search

tsinghua-fib-lab / UniST

Official implementation for "UniST: A Prompt-Empowered Universal Model for Urban Spatio-Temporal Prediction" (KDD 2024)
55 stars 7 forks source link

Question about four mask strategies. #2

Closed packer-c closed 6 days ago

packer-c commented 2 weeks ago

Hi, thank you for your nice work and for sharing the code.

I would like to know about the mask strategies. Are these four strategies working together or individually (['random','causal','frame','tube'])? I saw that in your code "main.py", you set mask_strategy = 'random', which seems to use only one random strategy but not use others. Which one is the best mask strategy effect?

Look forward to your reply. Thanks in advance!

YuanYuan98 commented 2 weeks ago

Thanks for your interest!

To clarify, the parameter mask_strategy determines the strategy used for masking during training. If mask_strategy_random is set to none, then mask_strategy dictates which specific strategy ('random', 'causal', 'frame', 'tube') is applied uniformly across all batches. On the other hand, if mask_strategy_random is set to batch, the masking strategy is randomly selected from the four options for each training batch.

packer-c commented 6 days ago

Clear! Thanks for your kindly reply.