def syn_community1v2(args, writer=None, export_graphs=False):
# data
graphs1 = datagen.gen_ba(
range(40, 60), range(4, 5), 500,
featgen.ConstFeatureGen(np.ones(args.input_dim, dtype=float)))
for G in graphs1:
G.graph['label'] = 0
if export_graphs:
util.draw_graph_list(graphs1[:16], 4, 4, 'figs/ba')
graphs2 = datagen.gen_2community_ba(
range(20, 30), range(4, 5), 500, 0.3,
[featgen.ConstFeatureGen(np.ones(args.input_dim, dtype=float))])
for G in graphs2:
G.graph['label'] = 1
if export_graphs:
util.draw_graph_list(graphs2[:16], 4, 4, 'figs/ba2')
graphs = graphs1 + graphs2
train_dataset, val_dataset, test_dataset, max_num_nodes, input_dim, assign_input_dim = prepare_data(
graphs, args)
if args.method == 'soft-assign':
print('Method: soft-assign')
model = encoders.SoftPoolingGcnEncoder(
max_num_nodes,
input_dim,
args.hidden_dim,
args.output_dim,
args.num_classes,
args.num_gc_layers,
args.hidden_dim,
assign_ratio=args.assign_ratio,
num_pooling=args.num_pool,
bn=args.bn,
linkpred=args.linkpred,
assign_input_dim=assign_input_dim).cuda()
elif args.method == 'base-set2set':
print('Method: base-set2set')
model = encoders.GcnSet2SetEncoder(input_dim,
args.hidden_dim,
args.output_dim,
2,
args.num_gc_layers,
bn=args.bn).cuda()
else:
print('Method: base')
model = encoders.GcnEncoderGraph(input_dim,
args.hidden_dim,
args.output_dim,
2,
args.num_gc_layers,
bn=args.bn).cuda()
train(train_dataset,
model,
args,
val_dataset=val_dataset,
test_dataset=test_dataset,
writer=writer)
def pkl_task(args, feat=None):
with open(os.path.join(args.datadir, args.pkl_fname), 'rb') as pkl_file:
data = pickle.load(pkl_file)
graphs = data[0]
labels = data[1]
test_graphs = data[2]
test_labels = data[3]
for i in range(len(graphs)):
graphs[i].graph['label'] = labels[i]
for i in range(len(test_graphs)):
test_graphs[i].graph['label'] = test_labels[i]
if feat is None:
featgen_const = featgen.ConstFeatureGen(np.ones(args.input_dim, dtype=float))
for G in graphs:
featgen_const.gen_node_features(G)
for G in test_graphs:
featgen_const.gen_node_features(G)
train_dataset, test_dataset, max_num_nodes = prepare_data(graphs, args, test_graphs=test_graphs)
model = encoders.GcnEncoderGraph(
args.input_dim, args.hidden_dim, args.output_dim, args.num_classes,
args.num_gc_layers, bn=args.bn).cuda()
train(train_dataset, model, args, test_dataset=test_dataset)
evaluate(test_dataset, model, args, 'Validation')
def benchmark_task_val(args, writer=None, feat='node-label'):
all_vals = []
graphs = load_data.read_graphfile(args.datadir,
args.bmname,
max_nodes=args.max_nodes)
example_node = util.node_dict(graphs[0])[0]
if feat == 'node-feat' and 'feat_dim' in graphs[0].graph:
print('Using node features')
input_dim = graphs[0].graph['feat_dim']
elif feat == 'node-label' and 'label' in example_node:
print('Using node labels')
for G in graphs:
for u in G.nodes():
util.node_dict(G)[u]['feat'] = np.array(
util.node_dict(G)[u]['label'])
else:
print('Using constant labels')
featgen_const = featgen.ConstFeatureGen(
np.ones(args.input_dim, dtype=float))
for G in graphs:
featgen_const.gen_node_features(G)
for i in range(10):
