D-DL4CV-Lec19c-Autoregressive_Model

Autoregressive Model

Concept

D-DL4CV-Lec19b-Explicit_Density_Estimation

We want to optimize learnable parameters in the network so that it has highest probability generating pixel patterns similar to that we feed in as training data

PixelRNN

D-DL4CV-Lec19ca-PixelRNN

PixelCNN

D-DL4CV-Lec19cb-PixelCNN

PixelCNN trains faster than PixelRNN

1. PixelCNN depends on neighbor’s pixel value
2. PixelRNN depends on neighbor’s hidden state

Hence, we need to wait for the previous hidden state be computed before computing current hidden state

Pros and Cons of PixelRNN & PixelCNN

Pros:

• Can explicitly compute likelihood $p(x)$
• Explicit likelihood is easy to do evaluation on the model

Cons:

• In test process, generate pixel is sequential process ⇒ slow