some deeptime.data only produce one state

[BissuelD]

While running deeptime tutorials, I noticed deeptime.data.ellipsoids produces samples in one single state, seemingly regardless of the random seed provided or the sample size.

I did some tests on other datasets by directly running the provided notebooks without any modification.
Here are the list of datasets with similar issues I noticed, together with screenshots showing the state population issues:

• for Birth-death chain model
  

• for Metropolis chain in 1D energy landscape
  

• for One-dimensional Prinz potential
  

• for Triple-well 1D
  

• for Ellipsoids dataset
  

• for Sqrt model
  

Note : I did not try all provided dataset notebooks, so I don't know if the list above is exhaustive. Also, being quite new into Markov Models, I am sorry in advance if the terminolgy I use is incorrect or if one of the issues above is not related to the others.

I run on Debian 12 (bookworm), and use deeptime==0.4.4 installed via pip on a uv venv (see the full requirements.txt).

[clonker]

Hmmm. That is definitely not right. Can you do me a favor and create a venv (via python -m venv .venv), activate it (via source .venv/bin/activate), clone the deeptime repository and install it into the venv with pip install . assuming the current working directory is inside the cloned repository? Then try again? I am just trying to rule out that there is something wrong with your python setup :)

[BissuelD]

Thanks for your answer!

It seems there is indeed something wrong with the python setup I used: with cloned repositories and pip install . for both the latest official release (deeptime-0.4.4) and the latest commit of the code (deeptime-0.4.4+11.g5f85b3e) showed no issue.

After a few tests, I believe the pip install deeptime command is the culprit in my case. Here is a summary of what I tried (with a constant Python version ie. 3.10.7)

• Building in a fresh venv from the cloned repository and pip install . (see the above paragraph) succeeds with no issue.
• Using pip install deeptime in a fresh venv fails as described in my first post.
• Using uv add deeptime (or uv pip install deeptime) in a fresh uv venv also fails the same way.
• Creating a fresh conda env and using conda install deeptime seems to work fine.

My intuition was then to look at my pip command, but as 1) the pip install . from source works and 2) uv pip uses its own pip, not the one from the host, I suppose the issue is not that simple.

I do not have another machine available right now, but I guess the next step would be to see if using pip install deeptime on another machine allows to replicate the issue.