train_dataset, val_dataset, max_num_nodes, input_dim, assign_input_dim = \
cross_val.prepare_val_data(graphs, args, i, max_nodes=args.max_nodes)
if args.method == 'amg-assign':
print('Method: amd-assign')
model = encoders.AmgPoolingGcnEncoder(
max_num_nodes,
input_dim,
args.hidden_dim,
args.output_dim,
args.num_classes,
args.num_gc_layers,
args.hidden_dim,
assign_ratio=args.assign_ratio,
num_pooling=args.num_pool,
bn=args.bn,
dropout=args.dropout,
linkpred=args.linkpred,
args=args,
assign_input_dim=assign_input_dim).cuda()
_, val_accs = train(train_dataset,
model,
args,
val_dataset=val_dataset,
test_dataset=None,
writer=writer)
all_vals.append(np.array(val_accs))
all_vals = np.vstack(all_vals)
all_vals = np.mean(all_vals, axis=0)
print(all_vals)
print(np.max(all_vals))
print(np.argmax(all_vals))
def gen_er(n_range, p, num_graphs, feature_generator=None):
graphs = []
for i in np.random.choice(n_range, num_graphs):
graphs.append(nx.erdos_renyi_graph(i, p))
if feature_generator is None:
feature_generator = featgen.ConstFeatureGen(0)
for G in graphs:
feature_generator.gen_node_features(G)
return graphs
def gen_ba(n_range, m_range, num_graphs, feature_generator=None):
graphs = []
for i in np.random.choice(n_range, num_graphs):
for j in np.random.choice(m_range, 1):
graphs.append(nx.barabasi_albert_graph(i, j))
if feature_generator is None:
feature_generator = featgen.ConstFeatureGen(0)
for G in graphs:
feature_generator.gen_node_features(G)
return graphs
def benchmark_task(args, writer=None, feat='node-label'):
graphs = load_data.read_graphfile(args.datadir,
args.bmname,
max_nodes=args.max_nodes)
if feat == 'node-feat' and 'feat_dim' in graphs[0].graph:
print('Using node features')
input_dim = graphs[0].graph['feat_dim']
elif feat == 'node-label' and 'label' in graphs[0].node[0]:
print('Using node labels')
for G in graphs:
for u in G.nodes():
G.node[u]['feat'] = np.array(G.node[u]['label'])
else:
print('Using constant labels')
featgen_const = featgen.ConstFeatureGen(
np.ones(args.input_dim, dtype=float))
for G in graphs:
featgen_const.gen_node_features(G)
train_dataset, val_dataset, test_dataset, max_num_nodes, input_dim, assign_input_dim = \
prepare_data(graphs, args, max_nodes=args.max_nodes)
if args.method == 'amg-assign':
print('Method: amg-assign')
model = encoders.AmgPoolingGcnEncoder(
max_num_nodes,
input_dim,
args.hidden_dim,
args.output_dim,
args.num_classes,
args.num_gc_layers,
args.hidden_dim,
assign_ratio=args.assign_ratio,
num_pooling=args.num_pool,
bn=args.bn,
dropout=args.dropout,
linkpred=args.linkpred,
args=args,
assign_input_dim=assign_input_dim).cuda()
train(train_dataset,
model,
args,
val_dataset=val_dataset,
test_dataset=test_dataset,
writer=writer)
evaluate(test_dataset, model, args, 'Validation')
def syn_community2hier(args, writer=None):
# data
feat_gen = [featgen.ConstFeatureGen(np.ones(args.input_dim, dtype=float))]
graphs1 = datagen.gen_2hier(1000, [2, 4], 10, range(4, 5), 0.1, 0.03,
feat_gen)
graphs2 = datagen.gen_2hier(1000, [3, 3], 10, range(4, 5), 0.1, 0.03,
feat_gen)
graphs3 = datagen.gen_2community_ba(range(28, 33), range(4, 7), 1000, 0.25,
