox: High accuracy can happen if one class dominates the data. The model might ignore smaller classes.
• Data leakage: If test images are too similar to training images, metrics look better but model won't generalize.
• Overfitting: Training accuracy is high but validation accuracy is low. The model memorizes training data but fails on new data.
• Ignoring class imbalance: Not checking precision and recall per class can hide poor performance on rare classes.

Self-check question

Your pre-trained model fine-tuned on a dataset has 98% accuracy but only 12% recall on the rare class. Is it good for production? Why or why not?

Answer: No, it is not good. The model misses most examples of the rare class (low recall), which could be critical if that class is important (like a disease). High accuracy is misleading because the rare class is small compared to others.

Key Result

Accuracy is key for overall performance, but precision and recall reveal class-specific strengths and weaknesses in pre-trained models.

Practice

1. Which of the following is a key benefit of using pre-trained models like VGG, ResNet, or MobileNet in TensorFlow?

easy

A. They require training from scratch for every new task.

B. They cannot be fine-tuned for new tasks.

C. They only work with text data, not images.

D. They save time by using models already trained on large datasets.

Solution

1. Step 1: Understand what pre-trained models are

   Pre-trained models are AI models trained on large datasets like ImageNet before you use them.

2. Step 2: Identify the main advantage

   Using these models saves time because you don't need to train from zero; you can use their learned features directly.

3. Final Answer:

   They save time by using models already trained on large datasets. -> Option D

4. Quick Check:

   Pre-trained models save time = B [OK]

Hint: Pre-trained means already trained, so saves time [OK]

Common Mistakes:

• Thinking pre-trained models need full retraining
• Believing they only work on text data
• Assuming they cannot be fine-tuned

2. Which of the following is the correct way to load the MobileNetV2 pre-trained model in TensorFlow with weights trained on ImageNet?

easy

A. tf.keras.MobileNetV2(weights='imagenet')

B. tf.keras.applications.MobileNetV2(weights='imagenet')

C. tf.keras.applications.MobileNet(weights='imagenet')

D. tf.keras.applications.MobileNetV2(weights='random')

Solution

1. Step 1: Recall the correct TensorFlow module for pre-trained models

   Pre-trained models are loaded from tf.keras.applications, not directly from tf.keras.

2. Step 2: Check the correct model name and weights argument

   The correct model name is MobileNetV2 and weights='imagenet' loads pre-trained weights.

3. Final Answer:

   tf.keras.applications.MobileNetV2(weights='imagenet') -> Option B

4. Quick Check:

   Use tf.keras.applications + correct model name + weights='imagenet' = A [OK]

Hint: Use tf.keras.applications.ModelName(weights='imagenet') [OK]

Common Mistakes:

• Using tf.keras.MobileNetV2 instead of tf.keras.applications.MobileNetV2
• Using wrong model name like MobileNet instead of MobileNetV2
• Setting weights='random' instead of 'imagenet'

3. What will be the output shape of the following code snippet using the ResNet50 model in TensorFlow?

import tensorflow as tf
model = tf.keras.applications.ResNet50(weights='imagenet')
input_shape = (1, 224, 224, 3)
sample_input = tf.random.uniform(input_shape)
output = model(sample_input)
print(output.shape)

medium

A. (1, 7, 7, 2048)

B. (224, 224, 3)

C. (1, 1000)

D. (1, 100)

Solution

1. Step 1: Understand ResNet50 output for ImageNet weights

   ResNet50 with weights='imagenet' outputs predictions for 1000 classes, so output shape is (batch_size, 1000).

2. Step 2: Check input batch size and output shape

   Input batch size is 1, so output shape is (1, 1000).

3. Final Answer:

   (1, 1000) -> Option C

4. Quick Check:

   ResNet50 output shape = (batch_size, 1000) = A [OK]

Hint: Output shape matches number of ImageNet classes (1000) [OK]

Common Mistakes:

• Confusing output with input shape
• Expecting feature map shape instead of final predictions
• Assuming output has spatial dimensions

4. You tried to load the VGG16 model with this code:

model = tf.keras.applications.VGG16(weights='imagenet', include_top=False, input_shape=(28,28,3))

But got an error. What is the likely cause?

medium

A. Input shape must be at least 32x32 with 3 channels when include_top=False.

B. Weights='imagenet' is not supported for VGG16.

C. include_top=False cannot be used with input_shape argument.

D. VGG16 requires grayscale images, so 3 channels is wrong.

Solution

1. Step 1: Check VGG16 input shape requirements

   When include_top=False, VGG16 requires input images at least 32x32 pixels with 3 color channels.

2. Step 2: Verify input shape given

   Input shape (28,28,3) is smaller than 32x32x3 minimum when include_top=False, causing the error.

3. Step 3: Identify common mistake

   Often error happens if input_shape is smaller than 32x32 or channels not 3. The question implies error due to input shape constraints.

4. Final Answer:

   Input shape must be at least 32x32 with 3 channels when include_top=False. -> Option A

5. Quick Check:

   VGG16 input shape min size = 32x32x3 [OK]

Hint: VGG16 needs input ≥32x32x3 if include_top=False [OK]

Common Mistakes:

• Thinking weights='imagenet' is unsupported
• Believing include_top=False forbids input_shape
• Assuming VGG16 uses grayscale images

5. You want to build a fast image classifier for mobile devices using TensorFlow. Which pre-trained model should you choose to balance speed and accuracy?

hard

A. MobileNet, because it is designed for lightweight and fast inference.

B. ResNet50, because it uses residual connections for better training.

C. InceptionV3, because it has the most complex architecture.

D. VGG19, because it has the deepest layers and highest accuracy.

Solution

1. Step 1: Understand model design goals

   VGG19 is large and slow, ResNet50 is accurate but heavier, MobileNet is optimized for mobile with smaller size and speed.

2. Step 2: Match model to mobile device needs

   Mobile devices need fast, lightweight models, so MobileNet fits best for balancing speed and accuracy.

3. Final Answer:

   MobileNet, because it is designed for lightweight and fast inference. -> Option A

4. Quick Check:

   MobileNet = lightweight + fast for mobile [OK]

Hint: MobileNet is made for mobile speed and size [OK]

Common Mistakes:

• Choosing VGG19 for mobile despite size
• Assuming ResNet50 is fastest
• Picking InceptionV3 without speed consideration