Generator and discriminator in PyTorch - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

Output of a simple Generator forward pass

Given the following PyTorch Generator model and input noise tensor, what is the shape of the output tensor after the forward pass?

PyTorch

import torch
import torch.nn as nn

class Generator(nn.Module):
    def __init__(self):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(100, 256),
            nn.ReLU(),
            nn.Linear(256, 784),
            nn.Tanh()
        )

    def forward(self, x):
        return self.net(x)

G = Generator()
noise = torch.randn(16, 100)
output = G(noise)
output_shape = output.shape

A (16, 784)

B (100, 256)

C (16, 256)

D (784, 16)

💡 Hint

The Generator transforms noise vectors of size 100 into images flattened to 784 pixels.

The input noise has batch size 16 and 100 features. The Generator outputs a tensor with shape (batch_size, 784) because the last Linear layer outputs 784 features per sample.

❓ Model Choice

intermediate

Choosing the correct Discriminator output layer

Which of the following PyTorch Discriminator output layers is correct for producing a single probability score for real/fake classification?

A nn.Linear(512, 2) followed by nn.LogSoftmax(dim=1)

B nn.Linear(512, 10) followed by nn.Softmax(dim=1)

C nn.Linear(512, 1) with no activation

D nn.Linear(512, 1) followed by nn.Sigmoid()

💡 Hint

The Discriminator outputs a probability between 0 and 1 for real or fake.

The Discriminator should output a single value per input, representing the probability of being real. Using nn.Sigmoid() converts the output to a probability between 0 and 1.

❓ Hyperparameter

advanced

Effect of learning rate on GAN training stability

In training a GAN, what is the most likely effect of setting the learning rate too high for both Generator and Discriminator?

A Training becomes unstable with oscillations and mode collapse

B Training converges faster and produces better images

C The Discriminator ignores fake images and outputs constant values

D The Generator stops learning but the Discriminator improves steadily

💡 Hint

High learning rates can cause the model weights to jump too much during updates.

Too high learning rates cause unstable training in GANs, leading to oscillations and mode collapse where the Generator produces limited variety.

❓ Metrics

advanced

Choosing the right loss function for GAN Discriminator

Which loss function is commonly used for the Discriminator in a vanilla GAN to distinguish real from fake images?

A Hinge Loss

B Mean Squared Error Loss

C Binary Cross Entropy Loss

D Categorical Cross Entropy Loss

💡 Hint

The Discriminator outputs a probability for two classes: real or fake.

Binary Cross Entropy Loss is used for binary classification tasks like real vs fake in vanilla GANs.

🔧 Debug

expert

Identifying the cause of NaN loss in GAN training

During GAN training, the Generator loss suddenly becomes NaN. Which of the following is the most likely cause?

A The batch size is too large causing memory overflow

B The Discriminator outputs values outside the range [0,1] causing log(0) in loss

C The Generator uses ReLU activation in the output layer instead of Tanh

D The optimizer learning rate is set to zero

💡 Hint

Check if the Discriminator output is properly squashed before computing log loss.

If the Discriminator outputs values outside [0,1], the log in Binary Cross Entropy can be log(0) causing NaN loss.