Data augmentation in pipeline in TensorFlow - Model Metrics & Evaluation

Metrics & Evaluation - Data augmentation in pipeline

Which metric matters for Data Augmentation in Pipeline and WHY

Data augmentation helps the model see more varied examples by changing training data slightly. This usually improves validation accuracy and generalization. So, accuracy on unseen data is key to check if augmentation helps.

Also, watch loss during training and validation. If loss on validation decreases and accuracy increases, augmentation is working well.

Confusion Matrix Example

Imagine a model classifying images into cats and dogs. After augmentation, the confusion matrix on validation might look like this:

      | Predicted Cat | Predicted Dog |
      |---------------|---------------|
      | True Cat: 45  | 5             |
      | True Dog: 7   | 43            |
    

Here, total samples = 45 + 5 + 7 + 43 = 100.

Precision for Cat = TP / (TP + FP) = 45 / (45 + 7) = 0.865

Recall for Cat = TP / (TP + FN) = 45 / (45 + 5) = 0.9

Precision vs Recall Tradeoff with Data Augmentation

Data augmentation can help balance precision and recall by making the model robust to variations.

For example, in medical image classification, high recall is critical to catch all positive cases. Augmentation can help the model recognize more varied positive examples, improving recall.

In contrast, for spam detection, high precision is important to avoid marking good emails as spam. Augmentation should be done carefully to not confuse the model.

Good vs Bad Metric Values for Data Augmentation

Good: Validation accuracy improves or stays stable, validation loss decreases, precision and recall both improve or remain balanced.

Bad: Validation accuracy drops, validation loss increases, or precision and recall become very unbalanced (e.g., very high precision but very low recall).

Also watch for overfitting signs: training accuracy very high but validation accuracy low.

Common Pitfalls in Metrics with Data Augmentation

• Accuracy Paradox: Accuracy might look good if data is imbalanced. Always check precision and recall.
• Data Leakage: Augmenting validation or test data leaks training info and inflates metrics.
• Overfitting: Augmentation that is too weak or too strong can cause overfitting or underfitting, seen in diverging training and validation metrics.
• Ignoring Validation Metrics: Only looking at training metrics can mislead about real performance.

Self Check

Your model with data augmentation has 98% accuracy but 12% recall on the positive class (e.g., fraud). Is it good for production?

Answer: No. Despite high accuracy, the model misses most positive cases (low recall). For fraud detection, catching fraud (high recall) is critical. This model would fail in real use.

Key Result

Data augmentation improves model generalization, best measured by validation accuracy and balanced precision-recall.

Practice

1. What is the main purpose of data augmentation in a TensorFlow training pipeline?

easy

A. To speed up the training process by skipping some images

B. To reduce the size of the training dataset

C. To create more varied training data by randomly changing original images

D. To convert images into grayscale only

Solution

1. Step 1: Understand data augmentation concept

   Data augmentation creates new training images by applying random changes like flips or rotations to original images.

2. Step 2: Identify the purpose in training pipeline

   This helps the model see more varied examples, improving learning and reducing overfitting.

3. Final Answer:

   To create more varied training data by randomly changing original images -> Option C

4. Quick Check:

   Data augmentation = varied training data [OK]

Hint: Augmentation adds variety to training images [OK]

Common Mistakes:

• Thinking augmentation reduces dataset size
• Believing augmentation speeds training by skipping data
• Assuming augmentation only converts images to grayscale

2. Which of the following is the correct way to add a random flip augmentation layer in a TensorFlow Sequential pipeline?

easy

A. tf.keras.Sequential([tf.keras.layers.RandomFlip('horizontal')])

B. tf.keras.Sequential([tf.keras.layers.FlipRandom('horizontal')])

C. tf.keras.Sequential([tf.keras.layers.RandomFlip(mode='vertical')])

D. tf.keras.Sequential([tf.keras.layers.RandomFlip('diagonal')])

Solution

1. Step 1: Recall TensorFlow augmentation syntax

   The correct layer is RandomFlip with argument 'horizontal' or 'vertical' as a string.

2. Step 2: Check each option

   tf.keras.Sequential([tf.keras.layers.RandomFlip('horizontal')]) uses correct class and argument. tf.keras.Sequential([tf.keras.layers.FlipRandom('horizontal')]) uses wrong class name. tf.keras.Sequential([tf.keras.layers.RandomFlip(mode='vertical')]) uses keyword argument 'mode' which is invalid. tf.keras.Sequential([tf.keras.layers.RandomFlip('diagonal')]) uses unsupported flip mode 'diagonal'.

