search

chen0706 / EWM

Code for A Gradual, Semi-Discrete Approach to Generative Network Training via Explicit Wasserstein Minimization (Chen et al., ICML 2019)
4 stars 2 forks source link

Question about the stopping criteria #1

Open ariel415el opened 1 year ago

ariel415el commented 1 year ago

Hi, Thanks for uploading the code for this interesting paper.

At the end of the paper (appendix E.3) you wrote this:

"Our empirical stopping criterion relies upon keeping a histogram of transportation targets in memory: if the histogram of targets is close to a uniform distribution (which is the distribution of training dataset), we stop OTS. This stopping criterion is grounded by our analysis in Section 3.1."

In the code this is implemented in : https://github.com/chen0706/EWM/blob/ee3c358bca08f5f91a5d91ecf826d8de47e2d014/main.py#L108-L112

This is practically does the opposite right? it seems to checks if the range of indices is small enough i.e the indices are in (n early_end[0]/n_bins, n early_end[1]/n_bins)

Is this a bug? Can you explain the actual stopping criteria used here?

ariel415el commented 1 year ago

I just noticed that hist.max() is the height of the highest bin (and histo.min() lowest respectively) So making sure these are around n/n_bins makes sure the histogram is uniform.

Maybe you can explain how you chose the parameters early_end = (200,320)

I think its better to use np.histogram with density=True and make sure the min and max bins are around 1/n_bins

chen0706 commented 1 year ago

Thanks for your interest in the paper and the code!

The choice of the early stopping criterion (200, 320) is mainly for running speed. You may run OTS for more iterations to get closer to uniform distribution at the cost of more running time.

And yes, I agree that using density as the stopping criterion sounds more intuitive here.