Horse

.ipynb_checkpoints/Untitled-checkpoint.ipynb

@@ -0,0 +1,6 @@
+{
+ "cells": [],
+ "metadata": {},
+ "nbformat": 4,
+ "nbformat_minor": 1
+}

README.md

@@ -35,9 +35,8 @@ For examples with RISGW:
 
 * SGW function both in CPU and GPU (with Pytorch):
 
-<p align="center">
-  <img src="https://github.com/tvayer/SGW/blob/master/sgw.png" width="600" >
-</p>
+![](horses.gif)
+
 
 * Entropic Gromov-Wasserstein in Pytorch.
 

lib/sgw_numpy2.py

@@ -29,31 +29,6 @@ def dist(x1, x2):
     x2p2 = np.sum(np.square(x2), 1)
     return x1p2.reshape((-1, 1)) + x2p2.reshape((1, -1)) - 2 * np.dot(x1, x2.T)
 
-def sw0(xs,xt,P): # xs.shape(1)<=xt.shape(1), P.shape[0]=xt.shape[1]
-
-    L=P.shape[1]
-    n=xs.shape[0]
-    xsp=np.dot(xs,P[:xs.shape[1],:])
-    xtp=np.dot(xt,P)
-
-    #xsp=np.sort(xsp,0)
-    #xtp=np.sort(xtp,0)
-
-    res=0
-
-    for l in range(L):
-
-        x1=np.sort(xsp[:,l])
-        x2=np.sort(xtp[:,l])
-
-        D1=dist(x1[:,None],x2[:,None])
-
-        l1=np.sum(np.square(D1))/n/n
-        #l2=np.sum(D1[::-1,::-1])/n/n
-
-        res+=l1
-
-    return res/L
 
 def sgw0(xs,xt,P): # xs.shape(1)<=xt.shape(1), P.shape[0]=xt.shape[1]
 
@@ -62,9 +37,6 @@ def sgw0(xs,xt,P): # xs.shape(1)<=xt.shape(1), P.shape[0]=xt.shape[1]
     xsp=np.dot(xs,P[:xs.shape[1],:])
     xtp=np.dot(xt,P)
 
-    #xsp=np.sort(xsp,0)
-    #xtp=np.sort(xtp,0)
-
     res=0
 
     for l in range(L):
@@ -189,61 +161,8 @@ def loss(delta):
 
     return loss(Xopt), Xopt
 
-def risw(xs,xt,P):
-
-    def loss(delta):
-        return sw0(np.dot(xs,delta),xt,P)
-
-    manifold = Stiefel(xt.shape[1], xs.shape[1])
 
 
-    problem = Problem(manifold=manifold, cost=loss)
-
-    # (3) Instantiate a Pymanopt solver
-    solver = SteepestDescent(logverbosity=0)
-
-    # let Pymanopt do the rest
-    Xopt = solver.solve(problem,x=np.eye( xs.shape[1],xt.shape[1]))
-
-    return loss(Xopt)
-
-def risw2(xs,xt,P,X0=None):
-
-    def loss(delta):
-        return sw0(np.dot(xs,delta),xt,P)
-
-    if X0 is None:
-        X0=np.eye( xs.shape[1],xt.shape[1])
-
-
-    manifold = Stiefel(xt.shape[1], xs.shape[1])
 
 
-    problem = Problem(manifold=manifold, cost=loss)
-
-    # (3) Instantiate a Pymanopt solver
-    solver = SteepestDescent(logverbosity=0)
-
-    # let Pymanopt do the rest
-    Xopt = solver.solve(problem,x=X0)
-
-    return loss(Xopt), Xopt
-
-
-def gw0(xs,xt): # xs.shape(1)<=xt.shape(1), P.shape[0]=xt.shape[1]
-
-    n=xs.shape[0]
-    u=ot.unif(n)
-    D1=dist(xs,xs)
-    D2=dist(xt,xt)
-
-    return ot.gromov.gromov_wasserstein2(D1,D2,u,u,'square_loss')[0]
-
-def w0(xs,xt): # xs.shape(1)<=xt.shape(1), P.shape[0]=xt.shape[1]
-
-    n=xs.shape[0]
-    u=ot.unif(n)
-    D1=dist(xs,xt)
-
-    return ot.emd2(u,u,D1)