3. Final Answer:

   tf.keras.Sequential([tf.keras.layers.RandomFlip('horizontal')]) -> Option A

4. Quick Check:

   Correct layer and argument = tf.keras.Sequential([tf.keras.layers.RandomFlip('horizontal')]) [OK]

Hint: Use RandomFlip('horizontal') exactly as named [OK]

Common Mistakes:

• Using wrong layer class name
• Passing arguments with wrong keywords
• Using unsupported flip modes

3. Given the following TensorFlow code snippet, what will be the output shape of the augmented images?

import tensorflow as tf
aug = tf.keras.Sequential([
  tf.keras.layers.RandomFlip('horizontal'),
  tf.keras.layers.RandomRotation(0.1)
])
input_image = tf.random.uniform([1, 128, 128, 3])
output_image = aug(input_image)
print(output_image.shape)

medium

A. (1, 128, 128, 3)

B. (128, 128, 3)

C. (1, 256, 256, 3)

D. (1, 128, 128)

Solution

1. Step 1: Understand input and augmentation layers

   Input shape is (1, 128, 128, 3) meaning batch size 1, 128x128 image with 3 color channels. RandomFlip and RandomRotation do not change image size.

2. Step 2: Check output shape after augmentation

   Augmentation layers keep the shape same, so output shape remains (1, 128, 128, 3).

3. Final Answer:

   (1, 128, 128, 3) -> Option A

4. Quick Check:

   Augmentation keeps shape = (1, 128, 128, 3) [OK]

Hint: Augmentation layers keep input shape unchanged [OK]

Common Mistakes:

• Assuming rotation changes image size
• Ignoring batch dimension in output
• Dropping color channels

4. Identify the error in this TensorFlow data augmentation pipeline code:

import tensorflow as tf
aug = tf.keras.Sequential([
  tf.keras.layers.RandomFlip('horizontal'),
  tf.keras.layers.RandomRotation(0.2, 0.3)
])

medium

A. Missing input shape in Sequential

B. RandomFlip does not accept 'horizontal' as argument

C. Sequential cannot contain augmentation layers

D. RandomRotation requires a single float or tuple, not two separate floats

Solution

1. Step 1: Check RandomRotation layer arguments

   RandomRotation expects either a single float or a tuple like (min_factor, max_factor). Passing two separate floats is invalid.

2. Step 2: Verify other parts

   RandomFlip('horizontal') is valid. Sequential can contain augmentation layers. Input shape is optional here.

3. Final Answer:

   RandomRotation requires a single float or tuple, not two separate floats -> Option D

4. Quick Check:

   RandomRotation argument format error = RandomRotation requires a single float or tuple, not two separate floats [OK]

Hint: RandomRotation needs one float or tuple, not two floats [OK]

Common Mistakes:

• Passing multiple floats instead of tuple to RandomRotation
• Thinking RandomFlip argument is invalid
• Believing Sequential can't hold augmentation layers

5. You want to build a TensorFlow data augmentation pipeline that randomly flips images horizontally, rotates them by up to 20%, and zooms in or out by up to 10%. Which of the following code snippets correctly implements this pipeline?

hard

A. tf.keras.Sequential([
tf.keras.layers.RandomFlip('horizontal'),
tf.keras.layers.RandomRotation(0.2),
tf.keras.layers.RandomZoom((0.1, 0.2))
])

B. tf.keras.Sequential([
tf.keras.layers.RandomFlip('horizontal'),
tf.keras.layers.RandomRotation(0.2),
tf.keras.layers.RandomZoom(0.1)
])

C. tf.keras.Sequential([
tf.keras.layers.RandomFlip('horizontal'),
tf.keras.layers.RandomRotation(0.02),
tf.keras.layers.RandomZoom(10)
])

D. tf.keras.Sequential([
tf.keras.layers.RandomFlip('vertical'),
tf.keras.layers.RandomRotation(20),
tf.keras.layers.RandomZoom((0.1, 0.1))
])

Solution

1. Step 1: Check flip and rotation parameters

   RandomFlip('horizontal') is correct. RandomRotation expects a float fraction (0.2 means 20%).

2. Step 2: Check zoom parameters

   RandomZoom(0.1) means zoom in/out by 10%. tf.keras.Sequential([ tf.keras.layers.RandomFlip('horizontal'), tf.keras.layers.RandomRotation(0.2), tf.keras.layers.RandomZoom((0.1, 0.2)) ]) uses zoom (0.1, 0.2) which is uneven zoom, not requested.

3. Final Answer:

   tf.keras.Sequential([ tf.keras.layers.RandomFlip('horizontal'), tf.keras.layers.RandomRotation(0.2), tf.keras.layers.RandomZoom(0.1) ]) -> Option B

4. Quick Check:

   Correct flip, rotation fraction, and zoom float = tf.keras.Sequential([ tf.keras.layers.RandomFlip('horizontal'), tf.keras.layers.RandomRotation(0.2), tf.keras.layers.RandomZoom(0.1) ]) [OK]

Hint: Use fractions for rotation and single float for zoom [OK]

Common Mistakes:

• Using degrees instead of fraction for rotation
• Passing large numbers to zoom
• Choosing wrong flip direction