CS 294-158 Deep Unsupervised Learning, Homework 4, Spring 2020 1
Homework 4: GAN Models
Deliverable: This PDF write-up by Tuesday March 31st, 23:59pm. Your PDF should be generated by simply replacing the placeholder images of this LaTeX document with the appropriate solution images that will be generated automatically when solving each question. The solution images are automatically generated and saved using the accompanying IPython notebook. Your PDF is to be submitted into Gradescope. This PDF already contains a few solution images. These images will allow you to check your own solution to ensure correctness.
Question 1: 1D Data
(a) [10pt] Minimax GAN Objective
figures/q1a_epoch1.png
figures/q1a_final.png
(a) Samples at epoch 1
(b) Final samples
figures/q1a_losses.png
(c) GAN loss curve
CS 294-158 Deep Unsupervised Learning, Homework 4, Spring 2020 2 (b) [10pt] Nonsaturating Objective
figures/q1b_epoch1.png
figures/q1b_final.png
(a) Samples at epoch 1
(b) Final samples
figures/q1b_losses.png
(c) GAN loss curve
CS 294-158 Deep Unsupervised Learning, Homework 4, Spring 2020 3 Question 2: WGAN-GP on CIFAR-10 [35pt]
Final inception score: TODO
Final Fr ́echet inception distance: TODO
(a) Samples
figures/q2_losses.png
(b) Training curve
CS 294-158 Deep Unsupervised Learning, Homework 4, Spring 2020 4 Question 3: Representation Learning with BiGAN on MNIST [45 pt]
Final supervised test accuracy: TODO Final unsupervised test accuracy: TODO
(a) Samples
figures/q3_gan_losses.png
figures/q3_supervised_losses.png
(b) BiGAN train curve (c) Supervised learning curve
(d) Reconstructions
CS 294-158 Deep Unsupervised Learning, Homework 4, Spring 2020 5 Bonus Questions (Optional)
1. [20pt] CycleGAN
figures/q4_mnist.png
(a) MNIST: original images, translations, and recon- structions
figures/q4_colored_mnist.png
(b) Colored MNIST: original images, translations, and reconstructions