feat_gen)
for G in graphs1:
G.graph['label'] = 0
for G in graphs2:
G.graph['label'] = 1
for G in graphs3:
G.graph['label'] = 2
graphs = graphs1 + graphs2 + graphs3
train_dataset, val_dataset, test_dataset, max_num_nodes, input_dim, assign_input_dim = prepare_data(
graphs, args)
if args.method == 'soft-assign':
print('Method: soft-assign')
model = encoders.SoftPoolingGcnEncoder(
max_num_nodes,
input_dim,
args.hidden_dim,
args.output_dim,
args.num_classes,
args.num_gc_layers,
args.hidden_dim,
assign_ratio=args.assign_ratio,
num_pooling=args.num_pool,
bn=args.bn,
linkpred=args.linkpred,
args=args,
assign_input_dim=assign_input_dim).cuda()
train(train_dataset,
model,
args,
val_dataset=val_dataset,
test_dataset=test_dataset,
writer=writer)
def benchmark_task_val(args, feat='node-label', pred_hidden_dims = [50], device='cpu'):
all_vals = []
data_out_dir = 'data/data_preprocessed/' + args.bmname + '/pool_sizes_' + args.pool_sizes
if args.normalize ==0:
data_out_dir = data_out_dir + '_nor_' + str(args.normalize)
data_out_dir = data_out_dir + '/'
if not os.path.exists(data_out_dir):
os.makedirs(data_out_dir)
graph_list_file_name = data_out_dir + 'graphs_list.p'
dataset_file_name = data_out_dir + 'dataset.p'
if os.path.isfile(graph_list_file_name) and os.path.isfile(dataset_file_name):
print('Files exist, reading from stored files....')
print('Reading file from', data_out_dir)
with open(dataset_file_name, 'rb') as f:
graphs = pickle.load(f)
with open(graph_list_file_name, 'rb') as f:
graphs_list = pickle.load(f)
print('Data loaded!')
else:
print('No files exist, preprocessing datasets...')
graphs = load_data.read_graphfile(args.datadir,args.bmname, max_nodes =args.max_nodes)
print('Data length before filtering: ', len(graphs))
dataset_copy = graphs.copy()
len_data = len(graphs)
graphs_list = []
pool_sizes = [int(i) for i in args.pool_sizes.split('_')]
print('pool_sizes: ', pool_sizes )
for i in range(len_data):
adj = nx.adjacency_matrix(dataset_copy[i])
# print('Adj shape',adj.shape)
if adj.shape[0] < args.min_nodes or adj.shape[0]> args.max_nodes or adj.shape[0]!= dataset_copy[i].number_of_nodes():
graphs.remove(dataset_copy[i])
# index_list.remove(i)
else:
# print('----------------------', i, adj.shape)
number_of_nodes = dataset_copy[i].number_of_nodes()
# if args.pool_ratios is not None:
# pool_sizes = []
# pre_layer_number_of_nodes = number_of_nodes
# for i in range(len(pool_ratios)):
# number_of_nodes_after_pool = int(pre_layer_number_of_nodes*pool_ratios[i])
# pool_sizes.append(number_of_nodes_after_pool)
# pre_layer_number_of_nodes = number_of_nodes_after_pool
# print('Test pool_sizes: ', pool_sizes)
coarsen_graph = gp(adj.todense().astype(float), pool_sizes)
# if args.method == 'wave':
coarsen_graph.coarsening_pooling(args.normalize)
graphs_list.append(coarsen_graph)
print('Data length after filtering: ', len(graphs), len(graphs_list))
print('Dataset preprocessed, dumping....')
with open(dataset_file_name, 'wb') as f:
pickle.dump(graphs, f)
with open(graph_list_file_name, 'wb') as f:
pickle.dump(graphs_list, f)
print('Dataset dumped!')
