CS 294-158 Deep Unsupervised Learning, Homework 2, Spring 2020 1
Homework 2: Flow Models
Deliverable: This PDF write-up by Tuesday February 25th, 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: 2D Data
(a) [15pt] Autoregressive Flow
Final test loss for dataset 1: 1.3092 nats / dim
(a) Training curve (b) Learned distribution (c) Latent Space
Figure 1: Results for Dataset 1 Final test loss for dataset 2: 0.0000 nats / dim
figures/q1_a_dset2_train_plfoitg.uprnegs/q1_a_dset2_densitiefsi.gpunrges/q1_a_dset2_latents.png
(a) Training curve (b) Learned distribution (c) Latent Space
Figure 2: Results for Dataset 2
CS 294-158 Deep Unsupervised Learning, Homework 2, Spring 2020
(b) [15pt] RealNVP
Final test loss for dataset 1: 2.0586 nats / dim
(a) Training curve (b) Learned distribution
Figure 3: Results for Dataset 1 Final test loss for dataset 2: 0.0000 nats / dim
2
(c) Latent Space
figures/q1_b_dset2_train_plfoitg.uprnegs/q1_b_dset2_densitiefsi.gpunrges/q1_b_dset2_latents.png
(a) Training curve (b) Learned distribution (c) Latent Space
Figure 4: Results for Dataset 2
CS 294-158 Deep Unsupervised Learning, Homework 2, Spring 2020 3 Question 2: Autoregressive Flows for Images [20pt]
Final test loss: 0.2258 nats / dim
figures/q2_train_plot.png
(a) Training curve
figures/q2_flooredsamples.png
(b) Samples
(c) Samples, removing noise from dequantization
CS 294-158 Deep Unsupervised Learning, Homework 2, Spring 2020 4
Question 3: RealNVP on Higher Dimensions (a) [40pt] RealNVP
Final test loss: 0.55 bits / dim
figures/q3_a_train_plot.png
(a) Training curve
(b) Samples
(c) Interpolations
CS 294-158 Deep Unsupervised Learning, Homework 2, Spring 2020 5 (b) [10pt] Bad Masking
Final test loss: 0.0000 nats / dim
figures/q3_b_train_plot.png
(a) Training curve
figures/q3_b_interpolations.png
(b) Samples
(c) Interpolations
CS 294-158 Deep Unsupervised Learning, Homework 2, Spring 2020 6
Bonus Questions (Optional)
1. [10pt] Multiscale RealNVP
Final test loss: 0.0000 nats / dim
figures/q4_a_train_plot.png
(a) Training curve
figures/q4_a_samples.png
figures/q4_a_interpolations.png
(b) Samples
(c) Interpolations
CS 294-158 Deep Unsupervised Learning, Homework 2, Spring 2020 7 2. [5pt] Glow
Final test loss: 0.0000 nats / dim
figures/q4_b_train_plot.png
(a) Training curve
figures/q4_b_samples.png
figures/q4_b_interpolations.png
(b) Samples
(c) Interpolations