Merge branch 'main' of https://github.com/meyerscetbon/LinearGromov i…

…nto main

README.md

@@ -5,11 +5,11 @@ Code of the [paper](https://arxiv.org/pdf/2106.01128.pdf) by Meyer Scetbon, Gabr
 
 ## A New Way to Regularize the GW Problem
 In this work, we propose to regularize the Gromov-Wasserstein (GW) problem by constraining the admissible couplings to have a low-nonnegative rank: we call it the Low-Rank Gromov-Wasserstein (LR-GW) Problem. In the following figure, we compare the couplings obtained by the Prior Art method based on an entropic regularization and ours.
-![figure](results/couplings_intro.jpg)
+![figure](figures/coupling_plot.pdf)
 
 
 Our regularization can take also advantage of the geometry of the problem, in particular when the cost matrices involved in the GW problem admits a low-rank factorization. In this case, our method is able to compute the the LR-GW cost in linear time with respect to the number of samples. We present the time-accuracy tradeoff between different methods when the samples are drawn from two Gaussian distributions evaluated on 5000 points in 2D and the underlying cost is the squared Euclidean distance.
-![figure](results/plot_accuracy_LR_vs_All.jpg)
+![figure](figures/acc_plot.pdf)
 
 In this [file](https://github.com/meyerscetbon/LinearGromov/blob/main/toy_examples.py) we provide some toy examples where we compare the Entropic GW scheme with our proposed method.