if feat == 'node-feat' and 'feat_dim' in graphs[0].graph:
print('Using node features')
input_dim = graphs[0].graph['feat_dim']
elif feat == 'node-label' and 'label' in graphs[0].node[0]:
print('Using node labels')
for G in graphs:
for u in G.nodes():
G.node[u]['feat'] = np.array(G.node[u]['label'])
else:
print('Using constant labels')
featgen_const = featgen.ConstFeatureGen(np.ones(args.input_dim, dtype=float))
for G in graphs:
featgen_const.gen_node_features(G)
# total_test_ac = 0
# total_test_best_ac = 0
# total_best_val_ac = 0
for i in range(10):
if i == args.shuffle:
if args.with_test:
train_dataset, val_dataset, test_dataset, max_num_nodes, input_dim = \
prepare_data(graphs, graphs_list, args, test_graphs = None,max_nodes=args.max_nodes, seed = i)
else:
train_dataset, val_dataset, test_dataset, max_num_nodes, input_dim = \
prepare_data(graphs, graphs_list, args, test_graphs = [],max_nodes=args.max_nodes, seed = i)
out_dir = args.bmname+ '/tar_' + str(args.train_ratio) + '_ter_' + str(args.test_ratio) + '/' + 'num_shuffle' + str(args.num_shuffle) + '/' + 'numconv_' + str(args.num_gc_layers) + '_dp_' + str(args.dropout) + '_wd_' + str(args.weight_decay) + '_b_' + str(args.batch_size) + '_hd_' + str(args.hidden_dim) + '_od_' + str(args.output_dim) + '_ph_' + str(args.pred_hidden) + '_lr_' + str(args.lr) + '_concat_' + str(args.concat)
out_dir = out_dir + '_ps_' + args.pool_sizes + '_np_' + str(args.num_pool_matrix) + '_nfp_' + str(args.num_pool_final_matrix) + '_norL_' + str(args.normalize) + '_mask_' + str(args.mask) + '_ne_' + args.norm + '_cf_' + str(args.con_final)
results_out_dir = args.out_dir + '/' + args.bmname + '/with_test' + str(args.with_test) + '/using_feat_' + args.feat + '/no_val_results/with_shuffles/' + out_dir + '/'
log_out_dir = args.out_dir + '/' + args.bmname + '/with_test' + str(args.with_test) + '/using_feat_' + args.feat + '/no_val_logs/with_shuffles/'+out_dir + '/'
if not os.path.exists(results_out_dir):
os.makedirs(results_out_dir, exist_ok=True)
if not os.path.exists(log_out_dir):
os.makedirs(log_out_dir, exist_ok=True)
results_out_file = results_out_dir + 'shuffle'+ str(args.shuffle) + '.txt'
log_out_file = log_out_dir + 'shuffle' + str(args.shuffle) + '.txt'
results_out_file_2 = results_out_dir + 'test_shuffle' + str(args.shuffle) + '.txt'
val_out_file = results_out_dir + 'val_result' + str(args.shuffle) + '.txt'
print(results_out_file)
with open(log_out_file, 'a') as f:
f.write('Shuffle ' +str(i) + '====================================================================================\n')
pool_sizes = [int(i) for i in args.pool_sizes.split('_')]
model = encoders.WavePoolingGcnEncoder(max_num_nodes, input_dim, args.hidden_dim, args.output_dim, args.num_classes, args.num_gc_layers, args.num_pool_matrix, args.num_pool_final_matrix,pool_sizes = pool_sizes, pred_hidden_dims = pred_hidden_dims, concat = args.concat,bn=args.bn, dropout=args.dropout, mask = args.mask,args=args, device=device)
if args.with_test:
_, val_accs, test_accs, best_val_result = train(train_dataset, model, args, val_dataset=val_dataset, test_dataset=test_dataset,
log_dir = log_out_file, device=device)
else:
_, val_accs, test_accs, best_val_result = train(train_dataset, model, args, val_dataset=val_dataset, test_dataset=None,
log_dir = log_out_file, device=device)
print('Shuffle ', i, '--------- best val result', best_val_result )
