search

facebookresearch / poincare-embeddings

PyTorch implementation of the NIPS-17 paper "Poincaré Embeddings for Learning Hierarchical Representations"
Other
1.68k stars 230 forks source link

How do you create the visualization of embeddings in hyperbolic space: wn-nouns.jpg? #4

Open alphamupsiomega opened 6 years ago

Helw150 commented 6 years ago

@mnick I am also curious about visualization methods for these embeddings. Is there anything special to take into account when visualizing embeddings from hyperbolic space?

Seeing your visualization code would of course be helpful, but I'm assuming there is a consideration that keeps you from posting that in this repo.

alex-bloom commented 5 years ago

same question here... what technique was used for mapping embeddings from, say, 5-dimensional Poincare ball (mammals) to 2-dimensional Euclidian ring for visualization? I tried SVD and while latent hierarchies survive to some extent after dimensionality reduction, the resulting plot is not that sharp as the one in the paper

alex-bloom commented 5 years ago

On the second thought, we all seem to be looking in a dark room for a black cat that is not there. Taking a closer look at the paper, it says explicitly "Figure 2 shows a visualization of a two-dimensional Poincaré embedding." And again, in the Figure 2 capture: "Figure 2: Two-dimensional Poincaré embeddings of transitive closure of the WORDNET mammals subtree... A Poincaré embedding with d = 5 achieves mean rank 1.26 and MAP 0.927 on this subtree" - apparently, to make it clear that the Figure 2 does not represent the metrics obtained with the proper number of dimensions of the Poincaré ball

lematt1991 commented 5 years ago

Yes, sorry for the late response, but @alex-bloom, you are correct. The figure is of an embedding that is trained in only 2 dimensions (no dimensionality reduction is done).

rtatgit commented 4 years ago

Yes, sorry for the late response, but @alex-bloom, you are correct. The figure is of an embedding that is trained in only 2 dimensions (no dimensionality reduction is done).

Hi, Can you tell me after how many epochs did you achieve the above mentioned metrics?

lematt1991 commented 4 years ago

Which metrics are you referring to?

rtatgit commented 4 years ago

Which metrics are you referring to?

A Poincaré embedding with d = 5 achieves mean rank 1.26 and MAP 0.927. These evaluation metrics is what I was referring to.

lematt1991 commented 4 years ago

Please create a separate issue. This issue is regarding the visualization provided in the README

joelkuiper commented 4 years ago

Just wondering if there was any update on sharing the methods or code that generated the picture in README!

adiell commented 4 years ago

+1

yiqingxyq commented 3 years ago

Same here!

jayachaturvedi commented 2 years ago

Same here, would like to know how to generate the picture please. Thanks!