yandex-research / ddpm-segmentation

Label-Efficient Semantic Segmentation with Diffusion Models (ICLR'2022)
https://yandex-research.github.io/ddpm-segmentation/
MIT License
662 stars 60 forks source link

Confusion in time steps mentioned in the paper! #25

Open HOMGH opened 1 year ago

HOMGH commented 1 year ago

Hi, Based on the paper, the later time steps, the more informative feature representations. But how can we see the effect of different time steps using the code? As far as I can see in the feature_extractors.py, in https://github.com/yandex-research/ddpm-segmentation/blob/9a13776bb78d4753df6a561749f3274cd158297a/src/feature_extractors.py#L72 with different values of "t", we would have different noisy images (with different level of noise, based on the t. The larger "t" corresponds to more noise in the image). So how can this "t" work as the time steps in reverse process? Thanks.