if args.with_test:
test_ac = test_accs[best_val_result['epoch']]
print('Test accuracy: ', test_ac)
best_val_ac = best_val_result['acc']
print('Best val on shuffle ', (args.shuffle), best_val_ac)
if args.with_test:
print('Test on shuffle', args.shuffle,' : ', test_ac)
np.savetxt(val_out_file, val_accs)
with open(results_out_file, 'w') as f:
f.write('Best val on shuffle '+ str(args.shuffle) + ': ' + str(best_val_ac) + '\n')
if args.with_test:
with open(results_out_file_2, 'w') as f:
f.write('Test accuracy on shuffle ' + str( args.shuffle ) + ':' + str(test_ac) + '\n')
with open(log_out_file,'a') as f:
f.write('Best val on shuffle ' + str(args.shuffle ) + ' : ' + str(best_val_ac) + '\n')
if args.with_test:
f.write('Test on shuffle ' + str( args.shuffle ) + ' : ' + str(test_ac) + '\n')
f.write('------------------------------------------------------------------\n')
def benchmark_task_val(args, writer=None, feat='node-label'):
all_vals = []
graphs = load_data.read_graphfile(args.datadir,
args.bmname,
max_nodes=args.max_nodes)
#args.max_nodes = 600, maxnodes
#print("maxnodes=",maxnodes),features
#compute_matching_degree(features)
if feat == 'node-feat' and 'feat_dim' in graphs[0].graph:
print('Using node features')
input_dim = graphs[0].graph['feat_dim']
elif feat == 'node-label' and 'label' in graphs[0].nodes[0]:
print('Using node labels')
for G in graphs:
for u in G.nodes():
G.nodes[u]['feat'] = np.array(G.nodes[u]['label'])
else:
print('Using constant labels')
featgen_const = featgen.ConstFeatureGen(
np.ones(args.input_dim, dtype=float))
for G in graphs:
featgen_const.gen_node_features(G)
for i in range(10):
print("****************", i)
train_dataset, val_dataset, max_num_nodes, input_dim, assign_input_dim = \
cross_val.prepare_val_data(graphs, args, i, max_nodes=args.max_nodes)
temp_input_dim = []
temp_assign_input_dim = []
for i in range(args.num_aspect):
temp_input_dim.append(input_dim)
temp_assign_input_dim.append(assign_input_dim)
input_dim_aspect = temp_input_dim
assign_input_dim_aspect = temp_assign_input_dim
#assign_input_di, max_graph_node_numm_aspect = [assign_input_dim, assign_input_dim, assign_input_dim]
if args.method == 'MxGNN':
print('Method: MxGNN')
if args.merge_method == 'cat':
model = MxGNNCat.SoftPoolingGcnEncoder(
max_num_nodes,
args.num_aspect,
args.multi_conv,
args.multi_pool,
input_dim_aspect,
args.hidden_dim,
args.output_dim,
args.num_classes,
args.num_gc_layers,
args.hidden_dim,
assign_ratio=args.assign_ratio,
num_pooling=args.num_pool,
bn=args.bn,
dropout=args.dropout,
linkpred=args.linkpred,
args=args,
assign_input_dim=assign_input_dim_aspect).cuda()
else:
model = MxGNNSum.SoftPoolingGcnEncoder(
max_num_nodes,
args.num_aspect,
args.multi_conv,
args.multi_pool,
input_dim_aspect,
args.hidden_dim,
args.output_dim,
args.num_classes,
args.num_gc_layers,
args.hidden_dim,
assign_ratio=args.assign_ratio,
num_pooling=args.num_pool,
bn=args.bn,
dropout=args.dropout,
linkpred=args.linkpred,
args=args,
assign_input_dim=assign_input_dim_aspect).cuda()
elif args.method == 'diffpool':
print('Method: diffpool')
model = encoders.SoftPoolingGcnEncoder(
max_num_nodes,
input_dim,
args.hidden_dim,
args.output_dim,
args.num_classes,
args.num_gc_layers,
args.hidden_dim,
assign_ratio=args.assign_ratio,
num_pooling=args.num_pool,
bn=args.bn,
dropout=args.dropout,
linkpred=args.linkpred,
