search

lmcinnes / umap

Uniform Manifold Approximation and Projection
BSD 3-Clause "New" or "Revised" License
7.37k stars 799 forks source link

Parametric umap worse than umap #657

Open balajiselvaraj1601 opened 3 years ago

balajiselvaraj1601 commented 3 years ago

I carried out experiments with parametric umap and umap

umap

  1. Clusters spread out throughout the entire space.
  2. There are datapoints in between the clusters, that are representing the transition from one cluster to another {which I can notice in the data}

parametric umap:

  1. The final map has few dense clusters that are present only in a small region
  2. I can't observe any transitions

I used the default parameters for both umap and parametric umap. I tried different model architectures. I tried tuning the hyper-parameters for training such as batch size and number of epochs

Is there any reason that we are not able to get a good performance from p-umap similar to that of the umap ?

timsainb commented 3 years ago

Hi @balajiselvaraj1601. Can you send an example? Like a colab notebook.

balajiselvaraj1601 commented 3 years ago

Hi @timsainb,

I used the default settings in umap and p-umap. I didnt do any modifications on it.

Thanks, S.Balaji

timsainb commented 3 years ago

Would you be able to send an example? I ask because I cannot help without being able to reproduce your issue.

timsainb commented 3 years ago

hi @balajiselvaraj1601 any update?