lr = 0.001
max_epoch = 200 # it takes a long epochs to converge, probably more than 2000
x_size = nfeature
hidden_size = intermediate_size = 32
num_attention_heads = 2
num_hidden_layers = 2
y_size = nclass
graph_size = ngraph
residual_type = 'graph_raw'
# --------------------------
print('************ Start ************')
print('GrapBert, dataset: ' + dataset_name + ', Pre-training, Node Attribute Reconstruction.')
# ---- objection initialization setction ---------------
data_obj = DatasetLoader()
data_obj.dataset_source_folder_path = './data/' + dataset_name + '/'
data_obj.dataset_name = dataset_name
data_obj.k = k
data_obj.load_all_tag = True
bert_config = GraphBertConfig(residual_type = residual_type, k=k, x_size=nfeature, y_size=y_size, hidden_size=hidden_size, intermediate_size=intermediate_size, num_attention_heads=num_attention_heads, num_hidden_layers=num_hidden_layers)
method_obj = MethodGraphBertNodeConstruct(bert_config)
method_obj.max_epoch = max_epoch
method_obj.lr = lr
method_obj.save_pretrained_path = './result/PreTrained_GraphBert/' + dataset_name + '/node_reconstruct_model/'
result_obj = ResultSaving()
result_obj.result_destination_folder_path = './result/GraphBert/'
result_obj.result_destination_file_name = dataset_name + '_' + str(k) + '_node_reconstruction'
max_graph_size = 40
elif dataset_name == 'PTC':
max_graph_size = 120
elif dataset_name == 'NCI1':
max_graph_size = 120
elif dataset_name == 'PROTEINS':
max_graph_size = 620
elif dataset_name == 'COLLAB':
max_graph_size = 500
#---- Step 1: Load Raw Graphs for Train/Test Partition ----
if 1:
print('************ Start ************')
print('Preprocessing dataset: ' + dataset_name)
# ---- objection initialization setction ---------------
data_obj = DatasetLoader()
data_obj.dataset_source_folder_path = './data/' + dataset_name + '/'
data_obj.dataset_name = dataset_name
data_obj.load_type = 'Raw'
method_obj = MethodProcessRaw()
result_obj = ResultSaving()
result_obj.result_destination_folder_path = './result/Preprocess/'
result_obj.result_destination_file_name = dataset_name
setting_obj = Settings()
evaluate_obj = None
# ------------------------------------------------------
ngraph = 2708
elif dataset_name == 'citeseer':
nclass = 6
nfeature = 3703
ngraph = 3312
elif dataset_name == 'pubmed':
nclass = 3
nfeature = 500
ngraph = 19717
#---- Step 1: WL based graph coloring ----
if 1:
print('************ Start ************')
print('WL, dataset: ' + dataset_name)
# ---- objection initialization setction ---------------
data_obj = DatasetLoader()
data_obj.dataset_source_folder_path = './data/' + dataset_name + '/'
data_obj.dataset_name = dataset_name
method_obj = MethodWLNodeColoring()
result_obj = ResultSaving()
result_obj.result_destination_folder_path = './result/WL/'
result_obj.result_destination_file_name = dataset_name
setting_obj = Settings()
evaluate_obj = None
# ------------------------------------------------------
# ---- running section ---------------------------------
x_size = nfeature
hidden_size = intermediate_size = 32
num_attention_heads = 2
num_hidden_layers = 2
y_size = nclass
graph_size = ngraph
residual_type = 'none'
# --------------------------
print('************ Start ************')
print('GrapBert, dataset: ' + dataset_name + ', residual: ' +
residual_type + ', k: ' + str(k) + ', hidden dimension: ' +
str(hidden_size) + ', hidden layer: ' + str(num_hidden_layers) +
', attention head: ' + str(num_attention_heads))
# ---- objection initialization setction ---------------
data_obj = DatasetLoader()
data_obj.dataset_source_folder_path = './result/Padding/'
data_obj.dataset_source_file_name = dataset_name
data_obj.k = k
bert_config = GraphBertConfig(residual_type=residual_type,
k=k,
x_size=nfeature,
y_size=y_size,
hidden_size=hidden_size,
intermediate_size=intermediate_size,
num_attention_heads=num_attention_heads,
num_hidden_layers=num_hidden_layers)
method_obj = MethodGraphBertNodeClassification(bert_config)
#---- set to false to run faster ----
method_obj.spy_tag = True
from code.DatasetLoader import DatasetLoader
import numpy as np
import scipy.sparse as sp
'''
To learn how data is loaded and passed to WL.
Also, how WL process the data and returns the output.
'''
if 0:
adj = np.random.randint(2, size=(5, 5))
print(adj)
adj = sp.coo_matrix(adj)
print(adj)
_adj_T = adj.T.multiply(adj.T > adj)
_adj = adj.multiply(adj.T > adj)
print(_adj_T.toarray())
print(_adj.toarray())
adj2 = adj + _adj_T - _adj
print(adj2.toarray())
assert False
dataset_name = 'cora'
data_obj = DatasetLoader()
data_obj.dataset_source_folder_path = './data/' + dataset_name + '/'
data_obj.dataset_name = dataset_name
data_obj.load()