Image generation basics in PyTorch - Model Metrics & Evaluation

Metrics & Evaluation - Image generation basics

Which metric matters for Image Generation and WHY

For image generation, we want to know how close the generated images are to real ones. Common metrics are:

• Inception Score (IS): Measures if images look clear and varied.
• Fréchet Inception Distance (FID): Compares statistics of generated and real images; lower is better.
• Perceptual Loss: Measures how similar images are in features humans notice.

These metrics help us understand if the model creates realistic and diverse images.

Confusion Matrix or Equivalent Visualization

Image generation does not use a confusion matrix like classification. Instead, we compare distributions of features from real and generated images.

Real Images Features:  Mean = μ_r, Covariance = Σ_r
Generated Images Features: Mean = μ_g, Covariance = Σ_g

FID = ||μ_r - μ_g||^2 + Trace(Σ_r + Σ_g - 2(Σ_r Σ_g)^(1/2))

Lower FID means generated images are closer to real ones.
    

Tradeoff: Diversity vs Quality

In image generation, we want images that are both diverse and high quality.

• High Quality, Low Diversity: Images look very clear but all look similar (mode collapse).
• High Diversity, Low Quality: Images vary a lot but look blurry or unrealistic.

Metrics like Inception Score reward both diversity and quality, while FID focuses on similarity to real images.

Good vs Bad Metric Values for Image Generation

• Inception Score (IS): Good: > 8 (varied, clear images). Bad: < 3 (blurry, repetitive images).
• FID: Good: < 50 (close to real images). Bad: > 100 (poor quality or unrealistic images).

Good metrics mean the model generates images that look real and different from each other.

Common Pitfalls in Image Generation Metrics

• Mode Collapse: Model generates few types of images, hurting diversity but may show good quality.
• Overfitting: Model memorizes training images, leading to low FID but poor generalization.
• Misleading Scores: IS can be high if images are sharp but unrealistic; always check images visually.
• Data Leakage: Using test images in training can falsely improve metrics.

Self-Check Question

Your image generation model has an Inception Score of 9 but a FID of 120. Is it good?

Answer: No. The high IS means images look clear and varied, but the very high FID means they are far from real images in distribution. The model likely generates unrealistic images despite variety. You should improve realism.

Key Result

Image generation quality is best judged by FID (lower is better) and Inception Score (higher is better) balancing realism and diversity.

Practice

1. What is the main role of a generator in image generation using PyTorch?

easy

A. To create new images from random noise

B. To classify images into categories

C. To reduce the size of images

D. To label images with tags

Solution

1. Step 1: Understand the generator's purpose

   A generator is a neural network designed to produce new data, often images, from random noise inputs.

2. Step 2: Differentiate from other network roles

   Unlike classifiers or reducers, generators focus on creating new images, not labeling or compressing them.

3. Final Answer:

   To create new images from random noise -> Option A

4. Quick Check:

   Generator = create images from noise [OK]

Hint: Generators make images from noise, not classify or compress [OK]

Common Mistakes:

• Confusing generator with classifier
• Thinking generator reduces image size
• Assuming generator labels images

2. Which PyTorch code snippet correctly initializes a noise tensor for image generation?

easy

A. noise = torch.randn(16, 100)

B. noise = torch.zeros(16, 100)

C. noise = torch.ones(16, 100)

D. noise = torch.randint(0, 1, (16, 100))

Solution

1. Step 1: Identify noise initialization for generators

   Generators typically use random noise from a normal distribution, which torch.randn provides.

2. Step 2: Check other options

   Zeros or ones do not provide randomness; randint with 0 to 1 range is invalid as upper bound is exclusive.

3. Final Answer:

   noise = torch.randn(16, 100) -> Option A

4. Quick Check:

   Random normal noise = torch.randn [OK]

Hint: Use torch.randn for random normal noise input [OK]

Common Mistakes:

• Using zeros or ones instead of random noise
• Incorrect use of torch.randint with invalid range
• Confusing noise shape dimensions

3. Given this PyTorch generator output code, what is the shape of the generated images?

import torch
noise = torch.randn(8, 100)
generator = torch.nn.Sequential(
    torch.nn.Linear(100, 256),
    torch.nn.ReLU(),
    torch.nn.Linear(256, 784),
    torch.nn.Tanh()
)
fake_images = generator(noise)

medium

A. (8, 28, 28)

B. (100, 256)

C. (8, 784)

D. (256, 784)

Solution

1. Step 1: Trace the generator output shape

   The last Linear layer outputs 784 features per input, so for batch size 8, output shape is (8, 784).

2. Step 2: Understand image reshaping

   784 corresponds to 28x28 pixels flattened; reshaping to (8, 28, 28) is done later, not shown here.

3. Final Answer:

   (8, 784) -> Option C

4. Quick Check:

   Output shape = (batch_size, 784) [OK]

Hint: Output shape matches last Linear layer output size [OK]

Common Mistakes:

• Assuming output is already 2D image shape
• Confusing input noise shape with output
• Mixing intermediate layer shapes

4. Identify the error in this PyTorch generator code snippet:

import torch
class Generator(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.fc = torch.nn.Linear(100, 784)
    def forward(self, x):
        x = self.fc(x)
        x = torch.sigmoid(x)
        return x
noise = torch.randn(16, 100)
gen = Generator()
output = gen(noise).reshape(16, 28, 28)

medium

A. Missing batch normalization layers

B. Not calling super().__init__() in constructor

C. Incorrect reshape dimensions for output

D. Using sigmoid activation instead of tanh

Solution

1. Step 1: Check activation function for image generation

   Image generation often uses tanh to scale outputs between -1 and 1, matching normalized image data.

2. Step 2: Analyze impact of sigmoid

   Sigmoid outputs between 0 and 1, which may mismatch expected scaling and reduce image quality.

3. Final Answer:

   Using sigmoid activation instead of tanh -> Option D

4. Quick Check:

   Activation should be tanh, not sigmoid [OK]

Hint: Use tanh for output scaling in image generation [OK]

Common Mistakes:

• Thinking reshape dimensions are wrong
• Believing batch norm is mandatory here
• Forgetting to call super().__init__()

5. You want to generate 64 images of size 32x32 with pixel values scaled between -1 and 1 using a PyTorch generator. Which output layer and activation setup is best?

hard

A. Linear layer with 1024 outputs followed by sigmoid activation

B. Linear layer with 1024 outputs followed by tanh activation

C. Linear layer with 1024 outputs followed by ReLU activation

D. Linear layer with 1024 outputs followed by softmax activation

Solution

1. Step 1: Calculate output size for 32x32 images

   32x32 = 1024 pixels, so output layer must have 1024 units to represent each pixel.

2. Step 2: Choose activation for scaling pixel values

   Tanh activation scales outputs between -1 and 1, matching the desired pixel value range.

3. Step 3: Evaluate other activations

   Sigmoid outputs 0 to 1, ReLU outputs 0 to infinity, softmax is for classification, so they are unsuitable here.

4. Final Answer:

   Linear layer with 1024 outputs followed by tanh activation -> Option B

5. Quick Check:

   Output size 1024 + tanh for [-1,1] scaling [OK]

Hint: Use tanh activation for output scaled between -1 and 1 [OK]

Common Mistakes:

• Using sigmoid instead of tanh for output scaling
• Choosing softmax for image pixel output
• Ignoring correct output size for image dimensions