args=args,
assign_input_dim=assign_input_dim).cuda()
else:
print('Method: base')
model = encoders.GcnEncoderGraph(input_dim,
args.hidden_dim,
args.output_dim,
args.num_classes,
args.num_gc_layers,
bn=args.bn,
dropout=args.dropout,
args=args).cuda()
_, val_accs = train(train_dataset,
model,
args,
val_dataset=val_dataset,
test_dataset=None,
writer=writer)
all_vals.append(np.array(val_accs))
all_vals = np.vstack(all_vals)
all_vals = np.mean(all_vals, axis=0)
print(all_vals)
print(np.max(all_vals))
print(np.argmax(all_vals))
result = "./result.txt"
f = open(result, 'a+')
f.write(str(np.max(all_vals)))
f.write("\n")
f.close()
def benchmark_task_val(args, writer=None, feat='node-label'):
all_vals = []
graphs = load_data.read_graphfile(args.datadir,
args.bmname,
max_nodes=args.max_nodes) # have value
example_node = util.node_dict(graphs[0])[0]
if feat == 'node-feat' and 'feat_dim' in graphs[0].graph:
print('Using node features')
input_dim = graphs[0].graph['feat_dim']
elif feat == 'node-label' and 'label' in example_node:
print('Using node labels')
for G in graphs:
for u in G.nodes():
util.node_dict(G)[u]['feat'] = np.array(
util.node_dict(G)[u]['label'])
else:
print('Using constant labels')
featgen_const = featgen.ConstFeatureGen(
np.ones(args.input_dim, dtype=float))
for G in graphs:
featgen_const.gen_node_features(G)
for i in range(
min([
10,
max([
int(
len(graphs) *
(1 - args.train_ratio - args.test_ratio)), 1
])
])):
'''I revised here 10-> min[...]'''
train_dataset, val_dataset, max_num_nodes, input_dim, assign_input_dim = \
cross_val.prepare_val_data(graphs, args, i, max_nodes=args.max_nodes)
if args.method == 'soft-assign':
print('Method: soft-assign')
model = encoders.SoftPoolingGcnEncoder(
max_num_nodes,
input_dim,
args.hidden_dim,
args.output_dim,
args.num_classes,
args.num_gc_layers,
args.hidden_dim,
assign_ratio=args.assign_ratio,
num_pooling=args.num_pool,
num_unpooling=args.num_unpool,
unpool_ratio=args.ratio_unpool,
bn=args.bn,
dropout=args.dropout,
linkpred=args.linkpred,
args=args,
assign_input_dim=assign_input_dim).cuda()
elif args.method == 'base-set2set':
print('Method: base-set2set')
model = encoders.GcnSet2SetEncoder(input_dim,
args.hidden_dim,
args.output_dim,
args.num_classes,
args.num_gc_layers,
bn=args.bn,
dropout=args.dropout,
args=args).cuda()
else:
print('Method: base')
model = encoders.GcnEncoderGraph(input_dim,
args.hidden_dim,
args.output_dim,
args.num_classes,
args.num_gc_layers,
bn=args.bn,
dropout=args.dropout,
args=args).cuda()
_, val_accs = train(train_dataset,
model,
args,
val_dataset=val_dataset,
test_dataset=None,
writer=writer)
all_vals.append(np.array(val_accs))
all_vals = np.vstack(all_vals)
all_vals = np.mean(all_vals, axis=0)
print('all_vals: ', all_vals)
print(np.max(all_vals))
print(np.argmax(all_vals))
def benchmark_task_val(args, writer=None, feat='node-feat'):
all_vals = []
#graphs = load_data.read_graphfile(args.datadir, args.bmname, max_nodes=args.max_nodes)
Hlist, New_G = load_data.read_graphfile2(args.datadir,
args.bmname,
max_nodes=args.max_nodes)
#example_node = util.node_dict(graphs[0])[0]
example_node = util.node_dict(New_G)[0]
#if feat == 'node-feat' and 'feat_dim' in graphs[0].graph:
if feat == 'node-feat' and 'feat_dim' in New_G.graph:
print('Using node features')
#input_dim =graphs[0].graph['feat_dim']
input_dim = New_G.graph['feat_dim']
