CS 294-158 Deep Unsupervised Learning, Homework 3, Spring 2020 1
Homework 3: Latent Variable Models
Deliverable: This PDF write-up by Tuesday March 10th, 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: VAEs on 2D Data [20pt]
(a) [10pt] Data from a Full Covariance Gaussian
Final Full -ELBO: 4.4388, Recon Loss: 2.7630, KL Loss: 1.6758 (Dataset 1)
(a) Training curve (b) Samples with Noise (c) Samples without Noise
Figure 1: Results for Dataset 1
Final Full -ELBO: FILL, Recon Loss: FILL, KL Loss: FILL (Dataset 2)
(a) Training curve (b) Samples with Noise (c) Samples without Noise
Figure 2: Results for Dataset 2
CS 294-158 Deep Unsupervised Learning, Homework 3, Spring 2020 2 (b) [10pt] Data from a Diagonal Gaussian
Final Full -ELBO: 4.4213, Recon Loss: 4.4094, KL Loss: 0.0119 (Dataset 1)
(a) Training curve (b) Samples with Noise (c) Samples without Noise
Figure 3: Results for Dataset 1
Final Full -ELBO: FILL, Recon Loss: FILL, KL Loss: FILL (Dataset 2)
(a) Training curve (b) Samples with Noise (c) Samples without Noise
Figure 4: Results for Dataset 2
Answer: Your answer to the reflection portion of part (b) reflection here (replace this text)
CS 294-158 Deep Unsupervised Learning, Homework 3, Spring 2020 3
Question 2: VAEs on Images [40pt] (a) [20pt] VAE
Final Full -ELBO: 104.0417, Recon Loss: 79.3798, KL Loss: 24.6620 (Dataset 1)
(a) Training Curve (b) Samples
(c) Reconstructions (d) Interpolations
Figure 5: Results for Dataset 1
CS 294-158 Deep Unsupervised Learning, Homework 3, Spring 2020 4 Final Full -ELBO: FILL, Recon Loss: FILL, KL Loss: FILL (Dataset 2)
(a) Training Curve (b) Samples
(c) Reconstructions (d) Interpolations
Figure 6: Results for Dataset 2
CS 294-158 Deep Unsupervised Learning, Homework 3, Spring 2020 5 (b) [20pt] VAE with AF Prior
Final Full -ELBO: 102.5659, Recon Loss: 80.2548, KL Loss: 22.3111 (Dataset 1)
(a) Training Curve (b) Samples
(c) Reconstructions (d) Interpolations
Figure 7: Results for Dataset 1
CS 294-158 Deep Unsupervised Learning, Homework 3, Spring 2020 6 Final Full -ELBO: FILL, Recon Loss: FILL, KL Loss: FILL (Dataset 2)
(a) Training Curve (b) Samples
(c) Reconstructions (d) Interpolations
Figure 8: Results for Dataset 2
CS 294-158 Deep Unsupervised Learning, Homework 3, Spring 2020 7 Question 3: VQ-VAE [40pt]
Final VQ-VAE Test Loss: 0.0286, PixelCNN Prior Test Los: 1.9440 (Dataset 1)
(a) VQ-VAE Training Curve (b) PixelCNN Prior Training Curve
(c) Samples
(d) Reconstructions
Figure 9: Results for Dataset 1
CS 294-158 Deep Unsupervised Learning, Homework 3, Spring 2020 8 Final VQ-VAE Test Loss: FILL, PixelCNN Prior Test Los: FILL (Dataset 2)
(a) VQ-VAE Training Curve (b) PixelCNN Prior Training Curve
(c) Samples
(d) Reconstructions
Figure 10: Results for Dataset 1
CS 294-158 Deep Unsupervised Learning, Homework 3, Spring 2020 9
Question 4: Bonus [10pt]
1. [5pt] Improving VQ-VAE Results
Final VQ-VAE Test Loss: FILL, PixelCNN Prior Test Los: FILL
(a) VQ-VAE Training Curve (b) PixelCNN Prior Training Curve
(c) Samples
(d) Reconstructions
Figure 11: Results for CIFAR10
CS 294-158 Deep Unsupervised Learning, Homework 3, Spring 2020 10 2. [5pt] PixelVAE
Final Full -ELBO: FILL, Recon Loss: FILL, KL Loss: FILL
(a) Training curve (b) Samples (c) Reconstructions
Figure 12: Results for MNIST