gpu parallelisation

examples/paper_results/figure_3/eigenvalues_example.py

@@ -1,7 +1,7 @@
 import networkx as nx
 import numpy as np
 import matplotlib.pyplot as plt
-from geocluster import curvature as cv
+from geometric_clustering import curvature as cv
 from tqdm import tqdm
 from scipy.sparse.linalg import eigsh
 import scipy as sp
@@ -10,9 +10,7 @@
 
 if __name__ == "__main__":
 
-    tau = 0.5
-    l = 2
-    g = 4000
+    l = 3
     seed = 1
     k_in, k_out = 4, 1 #this is in the sparse regime
 
@@ -70,7 +68,6 @@
     # for i in range(10):
     #     plt.plot(v[:,i])
 
-
     #when summing over eigenvectors in clusters and taking differences, the fluctuations vanish
     # g_range = np.arange(50,2000,500)
     g_range = [1000]
@@ -79,44 +76,49 @@
     for g_ in g_range:
         graph = nx.planted_partition_graph(l, g_, k_in/g_, k_out/g_, seed=seed)
         largest_cc = max(nx.connected_components(graph), key=len)
-        graph = graph.subgraph(largest_cc)
+        graph = graph.subgraph(largest_cc).copy()
+        mapping = dict(zip(graph, range(len(graph.nodes))))
+        graph = nx.relabel_nodes(graph, mapping)
 
         laplacian = cv._construct_laplacian(graph,use_spectral_gap=False)
 
         t = time.time()
-        w, v = sp.linalg.eig(laplacian.toarray())
-        # w,v = eigsh(laplacian.toarray(), k=500, which='SM')
+        # w, v = sp.linalg.eig(laplacian.toarray())
+        w,v = eigsh(laplacian.toarray(), k=200, which='SM')
         print(time.time() - t)
 
-        C_1 = [i for i, k in enumerate(largest_cc) if k<g_]
-        C_2 = [j for j, k in enumerate(largest_cc) if k>=g_]
+        # C_1 = [i for i, k in enumerate(largest_cc) if k<g_]
+        # C_2 = [j for j, k in enumerate(largest_cc) if k>=g_]
         diffs = []
+        pairs = np.vstack(list(graph.edges))
+        C_1 = list(pairs[:,0])
+        C_2 = list(pairs[:,1])
         for s in tqdm(range(v.shape[1])):
             diff = (v[C_1,s].sum() - v[C_2,s].sum())/np.sqrt(g_)
-            diffs.append(diff)#np.exp(-w[s]/lambda2)*
+            diffs.append(diff)
 
         sum_v_g.append(diffs)
 
+
     all_g = [g_ for i, g_ in enumerate(g_range) for j in range(len(sum_v_g[i]))]
 
-    plt.figure()
-    plt.scatter(all_g,abs(np.hstack(sum_v_g)))
-    plt.ylabel(r'$|\sum_{ij} (\phi_s(i) - \phi_s(j))\,\delta(C_i,C_j)|$')
-    plt.xlabel('number of nodes')
+    # plt.figure()
+    # plt.scatter(all_g,abs(np.hstack(sum_v_g)))
+    # plt.ylabel(r'$|\sum_{ij} (\phi_s(i) - \phi_s(j))\,\delta(C_i,C_j)|$')
+    # plt.xlabel('number of nodes')
 
-    ind = np.where(abs(np.hstack(sum_v_g))==max(abs(np.hstack(sum_v_g))))[0]
-    plt.figure()
-    plt.plot(v[:,ind])
-    plt.savefig('phi2.svg')
+    ind = np.argsort(np.abs(sum_v_g)).flatten()[::-1]
+    for i in range(l-1):
+        plt.figure()
+        plt.plot(v[:,ind[i]])
+    #plt.savefig('phi2.svg')
 
     plt.figure()
-    ind = np.argsort(w)
     # colors = cm.rainbow(np.linspace(0, 1, len(w)))
     color = []
     for i, y in enumerate(abs(np.array(sum_v_g).flatten())):
         plt.scatter(w[i], y, color=cm.turbo(y))
         color.append(cm.turbo(y))
-    # plt.scatter(w[ind],abs(np.array(sum_v_g)))
     plt.savefig('diffusion_differece.svg')
     #plt.axvline(lambda2,c='r')
 
@@ -127,52 +129,7 @@
     for s in range(v.shape[1]):
         corr.append(gt.dot(v[:,s]))
 
-    plt.figure()
-    for i, y in enumerate(corr):
-        plt.scatter(w[i],y,color=color[i])
-    plt.savefig('correlation.svg')
-
-
-    # diffs = []
-    # eigs = []
-    # for g_ in tqdm(range(5)):
-    #     graph = nx.planted_partition_graph(l, g, k_in/g, k_out/g, seed=seed)
-    #     largest_cc = max(nx.connected_components(graph), key=len)
-    #     graph = graph.subgraph(largest_cc)
-
-    #     laplacian = cv._construct_laplacian(graph,use_spectral_gap=False)
-    #     w, v = np.linalg.eig(laplacian.toarray())
-
-    #     C_1 = [i for i, k in enumerate(largest_cc) if k<g_]
-    #     C_2 = [j for j, k in enumerate(largest_cc) if k>=g_]
-    #     mean_v = []
-
-    #     for s in range(len(v)):     
-    #         diffs.append((v[C_1,s].sum() - v[C_2,s].sum())/np.sqrt(len(largest_cc)))
-    #         eigs.append(w[s])
-    #         # mean_v.append(np.mean(diffs))
-
-    #     # sum_v_g.append(np.array(mean_v))
-
     # plt.figure()
-    # plt.scatter(w,abs(sum_v_g[-1]))
-    # plt.axvline(lambda2,c='r')
-
-    # g=1000
-    # graph = nx.planted_partition_graph(l, g, k_in/g, k_out/g, seed=seed)
-    # largest_cc = max(nx.connected_components(graph), key=len)
-    # graph = graph.subgraph(largest_cc)
-
-    # laplacian = cv._construct_laplacian(graph,use_spectral_gap=False)
-    # _, v = np.linalg.eig(laplacian.toarray())
-
-    # C_1 = [i for i, k in enumerate(largest_cc) if k<g]
-    # C_2 = [j for j, k in enumerate(largest_cc) if k>=g]
-
-    # evec_diff = []
-    # pairs = [(x,y) for x in C_1 for y in C_2]
-    # # for s in range(len(v)):
-    # for pair in pairs:
-    #     evec_diff.append(v[pair[0],5]- v[pair[1],5])
-
-    # plt.hist(evec_diff,bins=50)
+    # for i, y in enumerate(corr):
+    #     plt.scatter(w[i],y,color=color[i])
+    # plt.savefig('correlation.svg')