search

lhoyer / DAFormer

[CVPR22] Official Implementation of DAFormer: Improving Network Architectures and Training Strategies for Domain-Adaptive Semantic Segmentation
Other
466 stars 91 forks source link

visuial of t-sne #56

Closed yuheyuan closed 1 year ago

yuheyuan commented 1 year ago

image I think the picture in your paper is so cool. I want to use t-sne to produce the same effiect.But I don't know how to realize it. Could you offer some code or detial about it. Maybe it's too long because you finish this work last year. If you can reply me. Very appreciated

lhoyer commented 1 year ago

We have generated these plots using https://github.com/DmitryUlyanov/Multicore-TSNE. Specifically, we have inferred the bottleneck features of the model, filtered out feature pixels that contain different ground truth classes within the feature pixel region (this can happen as the feature map has a lower resolution as the ground truth semantic segmentation), calculated the TSNE embedding of these features, and colored them according to the ground truth class of that feature pixel,

yuheyuan commented 6 months ago

We have generated these plots using https://github.com/DmitryUlyanov/Multicore-TSNE. Specifically, we have inferred the bottleneck features of the model, filtered out feature pixels that contain different ground truth classes within the feature pixel region (this can happen as the feature map has a lower resolution as the ground truth semantic segmentation), calculated the TSNE embedding of these features, and colored them according to the ground truth class of that feature pixel,

Hi， I sucee install multicore-TSNE. Then , I have some problems about the detials. I use tsne in the test.py. and the data are all too big, it spend too many time(9 hours), it dosen't visual. do you remeber you use data to visual. for example, 

the feature size = torch.size([2097152, 19])

it's too big. Do you resize the feature.