elif feat == 'node-label' and 'label' in example_node:
print('Using node labels')
for G in graphs:
for u in G.nodes():
util.node_dict(G)[u]['feat'] = np.array(
util.node_dict(G)[u]['label'])
else:
print('Using constant labels')
featgen_const = featgen.ConstFeatureGen(
np.ones(args.input_dim, dtype=float))
for G in graphs:
featgen_const.gen_node_features(G)
# related to cross_val.py 13d row
#cross validation
# for i in range(10):
# # train_dataset, val_dataset, max_num_nodes, input_dim, assign_input_dim = \
# # cross_val.prepare_val_data(graphs, args, i, max_nodes=args.max_nodes)
# train_dataset, val_dataset, max_num_nodes, input_dim, assign_input_dim = \
# cross_val.prepare_val_data2(Hlist,New_G, args, i, max_nodes=args.max_nodes)
# if args.method == 'soft-assign':
# print('Method: soft-assign')
# model = encoders.SoftPoolingGcnEncoder(
# max_num_nodes,
# input_dim, args.hidden_dim, args.output_dim, args.num_classes, args.num_gc_layers,
# args.hidden_dim, assign_ratio=args.assign_ratio, num_pooling=args.num_pool,
# bn=args.bn, dropout=args.dropout, linkpred=args.linkpred, args=args,
# assign_input_dim=assign_input_dim).cuda()
# elif args.method == 'base-set2set':
# print('Method: base-set2set')
# model = encoders.GcnSet2SetEncoder(
# input_dim, args.hidden_dim, args.output_dim, args.num_classes,
# args.num_gc_layers, bn=args.bn, dropout=args.dropout, args=args).cuda()
# else:
# print('Method: base')
# model = encoders.GcnEncoderGraph(
# input_dim, args.hidden_dim, args.output_dim, args.num_classes,
# args.num_gc_layers, bn=args.bn, dropout=args.dropout, args=args).cuda()
#
# _, val_accs = train(train_dataset, model, args, val_dataset=val_dataset, test_dataset=None,
# writer=writer)
# all_vals.append(np.array(val_accs))
train_dataset, val_dataset, max_num_nodes, input_dim, assign_input_dim = \
cross_val.prepare_val_data2(Hlist, New_G, args, i, max_nodes=args.max_nodes)
if args.method == 'soft-assign':
print('Method: soft-assign')
model = encoders.SoftPoolingGcnEncoder(
max_num_nodes,
input_dim,
args.hidden_dim,
args.output_dim,
args.num_classes,
args.num_gc_layers,
args.hidden_dim,
assign_ratio=args.assign_ratio,
num_pooling=args.num_pool,
bn=args.bn,
dropout=args.dropout,
linkpred=args.linkpred,
args=args,
assign_input_dim=assign_input_dim).cuda()
elif args.method == 'base-set2set':
print('Method: base-set2set')
model = encoders.GcnSet2SetEncoder(input_dim,
args.hidden_dim,
args.output_dim,
args.num_classes,
args.num_gc_layers,
bn=args.bn,
dropout=args.dropout,
args=args).cuda()
else:
print('Method: base')
model = encoders.GcnEncoderGraph(input_dim,
args.hidden_dim,
args.output_dim,
args.num_classes,
args.num_gc_layers,
bn=args.bn,
dropout=args.dropout,
args=args).cuda()
_, val_accs = train(train_dataset,
model,
args,
val_dataset=val_dataset,
test_dataset=None,
writer=writer)
all_vals.append(np.array(val_accs))
all_vals = np.vstack(all_vals)
all_vals = np.mean(all_vals, axis=0)
print(all_vals)
print(np.max(all_vals))
print(np.argmax(all_vals))
def benchmark_task_val(args, writer=None, feat='node-label'):
all_vals = []
graphs = load_data.read_graphfile(args.datadir, args.bmname, max_nodes=args.max_nodes)
if feat == 'node-feat' and 'feat_dim' in graphs[0].graph:
print('Using node features')
input_dim = graphs[0].graph['feat_dim']
elif feat == 'node-label' and 'label' in graphs[0].node[0]:
print('Using node labels')
for G in graphs:
