Skip to content

Latest commit

 

History

History
15 lines (12 loc) · 1.19 KB

README.md

File metadata and controls

15 lines (12 loc) · 1.19 KB

mnist_tsne

t-sne visualization of mnist images when feature is represented by raw pixels and cnn learned feature

something to say

  • the training code is from pytorch mnist example. The accuracy is 98% when use the original code, when bn is used in convolution and fully connected layer, the accuracy is 99. The training code here is with bn.
  • the code for t-sne visualization is from danielfrg/tsne
  • you can find the original mnist train raw data(60000x784), lable(60000x1), cnn learned feature(60000x50), t-sne generated feature(60000x2) for raw data and cnn learned feature, trained model in Baidu Pan or Google Drive
  • tsne_vis.ipynb is used to do tsne and visualization

visualization

t-sne of raw image pixel t-sne of cnn learned feature

from above visualization, it is shown that t-sne of cnn learned feature is more centered and cleaner than that of t-sne of raw image pixel