6.6.1 Generative Models Roadmap: Sample, Decode, Review

Generative models create new samples instead of only predicting labels. The practical loop is: sample a latent code, decode it, review the output, and compare versions.

Look at the Generation Flow First

Concept	First meaning
latent vector	compact hidden input used for generation
decoder / generator	turns latent code into an output
discriminator	judges real vs generated in GANs
VAE	learns a smoother latent space
review	generated output still needs human and metric checks

Run One Tiny Decoder

Create generative_first_loop.py and run it after installing torch.

import torch

torch.manual_seed(0)
latent = torch.randn(2, 4)
decoder = torch.nn.Sequential(torch.nn.Linear(4, 6), torch.nn.Tanh())
generated = decoder(latent)

print("latent_shape:", tuple(latent.shape))
print("generated_shape:", tuple(generated.shape))
print("value_range:", round(generated.min().item(), 3), round(generated.max().item(), 3))

Expected output:

latent_shape: (2, 4)
generated_shape: (2, 6)
value_range: -0.863 0.695

This is not a real generator yet. It shows the core shape idea: small latent vectors can be decoded into larger outputs.

Learn in This Order

Order	Read	What to focus on
1	6.6.2 GAN	generator, discriminator, adversarial balance
2	6.6.3 VAE	encoder, decoder, latent space

Evidence to Keep

Keep one generation review note:

Latent Shape

what compact code enters the generator/decoder

Output Shape

what sample-like object comes out

Quality Check

does it look plausible or reconstruct well?

Diversity Check

are outputs varied, or collapsing?

Trust Rule

generated output always needs review

Pass Check

You pass this roadmap when you can explain the difference between predicting a label and generating a sample, and describe why generated outputs need review rather than blind trust.

Check reasoning and explanation

1. A passing answer connects tensors, model layers, loss, backward(), and optimizer updates into one training loop.
2. The evidence should include a runnable mini experiment, tensor-shape checks, and a loss or validation curve you can explain.
3. A good self-check names one failure mode such as shape mismatch, no loss decrease, overfitting, data leakage, or using Attention/Transformer words without explaining the data flow.