for u in G.nodes():
G.node[u]['feat'] = np.array(G.node[u]['label'])
else:
print('Using constant labels')
featgen_const = featgen.ConstFeatureGen(np.ones(args.input_dim, dtype=float))
for G in graphs:
featgen_const.gen_node_features(G)
for i in range(10):
train_dataset, val_dataset, max_num_nodes, input_dim, assign_input_dim = \
cross_val.prepare_val_data(graphs, args, i, max_nodes=args.max_nodes)
if args.method == 'soft-assign':
print('Method: soft-assign')
model = encoders.SoftPoolingGcnEncoder(
max_num_nodes,
input_dim, args.hidden_dim, args.output_dim, args.num_classes, args.num_gc_layers,
args.hidden_dim, assign_ratio=args.assign_ratio, num_pooling=args.num_pool,
bn=args.bn, dropout=args.dropout, linkpred=args.linkpred, args=args,
assign_input_dim=assign_input_dim).cuda()
elif args.method == 'base-set2set':
print('Method: base-set2set')
model = encoders.GcnSet2SetEncoder(
input_dim, args.hidden_dim, args.output_dim, args.num_classes,
args.num_gc_layers, bn=args.bn, dropout=args.dropout, args=args).cuda()
else:
print('Method: base')
model = encoders.GcnEncoderGraph(
input_dim, args.hidden_dim, args.output_dim, args.num_classes,
args.num_gc_layers, bn=args.bn, dropout=args.dropout, args=args).cuda()
_, val_accs = train(train_dataset, model, args, val_dataset=val_dataset, test_dataset=None,
writer=writer)
all_vals.append(np.array(val_accs))
del train_dataset, val_dataset, model
torch.cuda.empty_cache()
all_vals = np.vstack(all_vals)
with open('log_', 'a+') as f:
f.write('{},{},{},{},{},{},{},{},{},{},{},{},{},{},{},{},{},{},{},{},{},{},{},{}\n'.format(
'method', 'gc', 'dim',
'10-mean', '10-std',
'20-mean', '20-std',
'30-mean', '30-std',
'40-mean', '40-std',
'50-mean', '50-std',
'60-mean', '60-std',
'70-mean', '70-std',
'80-mean', '80-std',
'90-mean', '90-std',
'100-mean', '100-std', 'std'))
f.write('{},{},{},{},{},{},{},{},{},{},{},{},{},{},{},{},{},{},{},{},{},{},{},{}\n'.format(
args.method, args.num_gc_layers, args.output_dim,
all_vals[:, 10-1].mean(), all_vals[:, 10-1].std(),
all_vals[:, 20-1].mean(), all_vals[:, 20-1].std(),
all_vals[:, 30-1].mean(), all_vals[:, 30-1].std(),
all_vals[:, 40-1].mean(), all_vals[:, 40-1].std(),
all_vals[:, 50-1].mean(), all_vals[:, 50-1].std(),
all_vals[:, 60-1].mean(), all_vals[:, 60-1].std(),
all_vals[:, 70-1].mean(), all_vals[:, 70-1].std(),
all_vals[:, 80-1].mean(), all_vals[:, 80-1].std(),
all_vals[:, 90-1].mean(), all_vals[:, 90-1].std(),
all_vals[:, 100-1].mean(), all_vals[:, 100-1].std(),
np.max(np.mean(all_vals, axis=0))))
all_vals = np.mean(all_vals, axis=0)
print(all_vals)
print(np.max(all_vals))
print(np.argmax(all_vals))
from graph_sampler import GraphSampler
from torch.autograd import Variable
'''
Run.
Attach to encoder
Change to average pooling
'''
# syn_community1v2
input_dim = 10
n_range = range(40, 60)
m_range = range(4, 5)
num_graphs = 500
feature_generator = featgen.ConstFeatureGen(np.ones(input_dim, dtype=float))
graphs1 = datagen.gen_ba(n_range, m_range, num_graphs, feature_generator)
for G in graphs1:
G.graph['label'] = 0
n_range = range(20, 30)
m_range = range(4, 5)
num_graphs = 500
inter_prob = 0.3
feature_generators = [featgen.ConstFeatureGen(np.ones(input_dim, dtype=float))]
graphs2 = datagen.gen_2community_ba(n_range, m_range, num_graphs, inter_prob,
feature_generators)
for G in graphs2:
G.graph['label'] = 1
graphs = graphs1 + graphs2
