search

atong01 / conditional-flow-matching

TorchCFM: a Conditional Flow Matching library
https://arxiv.org/abs/2302.00482
MIT License
875 stars 60 forks source link

Improvement suggestions for the tutorial on flow matching #88

Open kilianFatras opened 6 months ago

kilianFatras commented 6 months ago

This issue is opened for users to suggest improvements for the Flow Matching tutorial notebook.

radiradev commented 6 months ago

Is it possible to include a minimal example of how a trained model can be used to evaluate the density of a given sample?

atong01 commented 6 months ago

@radiradev Thats a great idea. I can work on this. If you have an immediate need or are interested in making one I can send over a dirty notebook for this.

radiradev commented 6 months ago

@atong01 That would be very much appreciated!

kilianFatras commented 5 months ago

We should give credit to stochastic interpolants and rectified flows at the beginning of the notebook. Basically saying that all 3 methods are similar and concurrent ICLR2023 papers

hinzflorian commented 4 months ago

Hi, there is a typo in Flow_matching_tutorial.ipynb: "optximizer.zero_grad()" ->"optimizer.zero_grad()"

kilianFatras commented 4 months ago

Hi, Thank you! I will correct that shortly.

radiradev commented 4 months ago

Is it possible to include a minimal example of how a trained model can be used to evaluate the density of a given sample?

Hi @atong01, could you please share your example notebook on how this is done? I saw here that to evaluate the density an ode must be solved, but I am not sure how to accomplish this in pytorch:

jax-fmx

harveymannering commented 3 months ago

Hello,

I cannot run the last cell in the Flow_matching_tutorial.ipynb notebook because the sample_xt function is not define. Does sample_xt serve the same purpose as sample_conditional_pt? Are the two functions interchangable?

Thank you for the tutorials!

kilianFatras commented 3 months ago

Oh that’s indeed a typo… I changed the name of the function last minute and forgot to change this. You are right! The correct function is sample_conditional_pt.

Edit: I have pushed a corrected tutorial.

csufangyu commented 1 month ago

We should give credit to stochastic interpolants and rectified flows at the beginning of the notebook. Basically saying that all 3 methods are similar and concurrent ICLR2023 papers

Your suggestion is fantastic, I'm a beginner and it looks like there is no difference between stochastic interpolants and rectified flows, except that stochastic interpolants have an extra random term and rectified flows seem to be a deterministic sampling, is my understanding Is my understanding correct? I look forward to your reply, thanks!