Why generative models create new data in TensorFlow - Why Metrics Matter

Metrics & Evaluation - Why generative models create new data

Which metric matters and WHY

For generative models that create new data, we care about how realistic and diverse the generated data is. Common metrics include Inception Score (IS) and Fréchet Inception Distance (FID). These metrics measure how close the generated data is to real data in terms of quality and variety. We want metrics that tell us if the model is making believable new examples, not just copying old ones.

Confusion matrix or equivalent visualization

Generative models don't use a confusion matrix like classifiers. Instead, we compare distributions of real and generated data. For example, FID measures the distance between feature distributions of real and generated images. A lower FID means generated data is closer to real data.

Real data distribution:  ●●●●●●●●●●
Generated data distribution:  ●●●●●●●●●●
FID = distance between these distributions (lower is better)

Precision vs Recall tradeoff with examples

In generative models, precision means how realistic the generated data is (no fake-looking samples). Recall means how diverse the generated data is (covers many types of real data).

Example: A model that only generates one type of flower perfectly has high precision but low recall (not diverse). A model that generates many types but some look fake has high recall but low precision.

Good generative models balance precision and recall to create many realistic and varied new data points.

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

For FID, a good score is low (e.g., below 50), meaning generated data is close to real data. A bad score is high (e.g., above 100), meaning generated data looks very different or poor quality.

For Inception Score, a good value is high (e.g., above 5 for images), showing clear and varied outputs. A bad score is low (close to 1), meaning outputs are blurry or repetitive.

Common pitfalls in metrics

• Mode collapse: Model generates limited types of data, hurting recall but may still have good precision.
• Overfitting: Model memorizes training data, generating copies instead of new data.
• Misleading scores: Metrics like IS can be fooled by models that generate sharp but unrealistic images.
• Data leakage: Using test data in training can inflate metric scores falsely.

Self-check question

Your generative model has a low FID score but generates very similar images repeatedly. Is it good? Why or why not?

Answer: No, because low FID alone doesn't guarantee diversity. The model likely suffers from mode collapse, producing limited new data. Good generative models must create both realistic and diverse outputs.

Key Result

Generative models are evaluated by how realistic and diverse their new data is, using metrics like FID and Inception Score.

Practice

1. Why do generative models create new data in machine learning?

easy

A. To only classify existing data into categories

B. To delete old data from the dataset

C. To learn patterns from existing data and generate similar new data

D. To speed up data collection from the internet

Solution

1. Step 1: Understand generative model purpose

   Generative models learn from existing data patterns to create new, similar data.

2. Step 2: Eliminate incorrect options

   Deleting data, classifying data, or speeding data collection are not goals of generative models.

3. Final Answer:

   To learn patterns from existing data and generate similar new data -> Option C

4. Quick Check:

   Generative models create new data by learning patterns = C [OK]

Hint: Generative models imitate data patterns to create new data [OK]

Common Mistakes:

• Confusing generative models with classifiers
• Thinking generative models delete data
• Assuming generative models only speed up data collection

2. Which TensorFlow code snippet correctly defines a simple generative model layer?

easy

A. tf.keras.layers.Dense(units=128, activation='relu')

B. tf.keras.layers.Conv2D(filters=32, kernel_size=3)

C. tf.keras.layers.Dropout(rate=0.5)

D. tf.keras.layers.MaxPooling2D(pool_size=2)

Solution

1. Step 1: Identify generative model layer type

   Dense layers with activation functions are commonly used in generative models to create outputs.

2. Step 2: Check other layers' roles

   Conv2D is for feature extraction, Dropout for regularization, MaxPooling2D for downsampling, not direct generation.

3. Final Answer:

   tf.keras.layers.Dense(units=128, activation='relu') -> Option A

4. Quick Check:

   Dense layer with activation is used in generative models = A [OK]

Hint: Dense layers with activation create new data features [OK]

Common Mistakes:

• Choosing convolution layers as generative output
• Confusing dropout with data generation
• Selecting pooling layers for generation

3. Given this TensorFlow code snippet for a simple generative model, what is the shape of the output tensor?

import tensorflow as tf
model = tf.keras.Sequential([
  tf.keras.layers.Dense(64, activation='relu', input_shape=(100,)),
  tf.keras.layers.Dense(28*28, activation='sigmoid'),
  tf.keras.layers.Reshape((28, 28))
])
output = model(tf.random.normal([1, 100]))
print(output.shape)

medium

A. (28, 28)

B. (100, 28, 28)

C. (1, 784)

D. (1, 28, 28)

Solution

1. Step 1: Analyze model layers and input

   Input shape is (100,), first Dense outputs 64 units, second Dense outputs 28*28=784 units, then reshaped to (28, 28).

2. Step 2: Check batch size and output shape

   Input batch size is 1, so output shape after reshape is (1, 28, 28).

3. Final Answer:

   (1, 28, 28) -> Option D

4. Quick Check:

   Output shape matches batch and reshape = (1, 28, 28) [OK]

Hint: Remember batch size stays in output shape [OK]

Common Mistakes:

• Ignoring batch size dimension
• Confusing flattened and reshaped output
• Assuming output shape is (28, 28) without batch

4. This TensorFlow code tries to generate new images but throws an error. What is the problem?

import tensorflow as tf
model = tf.keras.Sequential([
  tf.keras.layers.Dense(64, activation='relu', input_shape=(100,)),
  tf.keras.layers.Dense(28*28, activation='sigmoid'),
  tf.keras.layers.Reshape((28, 28))
])
noise = tf.random.normal([100, 28])
output = model(noise)
print(output.shape)

medium

A. Input batch size should be first dimension, so shape should be (batch_size, 100), not (100, 28)

B. Activation function 'sigmoid' is not allowed in generative models

C. Reshape layer cannot reshape to (28, 28)

D. Dense layers require integer inputs, not float tensors

Solution

1. Step 1: Analyze model input expectation

   The model has input_shape=(100,), so expects inputs of shape (batch_size, 100).

2. Step 2: Check provided input shape

   noise = tf.random.normal([100, 28]) has shape (100, 28), where the feature dimension is 28 instead of 100, causing a shape mismatch error.

3. Final Answer:

   Input batch size should be first dimension, so shape should be (batch_size, 100), not (100, 28) -> Option A

4. Quick Check:

   Model input_shape=(100,) requires feature dim=100; got 28 [OK]

Hint: Batch size is always first dimension in TensorFlow inputs [OK]

Common Mistakes:

• Mixing up batch size and feature dimensions
• Thinking activation functions cause errors
• Assuming reshape layer is invalid without checking

5. You want to build a generative model in TensorFlow that creates new handwritten digits similar to MNIST data. Which approach best explains why the model can create new images?

hard

A. The model memorizes all MNIST images and randomly picks one to output

B. The model learns the distribution of pixel patterns from MNIST and samples new points from this learned distribution

C. The model uses a fixed noise input and outputs the same image every time

D. The model only classifies digits and does not generate new images

Solution

1. Step 1: Understand generative model learning

   The model learns the underlying distribution of the training data (pixel patterns) to generate new samples.

2. Step 2: Eliminate incorrect options

   Memorizing images means no new data; fixed noise gives same output; classification does not generate data.

3. Final Answer:

   The model learns the distribution of pixel patterns from MNIST and samples new points from this learned distribution -> Option B

4. Quick Check:

   Generative models sample new data from learned distribution = A [OK]

Hint: Generative models sample new data from learned patterns [OK]

Common Mistakes:

• Thinking generative models memorize data
• Assuming fixed noise means varied output
• Confusing classification with generation